JP2020123229A - Abnormality sign detection system and abnormality sign detection method - Google Patents

Abstract

To detect a facility abnormality sign and to identify an abnormality factor.SOLUTION: A frequency calculation unit 5 of an abnormality sign detection system 1 calculates a power spectrum from oscillatory waveform data of a rotary machine 2 recorded in a data recording unit. A normal model creation unit 7 generates a normal model by Non-negative Matrix Factorization (NMF) on the basis of the power spectrum of the normal time collected in advance by a normal data recording unit 6. A comprehensive reconstruction error calculation unit 8 calculates an input-output error distribution when diagnostic data is input to the normal model, the oscillatory waveform data of the rotary machine 2 recorded in a diagnostic data recording unit 9 being used as the diagnostic data. When an error distribution of the diagnostic data deviates from the error distribution of the normal data, the abnormality determination unit 10 determines whether an abnormality sign occurred. A variable error output unit 1 identifies an oscillation frequency of the determined abnormality sign, and estimates an abnormality factor of the rotary machine 2 according to the identified oscillation frequency.SELECTED DRAWING: Figure 1

Description

The present invention relates to a technique for detecting a sign of abnormality based on sensing data of target equipment.

At present, while the number of engineers is insufficient due to population decline, a large amount of electrical equipment manufactured during the high economic growth period has reached the design life, and construction of an equipment diagnostic system utilizing "ICT/loT" is required.

At this time, failure of infrastructure equipment such as electric equipment is extremely rare, and therefore unsupervised learning in which only normal data is used as learning data is often used when performing data-driven diagnosis.

For example, Patent Document 1 proposes a diagnostic method using “One-class Support Vector Machine (OCSVM)” for an unsupervised learning approach.

JP, 2018-28845, A

Lee, D.D., Seung, H.S., “Algorithms for nonnegative matrix factorization”, Advances in Neural Information Processing Systeme 13,pp.556-562,(2000) Prognostics Center of Excellence of NASA:The prgnosticsdata repository.<URL:https://ti.arc.nasa.gov/tech/dash/pcoe/prognostic-data-repository/,> (2016)

In Patent Document 1, in order to detect an abnormality in a mechanical system of a rotating machine, only a data in a normal state is learned from a frequency component of vibration by "One-class Support Vector Machine (OCSVM)" to detect an abnormality sign.

However, since "OCSVM" performs classification in the feature space by the kernel method, it cannot be directly associated with the input dimension, it cannot be determined which frequency is abnormal, and it may not be possible to grasp the cause of abnormality of the target equipment. is there.

The present invention has been made in order to solve such a conventional problem, and an object of the invention is to identify an abnormality factor together with detection of a sign of equipment abnormality.

(1) One aspect of the present invention is an abnormality sign detection system for detecting an abnormality sign based on vibration waveform data of target equipment,
A normal model creation unit that creates a normal model by non-negative matrix factorization (NMF) based on the vibration waveform data collected in advance when the target facility is normal;
A vibration error data of the target facility is used as diagnostic data, and a reconstruction error calculation unit that obtains an input/output error distribution when the diagnostic data is input to the normal model,
If the error distribution of the diagnostic data deviates from the error distribution of the normal data, an abnormality determination unit that determines the occurrence of the abnormality sign,
An abnormality factor estimation unit that calculates an error for each frequency of the diagnostic data and estimates an abnormality factor of the target equipment according to the evaluation of the calculated error.

(2) Another aspect of the present invention is a method in which a computer detects an abnormality sign based on vibration waveform data of target equipment,
A normal model creating step for creating a normal model by non-negative matrix factorization (NMF) based on the vibration waveform data collected in advance when the target facility is normal;
Reconstruction error calculation step for obtaining the error distribution of input and output when the diagnostic equipment is the vibration waveform data of the target equipment and the diagnostic data is input to the normal model,
If the error distribution of the diagnostic data deviates from the error distribution of the normal data, an abnormality determination step of determining the occurrence of the abnormality sign,
An abnormality factor estimating step of identifying the vibration frequency of the determined abnormality sign and estimating the abnormality factor of the target facility according to the specified vibration frequency.

According to the present invention, it is possible to specify an abnormality factor together with detection of a sign of equipment abnormality.

1 is a system configuration diagram of an abnormality sign detection system according to an embodiment of the present invention. Schematic of NMF. The figure which shows an example of the relationship between a vibration frequency and an abnormal factor. (A) is a graph showing the basis matrix (H) of the embodiment, and (b) is a graph showing the same coefficient matrix (W). (A) is a graph which shows the distribution of the reconstruction error of an Example, (b) is the elements on larger scale of (a). (A) is a graph showing a reconstruction difference in a normal state of the embodiment, (b) is a graph showing a reconstruction difference of the sample number 705, and (c) is a graph showing a reconstruction difference immediately before a failure stop.

Hereinafter, an abnormality sign detection system (abnormal sign detection method) according to an embodiment of the present invention will be described. This abnormality detection system continuously collects vibration waveform data as sensing data from sensors (acceleration sensor, acoustic sensor, etc.) installed in target equipment such as electrical equipment, and detects abnormality of the target equipment based on the collected vibration waveform data. Detect and detect signs.

By catching this sign of abnormality, it is possible to take measures before the target equipment fails, and it is possible to reduce downtime of the infrastructure system and the like. At this time, since failure of infrastructure equipment is rare, it is difficult to obtain abnormal data in advance, and the main axis is unsupervised learning in which only normal data is used as learning data.

Further, as the abnormality sign detection system, it is preferable to output not only the abnormality degree but also information for specifying an abnormality factor, and a linear model is adopted as the learning model.

Prior art techniques for unsupervised anomaly detection with linear models often use Principal Component Analysis (PCA), which has a proven record in the field of statistical process management. However, since the amplitude spectrum of the vibration waveform data has a non-negative value, it cannot be directly matched with the physical event with the principal component analysis for taking a negative value, and it can be said that it is unsuitable in terms of model readability.

Therefore, the anomaly detection system uses the non-negative matrix factorization (hereinafter, referred to as NMF) of Non-Patent Document 1 to maintain the anomaly detection power and readability of the learned normal model. We propose a diagnostic method that explained the abnormality judgment result of.

This NMF is a non-negative constrained dimensional compression method that decomposes a non-negative matrix into two non-negative matrices of lower rank. It can clearly show the potential elements of the original matrix and An effective effect is obtained as a feature extraction method of a certain vibration amplitude spectrum. That is, the vibration spectrogram observed by the NMF is decomposed into the frequency components of the original signal source. At this time, the decomposed basis with non-negative values has the same dimension as the amplitude spectrum and has a physical meaning. Therefore, structural features can be analyzed in terms of interpretability and explanation of the model.

<< configuration example >>
The configuration of the abnormality sign detection system will be described with reference to FIG. Reference numeral 1 in FIG. 1 indicates the abnormality detection system. The anomaly detection system 1 executes anomaly detection using an optimization method having a non-negative constraint, and uses NMF for the feature amount extraction of the vibration spectrumgram and the anomaly detection method.

In the abnormality detection system 1, the rotating equipment (rotating machine) 2 is used as a diagnostic target equipment, and the rotating machine 2 is provided with a vibration sensor 3. The vibration frequency of the rotating machine 2 is detected by the vibration sensor 3, and an abnormality sign of the rotating machine 2 is detected based on the vibration frequency detected by the vibration sensor 3.

Specifically, the abnormality detection system 1 is configured by a computer, and includes normal computer hardware resources (eg, CPU, main storage device such as RAM/ROM, auxiliary storage device such as HDD/SSD).

As a result of the cooperation between the hardware resource and the software resource (OS, application, etc.), the abnormality sign detection system 1 includes the data recording unit 4, the frequency calculating unit 5, the normal data recording unit 6, the normal model creating unit 7, The total reconstruction error calculation unit 8, the diagnostic data recording unit 9, the abnormality determination unit 10, and the variable error output unit 11 are mounted. The recording units 4, 6 and 9 are respectively built in the storage device. Specifically, in the data recording unit 4, vibration waveform data obtained by A/D converting the vibration of the rotating machine 2 detected and acquired by the vibration sensor 3 is stored and recorded.

The frequency calculation unit 5 also calculates a power spectrum from the vibration waveform data recorded in the data recording unit 4. That is, the power for each frequency bandwidth is calculated from the power of the time signal of the vibration waveform data by using a fast Fourier transform (FFT), a constant Q transform (Constant-Q Transform), or the like.

Based on the power spectrum calculated here, the learning stages are executed by the respective units 6 to 8, while the diagnostic stages are executed by the respective units 8 to 11. This learning stage is executed before the diagnosis stage and generates a normal model based on the power spectrum calculated from the vibration waveform data of the rotating machine 2 during normal operation.

Further, the diagnosis stage diagnoses the presence/absence of an abnormality sign of the rotating machine 2 based on the vibration waveform data of the rotating machine 2 to be diagnosed, that is, the power spectrum calculated from the diagnosis data. The processing contents of the respective units 4 to 11 will be described below for each stage.

≪Learning stage≫
(1) The normal data recording unit 6 stores a power spectrum that is obvious that most of the data collected in advance are normal. That is, in the normal data recording unit 6, the power spectrum calculated from the vibration data during normal operation of the rotating machine 2 among the recording data of the data recording unit 4 is recorded. The power spectrum of the normal data recording unit 6 is called normal data.

(2) The normal model creating unit 7 creates a normal model (NMF model) by NMF using normal data as a learning sample, and normalizes the input data (standardization, normalization of 0 ^to 1 range, etc.) as necessary. To do.

The normal model created here is stored in the storage device, and the power spectrum is equivalent to the squared value of the amplitude. Therefore, the square root of the power spectrum is called the amplitude spectrum. The details of creating a normal model will be described below.

As shown in FIG. 2 and Expression (1), the NMF converts the vibration spectrogram V “i×μ” of the observed power spectrum into a non-negative coefficient matrix W(i×a) and a non-negative basis matrix H(a×μ). ) Consider the approximation with a linear sum of.

The square error criterion is adopted as the criterion for the degree of separation between the vibration spectrogram V and the matrix WH. Specifically, the coefficient matrix W and the basis matrix H are calculated so as to minimize the objective function represented by the equation (2-1) while keeping a nonnegative value. This matrix WH is used as the feature amount of the amplitude spectrum. At this point, a non-negative feature amount in the same dimension of the power spectrum can be obtained.

Here, "|||| _F "in Formula 2-1 has shown the Frobenius norm. Further, each element of the coefficient matrix W and the basis matrix H is calculated by iterative calculation of the update formulas shown in formulas (3-1) and (3-2).

(3) The total reconstruction error calculation unit 8 calculates the reconstruction error E _n from the coefficient matrix W and the basis matrix H for the amplitude spectrogram Vn in the normal state, as shown in Expression (4). This is defined as the error distribution of normal data (continuous probability distribution). Incidentally, since it is using the squared error criterion in equation (4), the reconstruction error E _n is assumed to follow a normal distribution.

The total reconstruction error calculation unit 8 also sets an appropriate first threshold value based on the error distribution of normal data, and determines abnormality of diagnostic data based on the set first threshold value. For example, assuming that the error distribution E _{n of} normal data follows the mean μ distribution δ ² , and as shown in Expression (5), the first threshold “1” is calculated from the mean μ and the standard deviation δ of the reconstruction errors E _n of each learning sample. Threshold (hereinafter referred to as S1)" can be set.

≪Diagnostic stage≫
(1) The diagnostic data recording unit 9 records the power spectrum calculated from the diagnostic waveform data of the diagnosis target in the recording data of the data recording unit 4. Here, the recording data (record) of the diagnostic data recording unit 9 is referred to as diagnostic data. The diagnostic data recording unit 9 can record diagnostic data for each working day of the rotating machine 2, for example.

(2) The total reconstruction error calculation unit 8 and the abnormality determination unit 10 determine the presence/absence of a sign of abnormality of the rotating machine 2 based on the diagnostic data of the diagnostic data recording unit 9. That is, the total reconstruction error calculation unit 8 inputs the diagnostic data for the normal model stored in the storage device to obtain the reconstruction error E _t .

At that time, as shown in Expression (6), with respect to the amplitude spectrogram Vt of the evaluation target based on the diagnostic data, the coefficient matrix W _t of the evaluation target is calculated by NMF while being fixed to the basis matrix H _n in the normal state. Then, the reconstruction error E _t is calculated. This is the error distribution of diagnostic data.

(3) The abnormality determination unit 10 compares the reconstruction error E _t of the equation (6) with the reconstruction error E _{n under} normal conditions. As a result of the comparison, if the reconstruction error E _t exceeds the first threshold value S1, it is determined that the error distribution E _n in the normal state is deviated, and an abnormality judgment is made on the diagnostic data, that is, the abnormality sign “There is an abnormality sign in the rotating machine 2. Is determined. By this abnormality determination, an abnormality sign of the rotating machine 2 is detected, and an alert is notified.

On the other hand, if the reconstruction error E _t does not exceed the first threshold value S1, it is determined that the error distribution E _{n is} in the normal state, and it is determined that the diagnostic data is normal, that is, "absent".

(4) Since the variable error output unit 11 indicates a portion that cannot be expressed by the combination with the basis matrix Hn in the normal state in Expression (6), that is, which frequency component is abnormal, input as shown in Expression (7) The difference between the evaluated object (V _t ) and the reconstructed matrix (W _t H _n ) is calculated and visualized. This difference is called the reconstruction difference D _t .

When the reconstruction difference D _t exceeds the second threshold value S2 preset for the error of the normal data, the frequency (variable) is evaluated as abnormal, and the frequency of the rotating machine 2 is determined from the frequency range evaluated as abnormal. Estimate the cause of abnormality. Hereinafter, a method of estimating an abnormal factor will be described.

That is, it is known that the failure of the mechanical system of the rotating machine 2 appears as an inherent vibration. FIG. 3 shows an example of the relationship between the vibration frequency and the cause of abnormality of the rotating machine 2. Here, since the modulation in the low frequency region includes the rotation frequency, it is suspected that imbalance or misbalance of the rotating body may occur. On the other hand, in the high frequency region, it is considered that the waveform of the shock system is included, and it is suspected that the scratches on the bearing and the local abnormality of the rotating body are caused.

Then, when the abnormality of the rotating machine 2 is detected, it is possible to estimate the cause of abnormality of the rotating machine 2 if it is known which frequency indicates the abnormality. Regarding this point, according to the abnormality sign detection system 1, when the reconstruction difference D _t exceeds the second threshold value S2 set to the error of the normal data, it is evaluated as the frequency abnormality. By comparing the frequency range evaluated to be abnormal with the relationship diagram of FIG. 3 or the like, it is possible to estimate the cause of abnormality of the rotating machine 2.

Further, the estimated abnormality factor is output to a monitor or the like together with the abnormality determination result and presented to the user. As a result, the user can identify and grasp the cause of abnormality of the rotating machine 2, and can repair the rotating machine 2 before the failure. As a result, it is possible to take countermeasures against failures in advance, and it is possible to contribute to reduction of downtime of infrastructure systems and the like.

<<Example>>
In this example, a detection test of a sign of abnormality was performed using the data set of Non-Patent Document 2. This data set relates to a bearing deterioration test provided by "National Aeronautics and Space Administration (NASA)".

This device under test had a shaft connected to an AC motor, a rotation speed of "2000 (rpm)", and four bearings attached to the shaft. Has been. Further, acceleration sensors are installed as channels 1 to 4 on each of the bearings. Here, the sampling frequency of vibration is "20 kHz", and measurement is performed for 1 second every 10 minutes.

In this embodiment, the data set No. of Non-Patent Document 2 is set. Two sensor channel 1 data were evaluated. At that time, the frequency calculation unit 5 performed a fast Fourier transform (FFT), and 100 samples of data from the start of measurement were used as learning data.

FIG. 4A shows a coefficient matrix W _n obtained by decomposing the learning data by NMF, and FIG. 4B shows a basis matrix H _n obtained by decomposing the learning data. Here, feature extraction of periodic peaks near 1000 Hz and 4000 Hz can be confirmed with the basis matrix H _n .

5A shows the reconstruction error of the data set, the horizontal axis shows the sample number (Sample) in the training data, the vertical axis shows the reconstruction error (Reconstruction error), and FIG. FIG. 6 shows a partially enlarged view in which “reconstruction error=0.0000 to 0.0007” in FIG. 5A is enlarged.

In FIGS. 5A and 5B, the vertical axis “sample numbers 1 to 100” represents the reconstruction error E _n , and the vertical axis “sample numbers 101 and later” represents the reconstruction error E _t . Here, according to FIGS. 5A and 5B, the reconstruction error E _t rapidly increases near the time of failure (sample number 1000).

However, as shown in part P of FIG. 5B, the first threshold “Threshold” is exceeded for the first time at sample number 512. Further, it can be confirmed that the reconstruction error E _t starts to increase from the sample number 531 and always exceeds the first threshold “Threshold” after that, and at this point, the sign is detected before the failure.

In addition, if the reconstruction difference D _t is evaluated, the range of the learning data (sample numbers 1 to 100), the sample number 705 at which the reconstruction error E _t peaks for the first time, and the time of failure (sample number 1000) are 3 Compare with one.

6A to 6C, the respective reconstruction differences to be compared are represented by colors, the horizontal axis represents time [second], and the vertical axis represents frequency [Hz]. At this time, in the sample number 705 shown in FIG. 6B, the color change is large near 4 kHz, and the minus difference increases, and in addition, the harmonic component plus difference can be confirmed. It is possible to suspect abnormalities and bearing abnormalities.

The present invention is not limited to the above embodiment, and the system configuration and the like can be modified and implemented within the scope described in each claim. For example, the equipment to be diagnosed is not limited to the rotating equipment (rotating machine) 2, and if the vibration waveform data can be collected, the abnormality sign can be detected.

Further, the abnormality sign detection system 1 may be configured as a program that causes a computer to function. According to this program, the computer functions as each of the units 4 to 11 and detects the abnormality sign of the target equipment.

1…Abnormal sign detection system 2…Rotary equipment (target equipment)
3... Vibration sensor 4... Data recording unit 5... Frequency calculation unit 6... Normal data recording unit 7... Normal model creation unit 8... Total reconstruction error calculation unit 9... Diagnostic data recording unit 10... Abnormality determination unit 11... Variable error calculation Department (abnormality factor estimation department)

Claims (6)

An abnormality sign detection system for detecting an abnormality sign based on vibration waveform data of target equipment,
A normal model creation unit that creates a normal model by non-negative matrix factorization (NMF) based on the vibration waveform data collected in advance when the target facility is normal;
A vibration error data of the target facility is used as diagnostic data, and a reconstruction error calculation unit that obtains an input/output error distribution when the diagnostic data is input to the normal model,
If the error distribution of the diagnostic data deviates from the error distribution of the normal data, an abnormality determination unit that determines the occurrence of the abnormality sign,
An error factor estimation unit that calculates an error for each frequency of the diagnostic data and estimates an error factor of the target equipment according to the evaluation of the calculated error,
An abnormality sign detection system comprising: Further comprising a frequency calculation unit for calculating a power spectrum from the vibration waveform data,
The normal model creation unit, with respect to the amplitude spectrogram of the power spectrum in a normal time as the normal model,
Calculate non-negative coefficient matrix and non-negative basis matrix as the feature amount using the non-negative matrix factorization,
The abnormality sign detection system according to claim 1, wherein an error distribution of the normal data is obtained based on the calculated non-negative feature amount. The abnormality determination unit,
For the amplitude spectrum gram of the power spectrum in the diagnostic data calculated by the frequency calculation unit,
The coefficient matrix of the diagnostic data is calculated by non-negative matrix factorization while being fixed in the basis matrix of the normal model, and the error distribution of the diagnostic data is obtained by using the calculated coefficient matrix as a weight. 2. The abnormality sign detection system described in 2. The abnormality determination unit determines that the error distribution of the diagnostic data deviates from the error distribution of the normal data when the error distribution of the diagnostic data exceeds a preset threshold value. Anomaly sign detection system described in. The abnormality factor estimation unit,
The difference between the amplitude spectrumgram of the diagnostic data and the multiplication value of the basis matrix and the coefficient matrix in the abnormality determination unit,
The abnormality sign detection system according to claim 3 or 4, wherein if the calculated difference exceeds a preset threshold, it is estimated as the abnormality factor. A method in which a computer detects an abnormality sign based on vibration waveform data of target equipment,
A normal model creating step of creating a normal model by non-negative matrix factorization (NMF) based on the vibration waveform data collected in advance when the target facility is normal;
Reconstruction error calculation step for obtaining an input/output error distribution when the diagnostic equipment is the vibration waveform data of the target equipment and the diagnostic data is input to the normal model,
If the error distribution of the diagnostic data deviates from the error distribution of the normal data, an abnormality determination step of determining the occurrence of the abnormality sign,
An abnormality factor estimating step of identifying the vibration frequency of the determined abnormality sign and estimating an abnormality factor of the target equipment according to the identified vibration frequency,
An abnormality sign detection method comprising:

Images