KR20220089576A - Method for diagnosing the fault of a rotary machine and Device thereof - Google Patents

Abstract

According to the present invention, by installing one or more vibration sensors in the rotating machine and analyzing the output signal thereof, it is possible to determine whether there is an abnormality in the rotating machine, and also performs image conversion on the signal of the vibration sensor and performs unsupervised learning and statistical processing for this. There is an advantage in that it is possible to determine whether there is an abnormality in the rotating machine through the
The rotating machine abnormality diagnosis method of the present invention includes a vibration signal measuring step of measuring a vibration signal through one or more vibration sensors attached to the rotating machine, a two-dimensional image conversion step of converting the vibration signal into a two-dimensional image, and a characteristic factor value in the two-dimensional image It consists of a characteristic factor extraction step of extracting , a defect index calculation step of calculating a defect index based on the extracted characteristic factor value, and an abnormality determination step of determining the presence or absence of an abnormality in the rotating machine through statistical testing of the defect index.

Description

Method for diagnosing the fault of a rotary machine and Device thereof

The present invention relates to a method and apparatus for diagnosing abnormalities in a rotating machine, and more particularly, to a rotating machine abnormality diagnosis method and apparatus capable of determining the presence or absence of an abnormality in a rotating machine based on image conversion for a vibration signal of the rotating machine.

In general, a rotating machine refers to a mechanical device in which a moving part revolves around a rotating shaft, such as an electric motor, a generator, a turbine, and the like.

There are various types of these rotary machines, and they are currently a key power unit that drives major social infrastructures such as production facilities, high-speed railways, and power plants.

In addition, the quality of the rotating machine must be secured in the production stage, and the reliability of the rotating machine in the field operation stage is important enough to influence the availability of the entire facility. There is a problem in that the entire process is stopped without being finished.

Therefore, in an industrial field using a rotating machine, research has been conducted to perform fault and abnormal diagnosis by mounting a sensor inside or outside the rotating machine to perform condition monitoring and preventive maintenance of the rotating machine.

As an example, in Korean Patent Laid-Open Publication No. 10-2012-0072614, a vibration sensor capable of simultaneously measuring X-axis vibration and Y-axis vibration is installed in a diagnostic target motor, and the rotation frequency and fault frequency of the diagnostic target motor are calculated. Thus, a fault diagnosis method using a motor vibration signal for diagnosing a motor defect by comparing the peak value of the vibration signal with a threshold value is disclosed.

However, in this case, since two sensors need to be installed, there is a problem in that the configuration of the entire system becomes complicated.

Korean Patent Laid-Open Publication No. 10-2012-0072614 (2012.07.04)

It is an object of the present invention to provide a method and apparatus for diagnosing abnormalities in a rotating machine by installing one or more vibration sensors in a rotating machine and analyzing an output signal thereof.

In addition, another object of the present invention is to provide a method and apparatus for diagnosing abnormality of a rotating machine by performing image conversion on a signal of a vibration sensor, and determining the presence or absence of abnormality in a rotating machine through unsupervised learning and statistical processing. .

The rotating machine abnormality diagnosis method according to the present invention includes a vibration signal measuring step of measuring a vibration signal through one or more vibration sensors attached to the rotating machine, a two-dimensional image conversion step of converting the vibration signal into a two-dimensional image, and a characteristic factor in the two-dimensional image It may include a characteristic factor extraction step of extracting a value, a defect index calculation step of calculating a defect index based on the extracted characteristic factor value, and an abnormality determination step of determining the presence or absence of an abnormality in the rotating machine through statistical testing of the defect index. .

Here, in the feature factor extraction step, the feature factor may be extracted from the two-dimensional image by learning to reduce the difference between the input image and the output image through encoding and decoding.

In addition, extracting the value of the characteristic factor may be performed by designing and training a convolutional neural network (CNN) based on unsupervised learning.

In addition, in the defect indicator calculation step, a defect indicator value may be calculated for the characteristic factor value based on a distance scale.

Meanwhile, the distance scale may be performed based on a Mahalanobis Distance scale.

In addition, in the abnormality determination step, it is possible to determine whether the rotating machine is abnormal based on a Sequential Probability Ratio Test (SPRT) algorithm for the defect index.

A rotating machine abnormality diagnosis apparatus according to another embodiment of the present invention includes a vibration signal measuring unit that measures a vibration signal through one or more vibration sensors attached to the rotating machine, a two-dimensional image conversion unit that converts the vibration signal into a two-dimensional image, and a two-dimensional It includes a characteristic factor extraction unit that extracts the characteristic factor value from the image, a defect indicator calculation unit that calculates a defect index based on the extracted characteristic factor value, and an abnormality determination unit that determines whether the rotating machine is abnormal through statistical testing of the defect index can do.

Here, the two-dimensional image converter may convert the vibration signal into a two-dimensional image having a time component and a frequency component as two axes based on a Wigner-Ville distribution function.

Also, the characteristic factor extractor may extract the characteristic factor by learning to reduce a difference between an input image and an output image through encoding and decoding from the two-dimensional image.

In addition, extracting the value of the characteristic factor may be performed by designing and training a convolutional neural network (CNN) based on unsupervised learning.

Meanwhile, the defect indicator calculation unit may calculate a defect indicator value based on a distance scale for the characteristic factor value.

Also, the distance scale may be performed based on a Mahalanobis Distance scale.

In addition, the abnormality determination unit may determine whether the rotating machine is abnormal based on a sequential probability ratio test (SPRT) algorithm for the defect index.

The rotating machine abnormality diagnosis method and apparatus according to the present invention has an advantage in that it is possible to determine whether the rotating machine is abnormal by installing one or more vibration sensors on the rotating machine and analyzing the output signal thereof.

In addition, the method and apparatus for diagnosing abnormality of a rotating machine according to the present invention have the advantage of performing image conversion on the signal of the vibration sensor, and determining whether there is an abnormality of the rotating machine through unsupervised learning and statistical processing.

1 is a flowchart illustrating a method for diagnosing abnormality in a rotating machine according to an embodiment of the present invention.
FIG. 2 is a diagram illustrating the two-dimensional image conversion step of FIG. 1 in detail.
FIG. 3 is a view showing in detail the step of extracting the characteristic factor of FIG. 1 .
FIG. 4 is a view showing in detail the step of calculating the defect index of FIG. 1 .
5 is a diagram illustrating in detail the step of determining whether there is an abnormality in FIG. 1 .
6 is a block diagram illustrating an apparatus for diagnosing abnormality of a rotating machine according to an embodiment of the present invention.

Detailed embodiments for carrying out the present invention will be described with reference to the accompanying drawings.

Since the present invention can have various changes and can have various embodiments, specific embodiments are illustrated in the drawings and described in detail in the detailed description. This is not intended to limit the present invention to specific embodiments, it can be understood to include all modifications, equivalents and substitutes included in the spirit and scope of the present invention.

Hereinafter, a method and apparatus for diagnosing abnormalities in a rotating machine according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a flowchart illustrating a method for diagnosing abnormalities in a rotating machine according to an embodiment of the present invention, and FIGS. 2 to 5 are detailed views for explaining FIG. 1 in detail.

Hereinafter, a method for diagnosing abnormalities in a rotating machine according to an embodiment of the present invention will be described with reference to FIGS. 1 to 5 .

First, referring to Figure 1, the rotating machine abnormality diagnosis method according to an embodiment of the present invention comprises the steps of measuring a vibration signal through one or more vibration sensors attached to the rotating machine (S100), and converting the vibration signal into a two-dimensional image step (S200), extracting a characteristic factor value from the two-dimensional image (S300), calculating a defect index based on the extracted characteristic factor value (S400), and the presence or absence of abnormality in the rotating machine through statistical testing of the defect index It is made of a step (S500) of determining.

That is, the present invention is to analyze the vibration signal generated from the rotating machine to determine whether there is an abnormality in the rotating machine, and the vibration caused by a change in the physical structure and mechanical motion of the bearing and shaft alignment in the rotating machine is measured through the vibration sensor. Then, it is determined whether there is an abnormality in the rotating machine through time-frequency analysis of the measured vibration data.

First, in the vibration signal measuring step (S100), a vibration signal for each time point is measured for an input load or an output load applied to the rotating body of the rotating machine.

Thereafter, in the two-dimensional image transformation step (S200), the acquired time series data is converted into an image form, and in the feature factor extraction step (S300), the characteristic factor is extracted through unsupervised learning.

Next, in the defect index calculation step (S400), the defect index is calculated according to the relative distance of the distribution of the characteristic factors, and finally, in the abnormality determination step (S500), the rotating body defect classification and severity are derived by using a statistical method determine whether there is an abnormality in

The present invention not only does not require design parameters of the rotating machine, but has the advantage of performing image conversion into one or more vibration sensor signals. Also, by using a time component, not only stationary but also non-stationary ) has the advantage of being applicable to time series data.

FIG. 2 is a view showing the two-dimensional image conversion step S200 of FIG. 1 in detail. FIG. 2 (a) is a graph showing a time axis signal and a frequency axis signal of a vibration signal, respectively, and FIG. 2(b) is a time of the vibration signal A three-dimensional graph having components, frequency components, and amplitude components as axes, FIG. 2C shows a two-dimensional graph having time components and frequency components of a vibration signal as axes.

As can be seen in FIG. 2 , in the two-dimensional image conversion step (S200), the vibration signal is converted into a two-dimensional image having two axes, a time component and a frequency component, based on a Wigner-Ville distribution function. do.

At this time, the vibration signal is input as a time component V210, and the time component V210 is transformed into a frequency component V220 by a Fourier transform.

That is, in this step, a two-dimensional image (V240) with the time component (V210) and the frequency component (V220) as two axes is generated using a Wigner-Ville distribution function for the vibration signal. .

As another embodiment, a three-dimensional image signal V230 with three axes including an amplitude component thereof along with a time component V210 and a frequency component V220 of the vibration signal may be generated.

As described above, in the present invention, a two-dimensional image can be generated using one or more vibration sensors, thereby simplifying the system.

FIG. 3 is a view showing the characteristic factor extraction step S300 of FIG. 1 in detail.

As can be seen from FIG. 3 , in the feature factor extraction step ( S300 ), the feature factor can be extracted from the two-dimensional image by learning to reduce the difference between the input image and the output image through encoding and decoding, and the feature factor value Extraction can be performed by designing and training a convolutional neural network (CNN) based on unsupervised learning.

Here, unsupervised learning is a method of finding out how data is structured. Unlike supervised learning or reinforcement learning, unsupervised learning is not given a target for an input value, and autonomous learning is performed. It has the advantage of being able to summarize and explain the main characteristics of the data.

3 is about a convolutional auto-encoder among unsupervised learning, where the number of training is 50 epochs, the learning rate is 0.0001, the batch size is 120, the active function is ReLU, and the optimization algorithm is Adam using The process of extracting features from a 500x300 two-dimensional image is shown.

As shown in Figure 3, the 500x300 two-dimensional image input in the two-dimensional image input step (S310) is a first convolution (convolution) and pooling (Pooling) step (S320), the second convolution and pooling step (S330) ), and by repeating convolution and pooling through the third convolution step (S340), a low-dimensional characteristic is sequentially derived from a high-dimensional characteristic, and then finally in the feature extraction step (S350) 15 features with characteristics are output and used as the extracted characteristic factor values.

As described above, in the present invention, the characteristic factor can be derived without inputting the characteristic parameter of the rotating machine by deriving the characteristic factor value by autonomous learning rather than arbitrary supervised learning.

FIG. 4 is a detailed view showing the defect index calculation step S400 of FIG. 1 .

As can be seen in FIG. 4 , in the defect index calculation step S400 , the defect index value can be calculated based on the distance scale for the characteristic factor value extracted in the characteristic factor extraction step S300 , and the distance scale is Mahala This can be done based on the Mahalanobis Distance scale.

Here, the Mahalanobis distance is a distance between two populations having different means, and is different from the Euclidean distance, which means a distance between two points.

That is, as can be seen from FIG. 4 , in the first population V401 and the second population V402, the first and second parameters having different characteristics, the average (V403) of the first population (V401) and the second Looking at the distance from the average (V404) of the second population (V402) to an arbitrary point (V405) belonging to the first population (V401), the Euclidean distance is the average (V403) of the first population (V401) ) is greater than the distance D420 from the mean V404 of the second population V402. On the other hand, Mahalanobis Distance is a characteristic that the distance (D440) from the average (V403) of the first population (V401) is smaller than the distance (D450) from the average (V404) of the second population (V402). have.

Therefore, when belonging to a population having different means, the Euclidean distance may be short, but the Mahalanobis distance may be large.

As described above, in the present invention, a plurality of defect index values can be obtained by identifying the characteristics by separating the population for the characteristic factor values.

FIG. 5 is a detailed view showing an abnormality determination step ( S500 ) of FIG. 1 .

As can be seen in FIG. 5 , in the abnormality determination step S500 , it is determined whether the rotating machine is abnormal based on a Sequential Probability Ratio Test (SPRT) algorithm for the defect index.

Here, the Sequential Probability Ratio Test (SPRT) algorithm is a statistical hypothesis testing method that diagnoses through a ratio between probabilities assuming that the measured values are normal or abnormal.

For example, when the first error distribution P510 and the second error distribution P520 are obtained for the defect index value, the error probability P540 for the first error distribution P510 for the arbitrary value V501. ) and the error probability P530 for the second error distribution P520 can be extracted, and the ratio of these two error probabilities P530 and P540 is defined to be normal when smaller than a predetermined diagnostic criterion and abnormal when larger than a predetermined diagnostic criterion. , it is possible to determine whether there is an abnormality in the rotating machine by setting the judgment pending in the case of a value in between.

As described above, in the present invention, it is possible to autonomously determine the presence or absence of an abnormality in the rotating machine based on abnormality diagnosis through a statistical test without defining parameters or any abnormality.

6 is a block diagram illustrating an apparatus for diagnosing abnormality of a rotating machine according to an embodiment of the present invention.

As can be seen in FIG. 6 , the apparatus for diagnosing abnormality in the rotating machine of the present invention includes a vibration signal measuring unit 100 that measures a vibration signal through one or more vibration sensors attached to the rotating machine, and converts the vibration signal into a two-dimensional image. A two-dimensional image conversion unit 200, a characteristic factor extraction unit 300 for extracting a characteristic factor value from a two-dimensional image, a defect index calculation unit 400 for calculating a defect index based on the extracted characteristic factor value, and a defect index It consists of an abnormality determination unit 500 that determines whether there is an abnormality in the rotating machine through a statistical test of .

Here, the two-dimensional image conversion unit 200 converts the vibration signal into a two-dimensional image having a time component and a frequency component as two axes based on a Wigner-Ville distribution function.

In addition, the characteristic factor extraction unit 300 may extract the characteristic factor by learning to reduce the difference between the input image and the output image through encoding and decoding from the two-dimensional image, and extracting the characteristic factor value is unsupervised. Based on unsupervised learning, a convolutional neural network (CNN) is designed, trained, and performed.

The defect indicator calculation unit 400 may calculate a defect indicator value based on a distance scale for the extracted characteristic factor value, and the distance scale may be performed based on a Mahalanobis Distance scale.

Thereafter, the abnormality determination unit 500 determines whether the rotating machine is abnormal based on a Sequential Probability Ratio Test (SPRT) algorithm for the defect index.

As such, the present invention does not require design parameters of the rotating machine, and has the advantage of performing image conversion into one or more vibration sensor signals. It has the advantage of being applicable to stationary time series data.

As described above, in the rotating machine abnormality diagnosis method and apparatus according to the present invention, by installing one or more vibration sensors in the rotating machine and analyzing the output signal thereof, it is possible to determine whether the rotating machine is abnormal, and also converts the image of the vibration sensor signal It has the advantage of being able to determine whether there is an abnormality in the rotating machine through unsupervised learning and statistical processing.

What has been described above includes examples of one or more embodiments. Of course, it is not possible to describe every possible combination of components or methods for purposes of describing the above-described embodiments, and those skilled in the art will recognize that many further combinations and permutations of various embodiments are possible. Accordingly, the described embodiments are intended to cover all alternatives, modifications and adaptations falling within the spirit and scope of the appended claims.

Claims (14)

A vibration signal measuring step of measuring a vibration signal through one or more vibration sensors attached to the rotator;
a two-dimensional image conversion step of converting the vibration signal into a two-dimensional image;
a feature factor extraction step of extracting a feature factor value from the two-dimensional image;
a defect index calculation step of calculating a defect index based on the extracted characteristic factor value; and
Rotating machine abnormality diagnosis method including; The method of claim 1,
In the two-dimensional image conversion step, based on a Wigner-Ville distribution function, the vibration signal is converted into a two-dimensional image having a time component and a frequency component as two axes. . The method of claim 1,
In the step of extracting the characteristic factor, the characteristic factor is extracted from the two-dimensional image by learning to reduce the difference between the input image and the output image through encoding and decoding. 4. The method of claim 3,
Extracting the characteristic factor value is a rotating machine abnormality diagnosis method, characterized in that it is performed by designing and training a convolutional neural network (CNN) based on unsupervised learning. The method of claim 1,
In the defect indicator calculation step, the defect indicator value is calculated based on a distance scale with respect to the characteristic factor value. 6. The method of claim 5,
The distance scale is a rotating machine abnormality diagnosis method, characterized in that it is performed based on the Mahalanobis Distance scale. The method of claim 1,
In the abnormality determination step, the rotating machine abnormality diagnosis method, characterized in that for determining the abnormality of the rotating machine based on a Sequential Probability Ratio Test (SPRT) algorithm for the defect index. a vibration signal measuring unit for measuring a vibration signal through one or more vibration sensors attached to the rotator;
a two-dimensional image converter converting the vibration signal into a two-dimensional image;
a characteristic factor extraction unit for extracting a characteristic factor value from the two-dimensional image; and
a defect indicator calculation unit for calculating a defect indicator based on the extracted characteristic factor value; and
Rotating machine abnormality diagnosis apparatus including; 9. The method of claim 8,
The two-dimensional image converter, based on a Wigner-Ville distribution function, converts the vibration signal into a two-dimensional image having a time component and a frequency component as two axes. 9. The method of claim 8,
and the characteristic factor extracting unit extracts the characteristic factor from the two-dimensional image by learning to reduce a difference between an input image and an output image through encoding and decoding. 11. The method of claim 10,
Extracting the characteristic factor value is a rotating machine abnormality diagnosis apparatus, characterized in that it is performed by designing and training a convolutional neural network (CNN) based on unsupervised learning. 9. The method of claim 8,
The defect indicator calculator is configured to calculate a defect indicator value based on a distance scale with respect to the characteristic factor value. 13. The method of claim 12,
The distance scale, Mahalanobis distance (Mahalanobis Distance) Rotator abnormality diagnosis apparatus, characterized in that performed based on the scale. 9. The method of claim 8,
The abnormality determination unit, Rotator abnormality diagnosis apparatus, characterized in that for determining the abnormality of the rotating machine based on a sequential probability ratio test (SPRT) algorithm with respect to the defect indicator.

Images