KR102327420B1 - Apparatus and method for generating and verifying training data - Google Patents

Abstract

The present invention relates to an apparatus and method for generating and verifying training data, capable of generating augmented data by using an auto-encoder and verifying the augmented data to determine whether to verify the augmented data as the training data. According to the present embodiment, the method for generating and verifying the training data may include the steps of: collecting first and second vibration signals from first and second devices; extracting first and second static feature values from the first and second vibration signals; performing comparison between the first static feature value and the second static feature value with respect to the same static items of the same axis to determine an axis and a static item having a difference of a threshold value or more; preparing for an abnormal data set labeled as being in an abnormal state and a normal data set labeled as being in a normal state, based on a first static feature value and a second static feature value of the determined axis and the determined static item; generating a defect detecting model by training the deep neural network by using, as the training data, a first abnormal data group of the labeled abnormal data set and the labeled normal data set; generating an augmented abnormal data group by applying at least a portion of the labeled abnormal data set to the auto encoder; verifying an accuracy of the defect detecting model by changing the ratio between a second abnormal data group of the labeled abnormal data set and the augmented abnormal data group; and determining whether to verify the training data for the augmented abnormal data group, based on a result obtained by verifying the accuracy of the defect detecting model.

Description

Apparatus and method for generating and validating training data

The present invention relates to an apparatus and method for generating and verifying training data for determining whether to authenticate as training data of a defect detection model by generating augmented data using an auto encoder and verifying it.

With the recent development of industrial technology, as rotary machines become larger and more precise, failures in these facilities can cause enormous economic loss or damage to human life, so fault monitoring of these rotary machines has become very important.

Since the machine including these rotating driving parts is driven under physical friction conditions, it is necessary to perform maintenance at frequent intervals to prevent breakdowns and accidents. As such, the rotating machine requires frequent maintenance, so there is a problem that maintenance cost and time are required.

Due to these problems, condition-based maintenance, which performs maintenance only when there is a possibility of a failure, has recently attracted attention. Because condition-based maintenance performs maintenance only when there is a possibility of a failure, it reduces operating costs and increases profits by reducing maintenance costs and time.

Accordingly, Korean Patent Application Laid-Open No. 2020-0101507 also disclosed a system for diagnosing and predicting a malfunction of a rotating machine using machine learning.

The fields where fault diagnosis can be applied to have great effects can be systems that require high maintenance costs, and gearboxes such as high-speed railways and wind turbines are representative.

Since the gearbox is a basic element used to transmit rotational power in a machine, research on detecting and diagnosing defects through various sensor signals is being actively conducted.

The above-mentioned background art is technical information possessed by the inventor for derivation of the present invention or acquired in the process of derivation of the present invention, and cannot necessarily be said to be a known technique disclosed to the general public prior to the filing of the present invention.

Domestic Patent Publication No. 10-2020-0101507 (2020.08.28)

An object of the present invention is to improve the detection accuracy of a defect detection model by generating augmented data using an autoencoder, verifying it, and authenticating it as training data of the defect detection model.

An object of the present invention is to improve the performance of a machine fault detection training process by increasing training data using an autoencoder.

An object of the present invention is to solve a problem in which computational resources are wasted by including unnecessary information in generating a model for detecting a device defect using statistical analysis and machine learning.

The problem to be solved by the present invention is not limited to the above-mentioned problems, and other problems and advantages of the present invention not mentioned can be understood by the following description, and more clearly understood by the embodiments of the present invention will be In addition, it will be appreciated that the problems and advantages to be solved by the present invention can be realized by means and combinations thereof indicated in the claims.

In a method for generating and verifying training data according to an embodiment of the present invention, the first and second devices are respectively attached to first devices determined to be defective and second devices determined to be free from defects. collecting first vibration signals of first devices and second vibration signals of second devices by vibration sensors arranged to sense vibration of at least three orthogonal axes during operation; extracting first statistical characteristic values with respect to the vibration signals of the axis of , and extracting second statistical characteristic values of the vibration signals of each axis from the second vibration signals; Comparing the characteristic values and the second statistical characteristic values to determine an axis and a statistical item having a difference greater than or equal to a threshold value; Preparing a normal data set in which the determined second statistical characteristic values of the determined axis and statistical items are in a steady state, and using the first abnormal data group and the labeled normal data set among the labeled abnormal data sets as training data Training a neural network to generate a defect detection model that determines whether a device is defective using statistical characteristic values of vibration signals generated by the device, and at least a part of the labeled abnormal data set to an auto encoder generating an augmented anomalous data group by applying it; and verifying the accuracy of the defect detection model by changing the ratio of the second abnormal data group and the augmented abnormal data group among the labeled abnormal data set; It may include determining whether to authenticate the training data for the augmented abnormal data group based on the verified result.

An apparatus for generating and verifying training data according to an embodiment of the present invention includes a processor and a memory operatively connected to the processor and storing at least one code executed by the processor, wherein the memory is executed by the processor , a processor attached to the first devices determined to be defective and the second devices determined to be free, respectively, arranged to sense vibration in at least three orthogonal axes while the first and second devices operate Collecting the first vibration signals of the first devices and the second vibration signals of the second devices by the sensors, extracting first statistical characteristic values for the vibration signals of each axis from the first vibration signals, and An axis having a difference greater than or equal to a threshold value by extracting second statistical characteristic values for vibration signals of each axis from the two vibration signals, and comparing the first statistical characteristic values and the second statistical characteristic values for the same statistical items on the same axis and determining a statistical item, and an abnormal data set in which the determined first statistical characteristic values of the determined axis and the statistical item are labeled as an abnormal state, and a normal data set in which the determined second statistical characteristic values of the determined axis and the statistical item are labeled as a normal state. Prepare and train a deep neural network using the first abnormal data group and the labeled normal data set among the labeled abnormal data sets as training data. Create a defect detection model to judge, apply at least a part of the labeled anomaly data set to an auto encoder to generate an augmented anomaly data group, and a second anomaly data group and augmented anomaly data from the labeled anomaly data set It is possible to store a code causing to verify the accuracy of the defect detection model by changing the ratio of the groups, and to determine whether to authenticate the training data for the augmented abnormal data group based on the result of verifying the accuracy of the defect detection model.

In addition to this, another method for implementing the present invention, another system, and a computer-readable recording medium storing a computer program for executing the method may be further provided.

Other aspects, features and advantages other than those described above will become apparent from the following drawings, claims, and detailed description of the invention.

According to the present invention, it is possible to improve the detection accuracy of the defect detection model by generating augmented data using an autoencoder, verifying it, and authenticating it as training data of the defect detection model.

In addition, by improving the detection accuracy of the defect detection model, it is possible to predict machine defects in advance and reduce the cost of detecting machine defects.

In addition, the autoencoder can be used to augment the training data to improve the performance of the machine fault detection training process.

In addition, by efficiently preparing more important and normal data as training data for generating a defect detection model using statistical analysis, it is possible to generate a defect detection model by using computational resources more effectively.

Effects of the present invention are not limited to those mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a block diagram schematically illustrating the configuration of an apparatus for generating and verifying training data according to the present embodiment.
2 is an exemplary view for explaining an output direction of a vibration signal output by the vibration sensor according to the present embodiment.
3 is an exemplary diagram illustrating a vibration signal output from a device in a normal state and a failure state according to the present embodiment.
4 is a diagram illustrating a result of preprocessing a vibration signal output from a device in a normal state and a failure state according to the present embodiment.
5 is a diagram illustrating a result of calculating a statistical characteristic value from a preprocessing result for a vibration signal output from a device in a normal state and a failure state according to the present embodiment.
FIG. 6 is a block diagram schematically illustrating the configuration of the generator in FIG. 1 .
7 is a block diagram schematically illustrating a configuration of an apparatus for generating and verifying training data according to another embodiment.
8 is a flowchart illustrating a method of generating and verifying training data according to the present embodiment.

Advantages and features of the present invention, and methods for achieving them will become apparent with reference to the detailed description in conjunction with the accompanying drawings. However, it should be understood that the present invention is not limited to the embodiments presented below, but may be implemented in a variety of different forms, and includes all transformations, equivalents, and substitutes included in the spirit and scope of the present invention. . The embodiments presented below are provided to complete the disclosure of the present invention, and to completely inform those of ordinary skill in the art to which the present invention pertains to the scope of the invention. In describing the present invention, if it is determined that a detailed description of a related known technology may obscure the gist of the present invention, the detailed description thereof will be omitted.

The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise. In the present application, terms such as “comprise” or “have” are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, but one or more other features It should be understood that this does not preclude the existence or addition of numbers, steps, operations, components, parts, or combinations thereof. Terms such as first, second, etc. may be used to describe various elements, but the elements should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another.

In the present application, “unit” may be a hardware component such as a processor or circuit, and/or a software component executed by a hardware component such as a processor.

Hereinafter, embodiments according to the present invention will be described in detail with reference to the accompanying drawings, and in the description with reference to the accompanying drawings, the same or corresponding components are given the same reference numerals, and overlapping descriptions thereof are omitted. decide to do

1 is a block diagram schematically illustrating the configuration of an apparatus for generating and verifying training data according to the present embodiment.

Referring to FIG. 1 , an apparatus 200 for generating and verifying training data may be connected to devices 100 that generate vibration, such as a rotating machine.

The devices 100 may include first devices 110 and second devices 120 , wherein the first devices 110 are defective and faulty devices, and the second devices 120 . ) are normal devices without defects. Here, the first devices 110 and the second devices 120 may be mechanical devices of the same model. For example, the first devices 110 may be device A, device B, and device C with faults, and the second devices 120 may be device A', device B', and device C' without faults. have.

These devices may include a gearbox used to transmit rotational power to a mechanical system. The gearbox may include a structure in which various shaft groups and gear groups (not shown) are disposed in order to transmit power required to rotate the drive shaft. Such a gearbox can be used in a wide variety of industries. And by using the gears of the gear group, a small rotational force can be converted into a large rotational force, or a low rotational force can be converted into a high-speed rotation. In the gearbox, a housing (not shown) of a certain shape forms the exterior. The housing may be formed to have a predetermined size and shape according to various mechanical devices or vehicles in which the gearbox is used. In addition, various gears (not shown), a driving shaft (not shown), a rotating shaft (not shown), etc. may be disposed inside the housing to transmit a driving motor (not shown) and driving force for generating driving force.

On the other hand, the gearbox generates various vibrations in the gearbox due to unstable movement in the process of generating and transmitting the driving force using the driving motor, and these vibrations may be transmitted to the drive shaft and the housing forming the gearbox. In addition, when the shaft vibration of the drive shaft is transmitted to the housing, in addition to the vibration of the housing itself, vibration of other components connected to the housing may be caused, and noise due to the vibration may also be induced.

In the past, experts analyzed the change in vibration amount and vibration characteristics of the gearbox to find the cause of the defect and set up appropriate countermeasures. However, due to the lack of experience of experts and changes in diversified facility characteristics, it is difficult to make an accurate diagnosis, so the method of setting the size of vibration and notifying the manager of an abnormal condition only when vibration exceeding the size occurs has been used.

However, this method can only determine the degree of failure of a device such as a rotating machine, and since the cause or progress of the failure cannot be known, it cannot be actively dealt with, so it is left to experts again. Therefore, when a defect occurs in a device such as a rotating machine, a technology is needed to prevent an accident in advance by identifying the cause of the defect in advance and predicting the progress at the stage before the failure. Accurate diagnosis is required.

Recently, with the development of machine learning technology, a technology capable of diagnosing using a deep neural network model is being used. However, in order to generate such a defect detection model, it is necessary to prepare a large amount of training data, and a considerable amount of computational resources and training period are required to train the model with such a large amount of training data.

In this embodiment, in order to efficiently and effectively generate a defect detection model, information useful for determining a defect may be determined, and a defect detection model may be generated using the useful information. According to the method of this embodiment, it is possible to generate an effective defect detection model with much less computational resources and time to generate the defect detection model.

The apparatus 200 for generating and verifying training data may generate a model capable of detecting whether a device is defective by performing statistical analysis and machine learning by receiving vibration signals generated from the devices 110 and 120 as input. . Here, whether the device is defective may be determined depending on whether the device is operating normally or when the device is broken. In this embodiment, the apparatus 200 for generating and verifying training data includes a preprocessor 210 , a calculator 220 , a first determiner 230 , a loading unit 240 , and a first generator 250 . , a second generation unit 260 , a verification unit 270 , a second determination unit 280 , and a control unit 290 .

Each of the devices 100 is provided with a vibration sensor (not shown) to detect and output a vibration signal generated by the device 100 . In this embodiment, the vibration sensor may include a first vibration sensor (not shown) to a third vibration sensor (not shown) arranged to sense vibration of at least three orthogonal axes. In addition, the vibration sensor may include a fourth vibration sensor (not shown) arranged to sense vibration of a diagonal axis defined by three orthogonal axes in addition to vibration signals of three axes.

Here, in the apparatus 200 for generating and verifying training data, a signal coming from the first devices 110 and a signal coming from the second devices 120 can be distinguished, and the first devices 110 are defective. It is known that the device exists and that the second devices 120 are defect-free devices.

The preprocessor 210 may convert the vibration signals in the time domain into the frequency domain. The calculator 220 may calculate statistical characteristic values for the vibration signals in the frequency domain. The first determining unit 230 compares the vibration signal from the defective device with the vibration signal from the non-defective device, and among these vibration signals, the statistical value of the vibration signal of the axis is greater than the threshold value (axis) and statistical items).

In the loading unit 240 , when the value of the selected statistical item of the vibration signal of the axis determined by the first determining unit 230 is labeled according to whether the device is defective, a training data set may be prepared. The first generator 250 may generate a defect detection model for detecting whether a device is defective by training a deep neural network using the prepared training data set.

The second generator 260 may generate an augmented abnormal data group by applying a part of the training data set to the autoencoder. The verification unit 270 may verify the accuracy of the defect detection model using another part of the training data set and the abnormal augmented data group. The second determiner may determine whether to authenticate the training data for the augmented abnormal data group based on the verified accuracy.

According to this training method, it is possible to train the model in a shorter time with less computational resources than when training the neural network model using the values of the vibration signals of all axes, and to detect machine faults by increasing the training data using an autoencoder. It can improve the performance of the training process.

FIG. 2 is an exemplary view illustrating an output direction of a vibration signal output by the vibration sensor of FIG. 1 . Referring to FIG. 2 , the first vibration sensor may output a first vibration signal (g_x vibration signal) by detecting vibration in the horizontal (x) direction generated in the device. The second vibration sensor may output a second vibration signal (g_y vibration signal) by detecting vibration in the vertical (y) direction generated in the device. The third vibration sensor may output a third vibration signal (g_z vibration signal) by detecting vibration in the axis (z) direction generated in the device. The fourth vibration sensor detects the vibration in the diagonal (T) direction, which is a combination of the horizontal (x) direction, the vertical (y) direction, and the axial (z) direction generated in the device to output a fourth vibration signal (g_T vibration signal). can

3 is an exemplary diagram illustrating a vibration signal output from a device in a normal state and a failure state according to the present embodiment. Referring to FIG. 3 , FIG. 3A shows a first vibration signal (g_x vibration signal) generated in the device 100 in a normal state, and FIG. 3B is in the devices 100 in a broken state. The generated first vibration signal (g_x vibration signal) is shown. In particular, FIG. 3B illustrates a first vibration signal generated in the device 100 when a tooth is broken.

The preprocessor 210 may convert a time domain vibration signal output from the vibration sensor into a frequency domain vibration signal. The vibration signal in the frequency domain includes more unique information that can be used to distinguish between normality and failure of the device 100 . In this embodiment, the preprocessor 210 performs fast Fourier transform (FFT) on the first to fourth vibration signals of the time domain output from the first vibration sensor to the fourth vibration sensor to generate the frequency domain. The first vibration signal to the fourth vibration signal may be output. The Fourier transform of a time function is a complex-valued function of frequency, the magnitude (absolute value) indicates the amount of frequency present in the original function, and the argument may be an offset of the fundamental sine wave at that frequency.

In this embodiment, the preprocessor 210 may include a first preprocessor 211 to a fourth preprocessor 214 . The first preprocessor 211 may convert a first vibration signal (g_x vibration signal) in the time domain output from the first vibration sensor into a first vibration signal (g_x FFT signal) in the frequency domain. The second preprocessor 212 may convert a second vibration signal (g_y vibration signal) of the time domain output from the second vibration sensor into a second vibration signal (g_y FFT signal) of the frequency domain. The third preprocessor 213 may convert a third vibration signal (g_z vibration signal) of the time domain output from the third vibration sensor into a third vibration signal (g_z FFT signal) of the frequency domain. The fourth preprocessing unit 214 may convert a fourth vibration signal (g_T vibration signal) of the time domain output from the fourth vibration sensor into a fourth vibration signal (g_T FFT signal) of the frequency domain.

4 is a diagram illustrating a result of preprocessing a vibration signal output from a device in a normal state and a failure state according to the present embodiment. Referring to FIG. 4 , FIG. 4A shows that a first vibration signal (g_x vibration signal) in the time domain generated in the device 100 in a normal state is preprocessed and output as a first vibration signal (g_x FFT signal) in the frequency domain. An example is shown, and FIG. 4b is a first vibration signal (g_x FFT signal) in the frequency domain by pre-processing the first vibration signal (g_x vibration signal) in the time domain generated in the devices 100 in the broken state. An example output is shown.

The calculator 220 may calculate statistical characteristic values for the vibration signal of the frequency domain output from the preprocessor 210 . The frequency domain vibration signal output from the preprocessor 210 can be directly input into the deep neural network model to be described later, but since it requires a complex structure and a long training time when working with high-dimensional input data, training the deep neural network model In order to reduce the input data for this purpose, a statistical characteristic value may be calculated. For the model trained by the method of the present embodiment, the same type of input value may be input in a subsequent application step.

Meanwhile, the preprocessor 210 may be configured to evaluate and selectively collect first to fourth vibration signals received from the first to fourth vibration sensors.

Among the received vibration signals, there may be vibration signals that are not suitable for evaluating the state of the machine due to noise or other external environment changes, internal processing errors, and the like. Therefore, it may be necessary to evaluate the suitability of the vibration signal and select and collect only the appropriate signal.

For example, the evaluation of the vibration signals is performed by comparing a combination of first to third vibration signals collected from the first to third vibration sensors with a fourth vibration signal collected from the fourth vibration sensor to determine the second vibration signal. It is performed by evaluating whether the first to fourth vibration signals are abnormal, and such selective collection may be performed when the first to fourth vibration signals are normal based on the evaluation.

For example, the first to third axes are x, y, and z axes orthogonal to each other. If the fourth axis is a diagonal axis defined by the x, y, and z axes, the square value of the magnitude of the vibration signal (first vibration signal) of the first axis in a set of vibration signals (that is, from one device) and the square of the magnitude of the vibration signal (second vibration signal) of the second axis and the square of the magnitude of the vibration signal (third vibration signal) of the third axis is the diagonal of the square of the magnitude of the vibration signal of the fourth axis should be equal to the value.

If, in one set of vibration signals, the sum of the square value of the magnitude of the first vibration signal, the square value of the magnitude of the second vibration signal, and the square value of the magnitude of the third vibration signal is the magnitude of the fourth vibration signal If there is a difference of more than a certain degree from the square value of , it can be seen that these sets of vibration signals are not accurately observed values. Accordingly, the processor ( 201 in FIG. 7 ) of the device 200 that generates and verifies the training data may not include the vibration signals in this case in the collection.

In this embodiment, the calculator 220 calculates the average value, the median value, the minimum value, the maximum value, the kurtosis value, and the skewness of the vibration signal of the frequency domain. A statistical characteristic value including at least one of a value, a standard deviation value, and a peak-to-peak value may be calculated. This is because using these statistical values rather than the vibration signal itself reflects the characteristics of the vibration signal while training the model with smaller size data.

Here, the kurtosis is an index indicating the degree of sharpness of the probability distribution, and the kurtosis may increase as the distribution of signal values gathers near a specific value to form a sharp shape. In addition, skewness is an indicator indicating the asymmetry of the probability distribution, and as the signal bias (the degree to which the distribution of signal values shifts to one side with respect to the average of the signal) increases, the skewness may also increase. The positive amplitude (peak-to-peak) is an index indicating the overall width of the signal and may indicate the difference between the smallest value and the largest value in the signal.

In this embodiment, the calculator 220 may include a first calculator 221 to a fourth calculator 224 . The first calculator 221 may calculate the above-described eight statistical characteristic values for the first vibration signal g_x FFT signal of the frequency domain output from the first preprocessor 211 . The second calculator 222 may calculate the above-described eight statistical characteristic values for the second vibration signal g_y FFT signal of the frequency domain output from the second preprocessor 212 . The third calculator 223 may calculate the above-described eight statistical characteristic values with respect to the third vibration signal g_z FFT signal of the frequency domain output from the third preprocessor 213 . The fourth calculation unit 224 may calculate the above-described eight statistical characteristic values for the fourth vibration signal g_T FFT signal of the frequency domain output from the fourth preprocessor 214 .

The first determination unit 230 compares the calculation results of the statistical characteristic values output from the first calculation unit 221 to the fourth calculation unit 224 , and has the largest value when the device 100 has or does not have a defect. It is possible to judge the data showing the difference (the statistical value of which statistical item of which axis vibration signal).

In the above-described example, since the first calculation unit 221 to the fourth calculation unit 224 each calculate 8 statistical characteristic values, the first determiner 230 receives 8 statistical characteristic values for each of the 4 axes. Thus, a total of 32 statistical characteristic values can be received per device.

The first determiner 230 may determine an axis and a statistical item having a difference greater than or equal to a threshold value by comparing the first statistical characteristic values and the second statistical characteristic values for the same statistical item on the same axis.

For example, the first determination unit 230 calculates each of the eight statistical characteristic values of the vibration signal of the x-axis from the devices A, B, and C of the vibration signal of the x-axis from the devices A', B', and C'. It can be compared with each of the statistical characteristic values.

That is, the first determiner 230 determines the range of kurtosis values of the vibration signal on the x-axis from devices A, B, and C, which are defective devices, and x from devices A', B', and C', which are devices without defects. The range of kurtosis values of the shaft vibration signal can be compared.

The first determination unit 230 may repeat the above comparison for all statistical items of the vibration signals of all axes. Through this iteration, the gap between the ranges of statistical values of the devices with and without defects being compared (if the ranges are far apart), or the size of the overlapping ranges (if the ranges overlap with each other) can be calculated. can

The first determining unit 230 determines the axes and statistical items in the order of the largest gap between the ranges of statistical values of defective devices and non-defective devices according to the result of this operation and the smallest overlapping range. can be listed.

Referring to FIG. 5 , a gap between statistical values of defective devices and non-defective devices in the kurtosis (kur) value among statistical characteristic values for the vibration signal on the x-axis (refer to FIG. 5(a)) This is the largest, followed by a large gap between the ranges of statistical values of defective and non-defective devices in skew value. Next, although they overlap each other, the overlapping range of the statistical values of the devices and the defect-free devices in the max value is small.

Accordingly, in the case of FIG. 5 , when the top three axes and statistical items are listed, the kurtosis value of the x-axis, the skewness value of the x-axis, and the maximum value of the x-axis may be arranged in the order.

The values of these three axes and statistical items can be selected as the index data that best shows the case where the device is defective and the case where there is no defect.

Here, the number of higher-order items to be determined may be an arbitrary number according to device performance such as processing power and memory capacity of the device 200 for generating and verifying training data. That is, the greater the processing power and memory capacity of the apparatus 200 for generating and verifying training data, the greater the number of items may be determined.

The loading unit 240 may prepare a training data set after the type of data to be used for training the defect detection model (which statistic value on which axis) is determined. Normal to each data because it is known to the apparatus 200 that generates and validates the training data that the first devices 110 are defective devices and the second devices 120 are defect-free devices. Or it can be labeled as a failure. In this embodiment, the loading unit 240 includes an abnormal data set in which the determined first statistical characteristic values of the determined axis and statistical items are labeled as abnormal, and the determined second statistical characteristic values of the determined axis and statistical items are labeled as normal. A normal data set can be prepared.

The first generator 250 may train the initially set deep neural network model in a supervised learning method using the labeled training data. Here, the initially set deep neural network model is an initial model designed to be configured as a model capable of detecting device defects, and parameter values are set to arbitrary initial values.

The initial model may be completed as a defect detection model capable of accurately detecting a device defect while optimizing parameter values while training through the above-described training data set.

Here, since the first generator 250 is trained to detect a device defect using statistical characteristic values calculated from the frequency domain vibration signal without using the time domain vibration signal, training complexity can be reduced. In addition, since training is performed using only data determined to be useful for judging defects, not all statistical characteristic values of all axes, training can be performed more quickly and at a lower cost.

For example, the deep neural network model includes a first vibration signal (g_x FFT signal) in the frequency domain, a second vibration signal (g_y FFT signal) in the frequency domain, and a third vibration signal (g_z FFT signal) in the frequency domain, It may be configured to determine whether a device has failed based on a maximum value and a kurtosis value among statistical characteristic values for the fourth vibration signal (g_T FFT signal) in the frequency domain.

5 is a diagram illustrating a result of calculating a statistical characteristic value from a preprocessing result for a vibration signal output from a device in a normal state and a failure state according to the present embodiment.

Referring to FIG. 5 , FIG. 5A is a first vibration signal in the frequency domain obtained by preprocessing the first vibration signal g_x vibration signal in the time domain generated in the device 100 in a normal state and a broken state. The distribution of statistical characteristics calculated from the x-axis vibration signal, g_x FFT signal) is shown.

5B is a second vibration signal (g_y) in the frequency domain obtained by pre-processing the second vibration signal (y-axis vibration signal, g_y vibration signal) of the time domain generated in the device 100 in the normal state and the broken state. The distribution of statistical characteristics calculated from the FFT signal) is shown.

5C is a first vibration signal (g_z) in the frequency domain obtained by preprocessing the third vibration signal (z-axis vibration signal, g_z vibration signal) of the time domain generated in the device 100 in the normal state and the broken state. The distribution of statistical characteristics calculated from the FFT signal) is shown.

5D is a fourth vibration signal in the frequency domain obtained by pre-processing the fourth vibration signal in the time domain (the vibration signal of the diagonal T axis, the g_T vibration signal) generated in the device 100 in the normal state and the broken state. The distribution of statistical characteristics calculated from (g_T FFT signal) is shown.

Among the distributions of statistical characteristics calculated from the first vibration signal (g_x FFT signal) in the frequency domain in FIG. 5A , the difference between the kurtosis values of the normal state and the broken state is shown in FIGS. 5B to 5B . A normal state calculated from the second vibration signal (g_y FFT signal) of the frequency domain, the third vibration signal (g_z FFT signal) of the frequency domain, and the fourth vibration signal (g_T FFT signal) of the frequency domain shown in 5d And it can be seen that it is larger than the difference between the kurtosis (max) values of the broken state.

In addition, in the distribution of statistical characteristics calculated from the first vibration signal (g_x FFT signal) in the frequency domain in FIG. 5A , the difference between the skewness values of the normal state and the broken state is shown in FIG. 5B . The normal calculated from the second vibration signal (g_y FFT signal) of the frequency domain, the third vibration signal (g_z FFT signal) of the frequency domain, and the fourth vibration signal (g_T FFT signal) of the frequency domain shown in FIGS. 5D to 5D . ) state and it can be seen that it is larger than the difference between the skewness values of the broken state.

From FIG. 5 , if the arbitrary number of data items serving as criteria for determining failure and normality is two, the kurtosis and skewness values of the x-axis vibration signal will be the criteria for determining whether the device 100 is defective. it can be seen that In addition, if the arbitrary number of data items serving as a criterion for determining a failure and a normality is three, the kurtosis value, skewness value, and maximum value of the x-axis vibration signal are the criteria for determining whether the device 100 is defective. can be

6 is a diagram illustrating distributions of maximum values and kurtosis values among statistical characteristic values from pre-processing results for vibration signals output from devices in steady state and failure state according to the present embodiment.

Referring to FIG. 6 , FIG. 6A is a first vibration signal in the frequency domain obtained by preprocessing the first vibration signal g_x vibration signal in the time domain generated in the devices 100 in the normal state and the broken state. The distribution of the maximum value and the kurtosis value among the statistical characteristic values calculated from (g_x FFT signal) is shown. 6B is a second vibration signal (g_y FFT signal) in the frequency domain obtained by pre-processing the second vibration signal (g_y vibration signal) in the time domain generated in the devices 100 in the normal state and the broken state. The distribution of the maximum value and the kurtosis value among the calculated statistical characteristic values is shown.

FIG. 6c shows a third vibration signal (g_z FFT signal) in a frequency domain obtained by preprocessing a third vibration signal (g_z vibration signal) in the time domain generated in the devices 100 in a normal state and a broken state. The distribution of the maximum value and the kurtosis value among the calculated statistical characteristic values is shown.

FIG. 6d shows the fourth vibration signal (g_T FFT signal) in the frequency domain obtained by pre-processing the fourth vibration signal (g_T vibration signal) in the time domain generated in the devices 100 in the normal state and the broken state. The distribution of the maximum value and the kurtosis value among the calculated statistical characteristic values is shown.

In FIG. 6A , the distribution of the maximum (max) value and the kurtosis value of the normal state and the broken state calculated from the first vibration signal (g_x FFT signal) of the frequency domain is identified in FIGS. 6B to The normal calculated from the second vibration signal (g_y FFT signal) of the frequency domain, the third vibration signal (g_z FFT signal) of the frequency domain, and the fourth vibration signal (g_T FFT signal) of the frequency domain shown in FIG. 6D It can be seen that it is clearer than the identification of the distribution of the max and kurtosis values of the state and the broken state.

From FIG. 6 , the fact that the first vibration signal (g_y FFT signal) in the frequency domain serves as a criterion for determining whether a device is defective is only an example derived through an experiment and is not limited thereto, and in another embodiment, the first vibration signal (g_y FFT signal) signal) to the fourth vibration signal (g_T FFT signal) may be a reference, and may be other statistical characteristic values of those vibration signals.

Whether a certain statistical characteristic value of a signal of a certain axis is a criterion for determining whether a device is defective may be determined by the method of a flowchart to be described below.

From the above, the loading unit 240 includes an abnormal data set in which the determined first statistical characteristic value of the determined axis and statistical item is labeled as an abnormal state, and the determined second statistical characteristic value of the determined axis and the statistical item is labeled as a normal state. A normal data set may be prepared, and the second generator 260 may generate an augmented abnormal data group by applying at least a portion of the labeled abnormal data set to an auto encoder.

Here, the autoencoder is a kind of machine learning method and may be included in unsupervised learning. Using a neural network algorithm, an input can be learned with the goal of making an output value that is passed through a neural network as similar to the input value as possible. At this time, if the number of neural network neurons is greater than or equal to the dimension of the input value, the meaning of learning may be lost. This is because it is enough to take the input value as it is and export it. That is, in order to have meaning, the number of neurons may be smaller than the dimension of the input value. As a result of this learning, it is possible to obtain the effect of compression that can restore the original value with a smaller number of values.

6 is a block diagram schematically illustrating the configuration of the second generator 260 of FIG. 1 . Referring to FIG. 6 , the second generator 260 as an autoencoder may include an encoder 261 and a decoder 262 .

Referring to Figure 6, the autoencoder 260 has a hidden layer (hidden layer) internally, the number of nodes of the input layer (input layer, input layer) and the output layer (output layer, output layer) is the same , the encoder 261 and the decoder 262 may be symmetric based on the hidden layer in the middle (weighting, bias, number of nodes, etc.). In general, the number of nodes of the hidden layer may be smaller than that of the input layer (size of input data).

The encoder 261 may encode at least a portion of the labeled abnormal data set to extract a feature, and compress the extracted features. In the encoding process of the encoder 261, an input (at least a part of the labeled abnormal data set) is compressed, and as the input value is compressed, a loss from the input to the output may occur. As part of this encoding process, the hidden layer can learn key features of the input under the assumption that the input is not random.

The decoder 262 may generate an augmented abnormal data group having the same size as at least a portion of the labeled abnormal data set by decoding the compressed features.

The verification unit 270 may verify the accuracy of the abnormal data discrimination neural network by changing a ratio between the second abnormal data group and the augmented abnormal data group generated by the second generation unit 260 among the labeled abnormal data sets.

Here, the first abnormal data group among the labeled abnormal data sets may be previously generated and loaded, and the second abnormal data group among the labeled abnormal data sets may be newly created and loaded. That is, it may include that training data for learning the defect detection model may be different.

The second determiner 280 may determine whether to authenticate the training data for the augmented abnormal data group based on a result of verifying the accuracy of the abnormal data determination neural network.

In the present embodiment, the verifier 270 may obtain a first output value by inputting the second abnormal data group and the augmented abnormal data group among the abnormal data sets labeled in the abnormal data discrimination neural network at a first ratio. Here, the first ratio may include, for example, a second abnormal data group of 100% of the labeled abnormal data set and an augmented abnormal data group of 0%. Also, the first output value may include a value indicating whether the second abnormal data group and the 0% augmented abnormal data group among 100% of the labeled abnormal data set are normal or abnormal.

The verifier 270 may obtain a second output value by inputting the second abnormal data group and the augmented abnormal data group among the abnormal data sets labeled in the abnormal data determination neural network at a second ratio. Here, the second ratio may include, for example, a second abnormal data group of 60% of the labeled abnormal data set and an augmented abnormal data group of 40%. In addition, the second output value may include a value indicating whether the second abnormal data group of 60% of the labeled abnormal data set and the 40% augmented abnormal data group are normal or abnormal.

In this embodiment, the ratio has been disclosed by limiting the ratio to the first ratio and the second ratio, but this is only an example and may be variously modified from the first ratio to the N-th ratio.

The verification unit 270 may compare the ratio of abnormal determination in the first output value and the second output value.

As a result of the comparison of the abnormal determination ratio of the verification unit 270 , the second determination unit 280 determines that the difference between the abnormal determination ratio in the first output value and the abnormal determination rate in the second output value is less than or equal to a predetermined threshold value of the augmented abnormal data. Training data certification may be granted.

However, when the difference between the abnormal determination ratio in the first output value and the abnormal determination ratio in the second output value exceeds a predetermined threshold, the second determiner 280 may disallow authentication of the augmented abnormal data training data.

The second determining unit 280 instructs the second generating unit 260 to generate augmented abnormal data by applying at least another part of the labeled abnormal data set to the autoencoder as the training data authentication of the augmented abnormal data is disallowed. can do.

The controller 290 may control the overall operation of the apparatus 200 for generating and verifying training data. The controller 290 may include any type of device capable of processing data, such as a processor. Here, the 'processor' may refer to a data processing device embedded in hardware having a physically structured circuit to perform a function expressed by, for example, a code or an instruction included in a program. As an example of the data processing apparatus embedded in the hardware as described above, a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated (ASIC) circuit) and a processing device such as a field programmable gate array (FPGA), but the scope of the present invention is not limited thereto.

7 is a block diagram schematically illustrating a configuration of an apparatus for generating and verifying training data according to another embodiment. In the following description, descriptions of parts overlapping with those of FIGS. 1 to 6 will be omitted. Referring to FIG. 7 , an apparatus 200 for generating and verifying training data according to another embodiment may include a processor 201 and a memory 202 .

In this embodiment, the processor 201 includes the preprocessor 210 , the calculator 220 , the first determiner 230 , the loading unit 240 , the first generator 250 , and the second generator of FIG. 1 . Functions performed by the 260 , the verification unit 270 , the second determination unit 280 , and the control unit 290 may be processed.

The processor 201 may control the overall operation of the apparatus 200 for generating and verifying training data. Here, the 'processor' may refer to a data processing device embedded in hardware having a physically structured circuit to perform a function expressed by, for example, a code or an instruction included in a program. As an example of the data processing apparatus embedded in the hardware as described above, a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated (ASIC) circuit) and a processing device such as a field programmable gate array (FPGA), but the scope of the present invention is not limited thereto.

The memory 202 may be operatively connected to the processor 201 and store at least one code in association with an operation performed by the processor 201 .

In addition, the memory 202 may perform a function of temporarily or permanently storing data processed by the processor 201 . Here, the memory 202 may include a magnetic storage media or a flash storage media, but the scope of the present invention is not limited thereto. Such memory 202 may include internal memory and/or external memory, such as volatile memory such as DRAM, SRAM, or SDRAM, one time programmable ROM (OTPROM), PROM, EPROM, EEPROM, mask ROM, flash ROM, NAND flash memory, or non-volatile memory such as NOR flash memory, SSD, compact flash (CF) card, SD card, Micro-SD card, Mini-SD card, Xd card, or flash drive such as a memory stick , or a storage device such as HDD.

8 is a flowchart illustrating a method of generating and verifying training data according to the present embodiment. In the following description, descriptions of parts overlapping with those of FIGS. 1 to 7 will be omitted.

In step S810 , the apparatus 200 for generating and verifying training data may collect vibration signals from vibration sensors attached to the devices 100 . Here, the devices 100 may include the first devices 110 pre-determined to be defective and the second devices 120 pre-determined as having no defects.

Vibration sensors attached to the devices may be arranged to sense vibrations in at least three orthogonal x-axis, y-axis, and z-axis directions. In addition, a vibration sensor for detecting vibration of a diagonal axis defined by the x-axis, the y-axis, and the z-axis (eg, an axis in the (1, 1, 1) direction) may be disposed.

Since it is known to the apparatus 200 for generating and verifying the defect training data whether a device is a defective device or a normal device, the apparatus 200 for generating and verifying the defect training data determines that the collected vibration signal is a defective device. It is possible to know whether it was collected from or from a normal device.

The apparatus 200 for generating and verifying defect training data may selectively collect vibration signals by determining whether they are normal vibration signals or not. The apparatus 200 for generating and verifying the defect training data may compare the combination of the vibration signals of the three orthogonal axes with the vibration signals of the diagonal axes to evaluate whether the vibration signal of the three axes is abnormal. The apparatus 200 for generating and verifying the defect training data may collect only the vibration signal when the vibration signals of the three axes are normal based on the result of such evaluation.

For example, if the sum of the square values of the sizes of the vibration signals on three orthogonal axes of the device A is different from the square values of the sizes of the vibration signals on the diagonal axis, the vibration signal from the device A may not be a normal signal. Accordingly, the apparatus 200 for generating and verifying the defect training data may exclude the signal from the device A in performing the subsequent process.

In step S820, the apparatus 200 for generating and verifying the training data may extract statistical characteristic values for the vibration signals of each axis from the vibration signals converted into the frequency domain from the collected vibration signals. After the vibration signals collected in the present embodiment are converted into vibration signals in the frequency domain through Fourier transform, statistical characteristic values may be extracted from the vibration signals in the frequency domain for vibration signals of each axis. In the frequency domain, signal characteristics can be analyzed more clearly.

For example, first statistical characteristic values are extracted with respect to the vibration signals of each axis from the first vibration signals converted into the frequency domain, and vibration signals of each axis are extracted from the second vibration signals converted into the frequency domain. It is possible to extract second statistical characteristic values for .

That is, a statistical characteristic value may be extracted for each vibration signal of each axis obtained from each device among the first devices. For example, if there are device A, device B, and device C, 8 statistical characteristic values are extracted for each vibration signal of each of the x, y, and z axes of device A, and this can be performed in both B and C. For example, among the vibration signals from defective devices, 8 statistical characteristic values for the vibration signal on the x-axis, 8 statistical characteristic values for the vibration signal on the y-axis, and 8 statistical characteristic values for the vibration signal on the z-axis Eight statistical characteristic values for the characteristic values and the vibration signal of the T-axis are extracted, and the same type of signals can be extracted from devices without defects.

In step S830, the apparatus 200 for generating and verifying the defect training data compares the statistical characteristic values of the vibration signal of the defective device and the vibration signal of the non-defective device for each same statistical item on the same axis, and the threshold It is possible to determine axes and statistical items that have a difference of more than a value.

For example, to compare the average value of the vibration signals on the x-axis of devices A, B, and C, which are previously determined to be defective, with the average value of the vibration signals on the x-axis of devices, A', B', C', which are previously determined to be free of defects. it means you can

Here, the step of determining the axis and the statistical item having a difference greater than or equal to the threshold value includes a range of a first type of statistical values for vibration signals of a first axis among the first statistical characteristic values, and a range of the first type of statistical values among the second statistical characteristic values. The range of the first kind of statistical values for the vibration signals of the first axis may be compared, and the step of comparing may be repeated for all kinds of statistical values of the vibration signals of all axes.

During this iterative step, the axes and statistics items are listed in the order of the largest gap between the ranges of statistical values compared and the smallest overlapping ranges, and the upper axis and statistics of any number of the listed axes and statistics items You can select an item.

For example, referring to the case of FIG. 5, among the statistical characteristic values for the vibration signal on the x-axis (refer to FIG. 5(a)), the statistics of devices with and without defects in the kurtosis (kur) value The gap between the ranges of values is the largest, followed by the gap between the ranges of statistical values of defective devices and non-defective devices in skew value. Next, although they overlap each other, the overlapping range of the statistical values of the devices and the defect-free devices in the max value is small.

Accordingly, in the case of FIG. 5 , when the top three axes and statistical items are listed, the kurtosis value of the x-axis, the skewness value of the x-axis, and the maximum value of the x-axis may be arranged in the order. The values of these three axes and statistical items can be selected as the index data that best shows the case where the device is defective and the case where there is no defect. Here, the number of higher-order items may be determined as an arbitrary number according to device performance such as processing power and memory capacity of the device for generating a defect detection model. That is, the larger the processing power and memory capacity of the model generating apparatus, the greater the number of items can be determined.

In step S840, after the type of data (which statistical value on which axis) to be used for training the defect detection model is determined, the apparatus 200 for generating and verifying the defect training data labels the determined type of data to generate a training data set. can be prepared Normal to each data because it is known to the apparatus 200 that generates and validates the training data that the first devices 110 are defective devices and the second devices 120 are defect-free devices. Or it can be labeled as a failure. According to the above example, normal or faulty labeling can be performed with respect to the kurtosis value, skewness value, and maximum value of the vibration signals of the x-axis among the vibration signals of the first devices 110 and the second devices 120 . have. In this embodiment, the apparatus 200 for generating and verifying training data includes an abnormal data set in which the determined first statistical characteristic values of the determined axis and statistical item are labeled as abnormal, and the determined second statistical characteristic value of the determined axis and statistical item. You can prepare a normal data set labeled as steady state.

In step S850, the apparatus 200 for generating and verifying training data trains a deep neural network using the first abnormal data group and the labeled normal data set among the labeled abnormal data sets as training data, and the vibration generated in the device A defect detection model for determining whether a device is defective can be generated by using the statistical characteristic value of the signal. Here, the initially set deep neural network model is an initial model designed to be configured as a model capable of detecting device defects, and parameter values are set to arbitrary initial values.

The initial model may be completed as a defect detection model capable of accurately detecting a device defect while optimizing parameter values while training through the above-described training data.

Meanwhile, information on the axes and statistical items determined above may be tagged in the defect detection model to inform the user of the axes and statistical items to be input for use of the subsequently generated defect detection model.

The user can collect only the vibration signal of the corresponding axis from the collection stage through the information tagged in the defect detection model and input it into the defect detection model, and thus the computational resources required to apply the model can be efficiently used.

In step S860 , the apparatus 200 for generating and verifying training data may generate an augmented abnormal data group by applying at least a portion of the labeled abnormal data set to an auto encoder. The apparatus 200 for generating and verifying training data may extract features by encoding at least a portion of the labeled abnormal data set, and compress the extracted features. In the encoding process, the input (at least some of the labeled abnormal data set) is compressed, and as the input value is compressed, loss from input to output may occur. As part of this encoding process, the hidden layer can learn key features of the input under the assumption that the input is not random. The apparatus 200 for generating and verifying training data may generate an augmented abnormal data group having the same size as at least a portion of the labeled abnormal data set by decoding the compressed features.

In step S870 , the apparatus 200 for generating and verifying the training data may verify the accuracy of the defect detection model by changing the ratio of the second abnormal data group and the augmented abnormal data group among the labeled abnormal data set.

The apparatus 200 for generating and verifying training data may obtain a first output value by inputting the second abnormal data group and the augmented abnormal data group among the abnormal data sets labeled in the defect detection model at a first ratio. The apparatus 200 for generating and verifying training data may obtain a second output value by inputting the second abnormal data group and the augmented abnormal data group among the abnormal data sets labeled in the defect detection model at a second ratio. The apparatus 200 for generating and verifying the training data may calculate the accuracy of the defect detection model by comparing the ratio of abnormal determinations in the first output value and the second output value.

In step S880, the apparatus 200 for generating and verifying the training data may determine whether to authenticate the training data for the augmented abnormal data group based on the result of verifying the accuracy of the defect detection model. The apparatus 200 for generating and verifying training data may allow authentication of training data of augmented abnormal data when the difference between the abnormal determination ratio in the first output value and the abnormal determination ratio in the second output value is less than or equal to a predetermined threshold. However, the apparatus 200 for generating and verifying training data may disallow training data authentication of augmented abnormal data when the difference between the abnormal determination ratio in the first output value and the abnormal determination ratio in the second output value exceeds a predetermined threshold. have.

The apparatus 200 for generating and verifying training data automatically performs at least another part of the labeled abnormal data set when the difference between the abnormal judgment ratio in the first output value and the abnormal judgment ratio in the second output value exceeds a predetermined threshold. It can be applied to an encoder to generate augmented anomalous data.

The embodiment according to the present invention described above may be implemented in the form of a computer program that can be executed through various components on a computer, and such a computer program may be recorded in a computer-readable medium. In this case, the medium includes a hard disk, a magnetic medium such as a floppy disk and a magnetic tape, an optical recording medium such as CD-ROM and DVD, a magneto-optical medium such as a floppy disk, and a ROM. , RAM, flash memory, and the like, hardware devices specially configured to store and execute program instructions.

Meanwhile, the computer program may be specially designed and configured for the present invention, or may be known and used by those skilled in the computer software field. Examples of the computer program may include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

In the specification of the present invention (especially in the claims), the use of the term "above" and similar referential terms may be used in both the singular and the plural. In addition, when a range is described in the present invention, each individual value constituting the range is described in the detailed description of the invention as including the invention to which individual values belonging to the range are applied (unless there is a description to the contrary). same as

The steps constituting the method according to the present invention may be performed in an appropriate order, unless there is an explicit order or description to the contrary. The present invention is not necessarily limited to the order in which the steps are described. The use of all examples or exemplary terms (eg, etc.) in the present invention is merely for the purpose of describing the present invention in detail, and the scope of the present invention is limited by the examples or exemplary terms unless defined by the claims. it's not going to be In addition, those skilled in the art will appreciate that various modifications, combinations, and changes may be made in accordance with design conditions and factors within the scope of the appended claims or their equivalents.

Therefore, the spirit of the present invention should not be limited to the above-described embodiments, and the scope of the spirit of the present invention is not limited to the scope of the scope of the present invention. will be said to belong to

100: devices
200: a device for generating and validating training data
201: processor 202: memory
210: preprocessor 220: calculation unit
230: first determining unit 240: loading unit
250: first generating unit 260: second generating unit
270: verification unit 280: second determination unit
290: control unit

Claims (20)

A method of generating and validating training data, performed by a processor of a device generating and validating the training data, the method comprising:
Vibration sensors attached to the first devices determined to be defective and the second devices determined to be free, respectively, and arranged to sense vibration of at least three orthogonal axes while the first and second devices operate collecting first vibration signals of the first devices and second vibration signals of the second devices by
extracting first statistical characteristic values for the vibration signals of each axis from the first vibration signals, and extracting second statistical characteristic values with respect to the vibration signals of each axis from the second vibration signals;
determining an axis and a statistical item having a difference equal to or greater than a threshold value by comparing the first statistical characteristic values with the second statistical characteristic values for the same statistical items on the same axis;
preparing an abnormal data set in which the determined first statistical characteristic values of the axes and statistical items are labeled as abnormal and a normal data set in which the determined second statistical characteristic values of the axes and statistical items are labeled as normal;
A deep neural network is trained using the first abnormal data group among the labeled abnormal data set and the labeled normal data set as training data, and whether the device is defective by using the statistical characteristic value of the vibration signal generated in the device. generating a defect detection model to detect;
generating an augmented anomaly data group by applying at least a portion of the labeled anomaly data set to an auto encoder;
changing a ratio of a second abnormal data group and the augmented abnormal data group in the labeled abnormal data set and verifying accuracy of the defect detection model; and
Determining whether to authenticate training data for the augmented abnormal data group based on a result of verifying the accuracy of the defect detection model,
The step of verifying the accuracy of the defect detection model comprises:
obtaining a first output value by inputting the second abnormal data group and the augmented abnormal data group among the labeled abnormal data set to the defect detection model at a first ratio;
obtaining a second output value by inputting the second abnormal data group and the augmented abnormal data group among the labeled abnormal data set to the defect detection model at a second ratio; and
Comprising the step of comparing the ratio of abnormal determination in the first output value and the second output value,
How to generate and validate training data. The method of claim 1,
The first vibration signals and the second vibration signals include vibration signals of a diagonal axis defined by the three orthogonal axes in addition to the vibration signals of the three axes,
The collecting step is
comparing the combination of the vibration signals of the three axes with the vibration signals of the diagonal axis to evaluate whether the vibration signal of the three axes is abnormal; and
Comprising the step of collecting only the vibration signal in the case of normal among the vibration signals of the three axes based on the result of the evaluation step,
How to generate and validate training data. The method of claim 1,
After the collecting step,
The method further comprises converting the collected first and second vibration signals into a frequency domain through a Fourier transform,
The extraction step is
extracting first statistical characteristic values from the first vibration signals converted into the frequency domain; and
Including the step of extracting second statistical characteristic values from the second vibration signals converted to the frequency domain,
How to generate and validate training data. 4. The method of claim 3,
The first statistical characteristic values and the second statistical characteristic values are a mean value, a median value, a minimum value, a maximum value, a kurtosis ( Including one or more of a kurtosis value, a skewness value, a standard deviation value, and a peak-to-peak value,
How to generate and validate training data. The method of claim 1,
The step of determining the axis and the statistical item is,
The range of the first kind of statistical values for the vibration signals on the first axis among the first statistical characteristic values and the statistical values of the first kind for the vibration signals on the first axis among the second statistical characteristic values comparing ranges;
repeating the comparing step for all kinds of statistical values of vibration signals of all axes;
listing the axes and statistical items in the order of the largest gap between the ranges of statistical values compared during the repeating step and the smallest overlapping range; and
selecting any number of top axes and statistics items from among the listed axes and statistics items;
How to generate and validate training data. The method of claim 1,
The method further comprising the step of tagging the axis and the statistical item determined in the determining step to the defect detection model so as to inform the user of the axis and the statistical item to be input in the process of applying the defect detection model,
How to generate and validate training data. The method of claim 1,
The step of generating the augmented abnormal data group comprises:
extracting features by encoding at least a portion of the labeled abnormal data set, and compressing the extracted features; and
decoding the compressed features to generate an augmented anomalous data group similar to at least a portion of the labeled anomalous data set;
How to generate and validate training data. delete The method of claim 1,
The step of determining whether to authenticate the training data is,
allowing training data authentication of the augmented abnormal data when a difference between the abnormality determination ratio in the first output value and the abnormal determination rate in the second output value is less than or equal to a predetermined threshold; and
disallowing authentication of training data of the augmented abnormal data when the difference between the abnormal determination ratio in the first output value and the abnormal determination ratio in the second output value exceeds a predetermined threshold,
How to generate and validate training data. 10. The method of claim 9,
When the difference between the abnormal determination ratio in the first output value and the abnormal determination ratio in the second output value exceeds a predetermined threshold, at least another part of the labeled abnormal data set is applied to the autoencoder to generate augmented abnormal data further comprising the step of
How to generate and validate training data. A device for generating and validating training data, comprising:
processor; and
a memory operatively coupled to the processor and storing at least one code executed by the processor;
The memory, when executed by the processor, is attached to the first devices determined to be defective and the second devices determined to be non-defective, respectively, so that the processor is at least as long as the first and second devices are in operation. Collecting the first vibration signals of the first devices and the second vibration signals of the second devices by vibration sensors arranged to sense the vibration of three orthogonal axes,
extracting first statistical characteristic values for the vibration signals of each axis from the first vibration signals, and extracting second statistical characteristic values for the vibration signals of each axis from the second vibration signals;
determining an axis and a statistical item having a difference greater than or equal to a threshold value by comparing the first statistical characteristic values and the second statistical characteristic values for the same statistical items on the same axis;
Prepare an abnormal data set in which the determined first statistical characteristic values of the axis and the statistical item are labeled as an abnormal state, and a normal data set in which the determined second statistical characteristic values of the determined axis and the statistical item are labeled as a normal state,
A deep neural network is trained using the first abnormal data group among the labeled abnormal data set and the labeled normal data set as training data, and whether the device is defective by using the statistical characteristic value of the vibration signal generated in the device. create a defect detection model to detect,
applying at least a portion of the labeled anomaly data set to an auto encoder to generate an augmented anomaly data group;
changing a ratio of a second abnormal data group and the augmented abnormal data group in the labeled abnormal data set and verifying accuracy of the defect detection model;
storing a code causing a determination of whether to authenticate training data for the augmented abnormal data group based on a result of verifying the accuracy of the defect detection model;
The memory causes the processor to
When verifying the accuracy of the defect detection model, inputting the second abnormal data group and the augmented abnormal data group among the labeled abnormal data set to the defect detection model at a first ratio to obtain a first output value,
obtaining a second output value by inputting the second abnormal data group and the augmented abnormal data group among the labeled abnormal data set to the defect detection model at a second ratio;
storing a code causing comparison of a ratio of abnormal determinations in the first output value and the second output value;
A device that generates and validates training data. 12. The method of claim 11,
The memory causes the processor to
The first vibration signals and the second vibration signals include vibration signals of a diagonal axis defined by the three orthogonal axes in addition to the vibration signals of the three axes,
When collecting the vibration signals, the combination of the vibration signals of the three axes is compared with the vibration signal of the diagonal axis to evaluate whether the vibration signal of the three axes is abnormal,
Storing a code that causes only the vibration signal in the normal case among the vibration signals of the three axes to be collected based on the result of the evaluation step,
A device that generates and validates training data. 12. The method of claim 11,
The memory causes the processor to
after collecting the vibration signals, further storing a code causing the collected first vibration signals and the second vibration signals to be converted into a frequency domain through a Fourier transform,
When extracting the characteristic values, first statistical characteristic values are extracted from the first vibration signals converted to the frequency domain, and second statistical characteristic values are extracted from the second vibration signals converted to the frequency domain to store the code to
A device that generates and validates training data. 14. The method of claim 13,
The memory causes the processor to
The first statistical characteristic values and the second statistical characteristic values are a mean value, a median value, a minimum value, a maximum value, a kurtosis ( storing a code causing it to include one or more of a kurtosis value, a skewness value, a standard deviation value, and a peak-to-peak value;
A device that generates and validates training data. 12. The method of claim 11,
The memory causes the processor to
When determining the axis and the statistical item, the range of the first kind of statistical values for the vibration signals of the first axis among the first statistical characteristic values and the vibration signals of the first axis among the second statistical characteristic values comparing the ranges of the first kind of statistical values for
repeating the comparing step for all kinds of statistical values of vibration signals of all axes,
In the repeating step, the axes and statistical items are listed in the order of the largest gap between the ranges of statistical values compared and the smallest overlapping range,
storing code that causes the selection of any number of top axes and statistics items among the listed axes and statistics items;
A device that generates and validates training data. 12. The method of claim 11,
The memory causes the processor to
Further storing a code causing tagging of the determined axes and statistical items in the defect detection model so as to inform the user of the axes and statistical items to be input in the process of applying the defect detection model,
A device that generates and validates training data. 12. The method of claim 11,
The memory causes the processor to
When generating the augmented anomaly data group, encoding at least a part of the labeled anomaly data set to extract a feature, compressing the extracted features,
storing code that causes decoding of the compressed features to generate an augmented anomalous data group similar to at least a portion of the labeled anomalous data set;
A device that generates and validates training data. delete 12. The method of claim 11,
The memory causes the processor to
When determining whether to authenticate the training data, if the difference between the abnormal determination rate in the first output value and the abnormal determination rate in the second output value is less than or equal to a predetermined threshold, the training data authentication of the augmented abnormal data is allowed,
storing a code for causing disallowing training data authentication of the augmented abnormal data when a difference between the abnormal determination ratio in the first output value and the abnormal determination ratio in the second output value exceeds a predetermined threshold;
A device that generates and validates training data. 20. The method of claim 19,
The memory causes the processor to
When the difference between the abnormal determination ratio in the first output value and the abnormal determination ratio in the second output value exceeds a predetermined threshold, at least another part of the labeled abnormal data set is applied to the autoencoder to generate augmented abnormal data to store more code that causes it to
A device that generates and validates training data.

Images