KR102378729B1 - Method and apparatus for determining optimized training scheme - Google Patents

Abstract

The present invention relates to a method and apparatus for determining an optimized training scheme. According to an embodiment of the present invention, the method for determining an optimized training scheme may comprise the steps of: loading an original data set related to a vibration signal generated by a machine; generating fake abnormal data using a generative adversarial network; generating a plurality of training data sets by changing the ratio of a part of the original data set and the fake abnormal data; extracting features by inputting the plurality of training data sets to a plurality of different feature extraction algorithms; generating a plurality of machine defect discrimination training models by training a plurality of training models implemented with different algorithms using the feature extraction result as training data; verifying the accuracy of the plurality of machine defect discrimination training models using other parts of the original data set; and determining an optimized training model among the plurality of machine defect discrimination training models, and an optimized ratio of the part of the original data set and the fake abnormal data according to each feature extraction algorithm, based on the result of verifying the accuracy of the machine defect discrimination training model. The present invention can prevent mechanical defects in advance.

Description

Method and apparatus for determining the optimal training method {METHOD AND APPARATUS FOR DETERMINING OPTIMIZED TRAINING SCHEME}

The present invention is based on the results of verifying the accuracy of the machine defect identification learning model, and the optimal ratio of original vibration data and fake data generated using a generative adversarial network (GAN) for each feature extraction algorithm and , relates to a method and apparatus for determining an optimal training method for determining an optimal learning model among a plurality of machine defect discrimination learning models.

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. Since 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.

Areas 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 to detect and diagnose faults through various sensor signals is being actively conducted.

The above-mentioned background art is technical information that the inventor possessed for the 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, based on the result of verifying the accuracy of the machine defect identification learning model, the optimal ratio of original vibration data and fake data generated using an adversarial neural network (GAN) for each feature extraction algorithm and , is to determine the optimal learning model among a plurality of machine defect discrimination learning models.

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 understood that the problems and advantages to be solved by the present invention can be realized by means and combinations thereof indicated in the claims.

A method for determining an optimal training method according to an embodiment of the present invention comprises the steps of: loading an original data set including a labeled normal data set and a labeled abnormal data set in relation to a vibration signal generated by a machine; generating a fake anomalous data group by applying at least a portion of the acquired anomaly data set to a generative adversarial network (GAN); generating a plurality of training data sets while changing Using the results as training data, a plurality of learning models implemented with different algorithms are trained, and a plurality of machine fault discrimination learning implemented with different algorithms to determine whether the machine is defective using the characteristics of the vibration signal generated by the machine. Creating a model, verifying the accuracy of a plurality of machine defect determination learning models using a second data group from the original data set, and verifying the accuracy of the machine defect determination learning model. The method may include determining an optimal ratio of the first data group and the simulated abnormal data group in the original data set for each extraction algorithm, and an optimal learning model among a plurality of machine defect discrimination learning models.

An apparatus for determining an optimal training method according to an embodiment of the present invention includes a processor and a memory operably connected to the processor and storing at least one code executed by the processor, wherein the memory is executed through the processor, the processor With respect to the vibration signal generated by the machine, it loads an original data set including a labeled normal data set and a labeled anomalous data set, and generates at least a portion of the labeled anomalous data set in a generative adversarial network (GAN). network) to create a fake abnormal data group, change the ratio of the first data group and the simulated abnormal data group in the original data set to create a plurality of training data sets, and It is generated in the machine by training a plurality of learning models implemented with different algorithms by inputting them into a plurality of different feature extraction algorithms to extract features, and using the results of inputting features into a plurality of different feature extraction algorithms as training data. Generates a plurality of machine defect determination learning models implemented with different algorithms to determine whether a machine is defective by using the characteristics of the vibration signal, Based on the results of verifying the accuracy of and validating the accuracy of the machine defect discrimination learning model, the optimal ratio of the first data group and the simulated abnormal data group among the original data set for each feature extraction algorithm, and multiple machine defect discrimination learning It is possible to store the code that causes the model to determine the optimal learning 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, based on the result of verifying the accuracy of the machine fault determination learning model, the optimal ratio of original vibration data and fake data generated using an adversarial neural network (GAN) for each feature extraction algorithm, It is possible to determine the optimal learning model among the machine fault discrimination learning models of

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

In addition, it is possible to increase the training data using a generative adversarial network (GAN) to improve the performance of the machine fault determination training process.

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 and 2 are block diagrams schematically illustrating the configuration of an apparatus for determining an optimal training method according to the present embodiment.
3 is an exemplary view for explaining an output direction of a vibration signal output from the machine according to the present embodiment.
4 is an exemplary view illustrating vibration signals output from the normal machine and the abnormal machine according to the present embodiment.
FIG. 5 is a block diagram schematically illustrating the configuration of the first generator in FIG. 1 .
6 is a diagram illustrating a result of converting a vibration signal output from a normal machine and an abnormal machine into a frequency domain according to the present embodiment.
7 is a view showing the result of calculating statistical characteristic values from vibration signals output from the normal machine and the abnormal machine converted to the frequency domain according to the present embodiment.
8 is a block diagram schematically illustrating a configuration of an apparatus for determining an optimal training method according to another exemplary embodiment.
9 is a flowchart illustrating a method for determining an optimal training method according to the present embodiment.

Advantages and features of the present invention, and a method for achieving them will become apparent with reference to the detailed description in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments presented below, it can be implemented in a variety of different forms, and should be understood to include all transformations, equivalents, and substitutes included in the spirit and scope of the present invention. . The embodiments presented below are provided so that the disclosure of the present invention is complete, 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.

Also, 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. decide to do

1 and 2 are block diagrams schematically illustrating the configuration of an apparatus for determining an optimal training method according to the present embodiment. 1 and 2 , the apparatus 100 for determining the optimal training method includes a storage unit 110 , a loading unit 120 , a first generation unit 130 , a second generation unit 140 , and an extraction unit ( 150 ), a third generation unit 160 , a verification unit 170 , a determination unit 180 , and a control unit 190 may be included.

The storage unit 110 may store an original data set including a labeled normal data set and a labeled abnormal data set in relation to a vibration signal generated by a machine (eg, a rotating machine, not shown).

The loading unit 120 may load the original data set including the labeled normal data set and the labeled abnormal data set. The loading unit 120 includes a normal data set in which a first vibration signal generated in a normal machine (eg, a normal rotating machine, not shown) is labeled as a normal state, and an abnormal machine (eg, abnormal rotation) An original data set including an abnormal data set in which a second vibration signal generated in a machine (not shown) is labeled as a broken state may be loaded.

In this embodiment, loading may be interpreted as receiving a labeled normal data set from a normal machine and receiving a labeled abnormal data set from an abnormal machine. Also, the meaning of loading may be interpreted as receiving an original data set including the labeled normal data set and the labeled abnormal data set stored in the storage unit 110 .

In this embodiment, the vibration signal generated in a normal machine or an abnormal machine is a vibration signal in the horizontal (x) direction (g_x vibration signal) and a vibration signal in the vertical (y) direction (g_y vibration) as shown in FIG. 3 . signal), a vibration signal in the axis (z) direction (g_z vibration signal), and a vibration signal in the diagonal (T) direction (g_T vibration signal) in the horizontal (x) direction, vertical (y) direction, and axis (z) direction. ) may be included.

4 is an exemplary diagram illustrating vibration signals output from a normal machine and an abnormal (failure) machine according to the present embodiment. Referring to FIG. 4, FIG. 4A shows a vibration signal (g_x) in the horizontal (x) direction generated in a normal machine, and FIG. 4B is a horizontal (x) generated in an abnormal (broken) machine. The vibration signal (g_x vibration signal) of the direction is shown. In particular, Figure 4b shows a vibration signal (g_x vibration signal) in the horizontal (x) direction generated in the rotating machine when the tooth (tooth) is broken.

The first generator 130 may generate a fake abnormal data group by applying at least a portion of the labeled abnormal data set to a generative adversarial network (GAN).

Here, the generative adversarial neural network is an artificial intelligence algorithm used for unsupervised learning, and is composed of a generator and a discriminator. By iteratively learning the process of probabilistically examining whether the simulated data is real or fake based on the data, it may mean an algorithm in which the generator is configured to generate simulated data that is almost similar to the real data.

FIG. 5 is a block diagram schematically illustrating the configuration of the first generator 130 of FIG. 1 . Referring to FIG. 6 , the first generator 130 may include a generator 131 and an identifier 132 .

The generator 131 may generate simulated abnormal data as a predetermined input value. Here, the input value may be a random value. In an embodiment, the random value may be random noise generated using a Gaussian distribution.

The identifier 132 may identify whether the simulated anomalous data is true or false by comparing at least a portion of the simulated abnormal data generated by the generator 131 and the labeled abnormal data set.

In this embodiment, the identifier 132 may be a neural network trained using the training data generated by the generator 131, that is, simulated abnormal data. In addition, the generator 131 is a neural network trained using the simulated abnormal data determined to be false by the identifier 132 as training data, and may generate simulated abnormal data close to the actual abnormal data. From this, the generative adversarial neural network can be said to be an algorithm that learns through competition between two neural networks and derives results.

A neural network including a generator 131 and an identifier 132 typically performs dense processing, batch normalization processing, activation processing, input reshaping processing, and Gaussian drop processing. Multi-layers consisting of multiple processing layers, such as out processing (gaussian dropout processing), Gaussian noise processing (gaussian noise processing), two-dimensional convolution, and two-dimensional up sampling (two-dimensional up sampling) It may be implemented by a layer network.

Returning to FIG. 1 , the second generator 140 may generate a plurality of training data sets by changing the ratio of the first data group and the simulated abnormal data group among the original data sets.

In this embodiment, the second generator 140 may generate a first training data set of a first ratio in which the first data group of the original data set is 100% and the simulated abnormal data group is 0%. In this embodiment, the second generator 140 may generate a second training data set having a second ratio in which the first data group among the original data sets is 80% and the simulated abnormal data group is 20%. In the present embodiment, the second generator 140 may generate a third training data set having a third ratio in which the first data group among the original data sets is 60% and the simulated abnormal data group is 40%. In the present embodiment, the second generator 140 may generate a fourth training data set having a fourth ratio in which the first data group among the original data sets is 40% and the simulated abnormal data group is 60%. In this embodiment, the second generator 140 may generate a fifth training data set with a fifth ratio in which the first data group among the original data sets is 20% and the simulated abnormal data group is 80%. In this embodiment, the second generator 140 may generate a sixth training data set with a sixth ratio in which the first data group among the original data sets is 0% and the simulated abnormal data group is 100%.

In the present embodiment, the ratio of the first data group and the simulated abnormal data group in the original data set is limited to six first to sixth ratios, but the application of various embodiments such as six or more or less than six is not limited thereto. This may be possible.

The extractor 150 may extract features by inputting a plurality of training data sets into a plurality of different feature extraction algorithms. In this embodiment, the plurality of training data sets may include a first training data set to a sixth training data set.

The extractor 150 may extract, as a first feature, a data group of a frequency domain generated by applying a first feature extraction algorithm to a plurality of training data sets (a first training data set to a sixth training data set).

Here, the first feature extraction algorithm may include a fast Fourier transform (FFT) algorithm for converting a time domain vibration signal into a frequency domain vibration signal.

Since the vibration signal in the frequency domain contains more unique information that can be used to distinguish normal and failure of the machine, the first feature extraction algorithm, FFT algorithm, is used to convert the time domain vibration signal into a frequency domain vibration signal. can be converted

6 is a diagram illustrating a result of converting a vibration signal output from a normal machine and an abnormal machine into a frequency domain according to the present embodiment. Referring to FIG. 6 , FIG. 4A shows a horizontal (x) direction vibration signal (g_x FFT) in the frequency domain by applying a time domain horizontal (x) vibration signal (g_x vibration signal) generated in a normal machine to the first feature extraction algorithm. An example of extracting signal) is shown. Figure 4b is an example of extracting the horizontal (x) direction vibration signal (g_x FFT signal) of the frequency domain by applying the horizontal (x) direction vibration signal (g_x vibration signal) of the time domain generated in the abnormal machine to the first feature extraction algorithm is showing

In this embodiment, the extractor 150 may extract six sets of first features by applying the first to sixth training data sets to the first feature extraction algorithm.

In addition, the extraction unit 150 is a statistical analysis calculated by applying the second feature extraction algorithm to the data group of the frequency domain (that is, the FFT-transformed first training data set to the sixth training data set) It can be extracted as the second feature.

Here, the second feature extraction algorithm is the average value, median value, minimum Algorithm to calculate at least one of (min) value, maximum (max) value, kurtosis value, skewness value, standard deviation value, and peak-to-peak value may include

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 index indicating the asymmetry of the probability distribution, and as the signal bias (the degree to which the distribution of signal values converges 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.

7 is a view showing the result of calculating statistical characteristic values from vibration signals output from the normal machine and the abnormal machine converted to the frequency domain according to the present embodiment. Referring to FIG. 7 , FIG. 7A shows the distribution of statistical characteristics calculated from the horizontal (x) direction vibration signal (g_x FFT signal) in the frequency domain. 7B shows the distribution of statistical characteristics calculated from the vertical (y) direction vibration signal (g_y FFT signal) in the frequency domain. 7C shows the distribution of statistical characteristics calculated from a vibration signal (g_z FFT signal) in the axis (z) direction of the frequency domain. 7D shows the distribution of statistical characteristics calculated from a vibration signal (g_T FFT signal) in the diagonal (T) direction of the frequency domain.

In this embodiment, the extractor 150 may extract six sets of second features by applying the first to sixth training data sets converted to the frequency domain to the second feature extraction algorithm. Here, the second feature output by the second feature extraction algorithm is at least 1 (one of the mean value, median value, minimum value, maximum value, kurtosis value, skewness value, standard deviation value, and positive amplitude value) to a maximum of 8 (average value). , median, minimum, maximum, kurtosis, skewness, standard deviation, and positive amplitude), and at least 6 to 48 sets of second features can be extracted for each training data set. there is.

The extraction unit 150 may extract a reduced-dimensional data group generated by applying a third feature extraction algorithm to a plurality of training data sets (the first training data set to the sixth training data set) as the third feature. .

Here, the third feature extraction algorithm may include a principal component analysis (PCA) algorithm that generates data with reduced dimensions.

Principal component analysis may be a representative algorithm among techniques for reducing high-dimensional data to low-dimensional data in order to analyze multivariate data. An orthogonal transformation can be used to transform samples in a high-dimensional space that are likely to be related to each other into samples in a low-dimensional space (principal components) that do not have linear correlation. This algorithm extracts and summarizes key information by reducing the dimension of the data, and minimizes the loss of information by selecting the first few principal components that express the most variability in the data. Therefore, the number of dimensions of the principal component can be less than or equal to the number of dimensions of the original sample. In particular, the characteristic of principal component analysis may include transforming samples in a high-dimensional space into a low-dimensional space without linear correlation by finding new basis orthogonal to each other while maximally preserving the variance of the data. Therefore, in principal component analysis, when data is mapped to one axis, the data can be linearly transformed into a new coordinate system so that the axis with the largest variance is the first principal component and the second axis is the second principal component. It can be applied in various ways by decomposing the sample difference into the components that best represent the difference.

In this embodiment, the extractor 150 may extract six sets of third features by applying the first to sixth training data sets to the third feature extraction algorithm.

As a result, the extraction unit 150 generates a total of 6 sets of first features, a minimum of 6 sets to a maximum of 48 sets of second features, and a total of at least 18 sets to a maximum of 60 sets of different 3 sets of 6 sets. features can be extracted.

In the present embodiment, the feature extraction algorithm is limited to three, but it is not limited thereto, and various embodiments such as three or more or less than three may be applied.

The third generation unit 160 trains a plurality of learning models implemented with different algorithms using a plurality of different features extracted by the extraction unit 150 as training data, and uses the characteristics of the vibration signal generated by the machine. It is possible to create a plurality of machine defect determination learning models implemented with different algorithms for determining whether a machine is defective.

The third generator 160 trains the first learning model implemented with the first algorithm using the first to third features as training data, respectively, and uses the features of the vibration signal generated by the machine to determine whether the machine is defective. It is possible to generate a plurality of first machine defect determination learning models to determine. Here, the training data for training the first learning model may include a total of a minimum of 18 sets to a maximum of 60 sets of training data, and the first learning model may be trained with the training data of each set. In addition, the plurality of first machine defect determination learning models may include a total of at least 18 to a maximum of 60 first machine defect determination learning models for each set.

In this embodiment, the first learning model may include an artificial neural network (ANN). An artificial neural network is an algorithm inspired by machine learning and biological brain structure, and it is a method for neurons that form a network by combining synapses to change the strength of synapses through data learning and to have problem-solving ability. The neural network may include a non-linear structure of statistical data composed of a modeling tool.

The third generator 160 trains the second learning model implemented with the second algorithm using the first to third features as training data, respectively, and uses the features of the vibration signal generated by the machine to determine whether the machine is defective. It is possible to generate a plurality of second machine defect determination learning models to determine. Here, the training data for training the second learning model may include a total of minimum 18 sets to a maximum of 60 sets of training data, and the second learning model may be trained with each set of training data. In addition, the plurality of second machine defect determination learning models may include a total of at least 18 to a maximum of 60 second machine defect determination learning models for each set.

In this embodiment, the second learning model may include K-means clustering. K-means clustering is an algorithm that groups given data into k clusters, and can operate in a way that minimizes the variance of each cluster and distance difference. This algorithm is a kind of self-learning, and can play the role of labeling unlabeled input data.

The third generator 160 trains the third learning model implemented with the third algorithm using the first to third features as training data, respectively, and uses the features of the vibration signal generated by the machine to determine whether the machine is defective. It is possible to generate a plurality of third machine defect determination learning models to determine. Here, the training data for training the third learning model may include a total of minimum 18 sets to a maximum of 60 sets of training data, and the third learning model may be trained with each set of training data. In addition, the plurality of third machine defect determination learning models may include a total of at least 18 to a maximum of 60 third machine defect determination learning models for each set.

In this embodiment, the third learning model may include a support vector machine (SVM). A support vector machine is an algorithm that transforms input data into a high-dimensional space and finds an optimal decision boundary that can be linearly separated in a high-dimensional space. Here, the optimal boundary can play a role in predicting and classifying data classes. At this time, different data must be classified with the most margin so that any data can be classified well.

In this embodiment, the number of learning models is limited to three, but the number of learning models is not limited thereto, and various embodiments such as three or more or less than three may be applied.

The third generation unit 160 may output a first classification value for whether a total of 18 to 60 machine defects in total from the first machine defect determination learning model, and a total minimum of 18 from the second machine defect determination learning model. It is possible to output a second classification value for whether or not there are up to 60 machine defects, and output a third classification value for a total of at least 18 to up to 60 machine defects from the third machine defect determination learning model. . Accordingly, the third generation unit 160 may output a classification value for whether a total of at least 54 to a maximum of 180 machine defects.

The verification unit 170 may verify the accuracy of the plurality of machine defect determination learning models by using the second data group among the original data sets.

Here, the first data group among the original data sets may include 70% of the original data sets among 100% of the original data sets. In addition, the second data group of the original data set may include 30% of the original data set among 100% of the original data set, and may be test data for verification.

In this embodiment, verifying the accuracy means using the first data group and the simulated abnormal data group in the original data set, the classification value output from each of a plurality of machine defect determination learning models, and the second data group in the original data set. It may include analyzing the degree of agreement by comparing the classification values output from each of a plurality of machine defect identification learning models using .

The verification unit 170 may verify the first accuracy of the plurality of first machine defect determination learning models by using the second data group among the original data sets. The verification unit 170 may verify the second accuracy of the plurality of second machine defect determination learning models by using the second data group among the original data sets. The verification unit 170 may verify the third accuracy of the plurality of third machine defect determination learning models by using the third data group among the original data sets.

The determiner 180 determines the optimal ratio of the first data group and the simulated abnormal data group in the original data set for each feature extraction algorithm, based on the result of verifying the accuracy of the machine defect determination learning models, and the plurality of machine defect determination learning models Among the models, the optimal learning model can be determined.

The determination unit 180 includes a first data group in a training data set having the lowest ratio of a first data group among a training data set that derives more than a predetermined threshold accuracy from a result of verifying the accuracy of the machine fault determination learning models; The proportion of simulated anomalous data groups can be determined as an optimal proportion.

The determiner 180 may determine any one machine defect determination learning model having the highest accuracy among the first machine defect determination learning model to the third machine defect determination learning model as the optimal learning model.

The determiner 180 may determine, as an optimal feature extraction algorithm, any one of the first feature extraction algorithms to the third feature extraction algorithms that generates training data input to the machine defect identification learning model having the highest accuracy. .

Effective feature extraction algorithms and artificial intelligence models are different for each type of machine, and the ratio between original data and simulated abnormal data may be different. and a feature extraction algorithm may be selected (determined).

Here, the selection may include any one of the above-described first to sixth ratios, and any one of a plurality of machine defect determination learning models outputting a minimum of 54 to a maximum of 180 classification values. may include, and any one algorithm among a plurality of feature extraction algorithms for extracting a minimum of 18 to a maximum of 60 features may be selected.

In the ratio of the original data to the simulated data, increasing the size of the training data by adding the simulated data and decreasing the accuracy of the training data due to the addition of the simulated data may be a trade-off relationship.

According to the determination of the optimal training method of the present invention, it is possible to find the maximum simulated data ratio that does not decrease the accuracy of the training data while increasing the size of the training data (or does not drop more than a certain amount).

The controller 190 may control the overall operation of the apparatus 100 for determining the optimal training method. The controller 190 may include all kinds of devices 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.

8 is a block diagram schematically illustrating a configuration of an apparatus for determining an optimal training method according to another exemplary embodiment. In the following description, descriptions of parts overlapping with those of FIGS. 1 to 7 will be omitted.

Referring to FIG. 8 , the apparatus 100 for determining an optimal training method according to another embodiment may include a processor 101 and a memory 102 .

In the present embodiment, the processor 101 includes the storage unit 110 , the loading unit 120 , the first generation unit 130 , the second generation unit 140 , the extraction unit 150 , and the third generation unit of FIG. 1 . Functions performed by 160 , the verification unit 170 , the determiner 180 , and the control unit 190 may be processed.

The processor 101 may control the overall operation of the apparatus 100 for determining the optimal training method. 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 102 may be operatively connected to the processor 101 and store at least one code in association with an operation performed by the processor 101 .

In addition, the memory 102 may perform a function of temporarily or permanently storing data processed by the processor 101 . Here, the memory 102 may include a magnetic storage media or a flash storage media, but the scope of the present invention is not limited thereto. Such memory 102 may include internal memory and/or external memory, including 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.

9 is a flowchart illustrating a method for determining an optimal training method according to the present embodiment. In the following description, descriptions of parts overlapping with those of FIGS. 1 to 8 will be omitted.

Referring to FIG. 9 , in step S910 , the apparatus 100 for determining the optimal training method loads the original data set including the labeled normal data set and the labeled abnormal data set in relation to the vibration signal generated by the machine. can In this embodiment, the optimal training method determination apparatus 100 breaks the normal data set in which the first vibration signal generated in the normal machine is labeled as normal and the second vibration signal generated in the abnormal machine is broken. ), we can load the original data set containing the anomalous data set labeled with the state.

In step S920 , the apparatus 100 for determining the optimal training method may generate a fake abnormal data group by applying at least a portion of the labeled abnormal data set to a generative adversarial network (GAN). In this embodiment, the optimal training method determination apparatus 100 generates simulated abnormal data from random noise generated using a Gaussian distribution, and simulates abnormality through comparison of at least some of the simulated abnormal data and the labeled abnormal data set. It can identify whether data is true or false.

In operation S930, the optimal training method determining apparatus 100 may generate a plurality of training data sets by changing the ratio of the first data group and the simulated abnormal data group among the original data sets. In this embodiment, the apparatus 100 for determining the optimal training method may generate the training data set while changing the first to sixth ratios.

In operation S940 , the apparatus 100 for determining the optimal training method may extract a feature by inputting a plurality of training data sets into a plurality of different feature extraction algorithms.

In the present embodiment, the apparatus 100 for determining the optimal training method may extract, as a first feature, a data group of a frequency domain generated by applying a first feature extraction algorithm to a plurality of training data sets. Here, the first feature extraction algorithm may include a fast Fourier transform (FFT) algorithm for converting a time domain vibration signal into a frequency domain vibration signal.

Also, the apparatus 100 for determining the optimal training method may extract a statistical analysis calculated by applying the second feature extraction algorithm to the data group of the frequency domain as the second feature. Here, the second feature extraction algorithm is the average value, median value, To calculate one or more of the minimum (min) value, the maximum (max) value, the kurtosis value, the skewness value, the standard deviation value, and the peak-to-peak value Algorithms may be included.

Also, the apparatus 100 for determining the optimal training method may extract, as a third feature, a data group with reduced dimensions generated by applying a third feature extraction algorithm to a plurality of training data sets. Here, the third feature extraction algorithm may include a principal component analysis algorithm that generates data with reduced dimensions.

In the present embodiment, the apparatus 100 for determining the optimal training method may extract a minimum of 18 to a maximum of 60 features by using the data of the first to sixth ratios.

In step S950, the apparatus 100 for determining the optimal training method trains a plurality of learning models implemented with different algorithms by inputting a plurality of different feature extraction algorithms and using the result of feature extraction as training data, and generating in the machine It is possible to generate a plurality of machine defect determination learning models implemented with different algorithms for determining whether a machine is defective by using the characteristics of the vibration signal.

In this embodiment, the optimal training method determination apparatus 100 trains the first learning model implemented with the first algorithm by using the first to third characteristics as training data, respectively, and the characteristics of the vibration signal generated by the machine It is possible to generate a plurality of first machine defect determination learning models that determine whether the machine is defective by using . Here, the training data for training the first learning model may include a total of a minimum of 18 sets to a maximum of 60 sets of training data, and the first learning model may be trained with the training data of each set. In addition, the plurality of first machine defect determination learning models may include a total of at least 18 to a maximum of 60 first machine defect determination learning models for each set.

In addition, the optimal training method determination apparatus 100 trains a second learning model implemented with a second algorithm using the first to third features as training data, respectively, and uses the features of the vibration signal generated by the machine. A plurality of second machine defect determination learning models for determining whether a machine is defective may be generated. Here, the training data for training the second learning model may include a total of minimum 18 sets to a maximum of 60 sets of training data, and the second learning model may be trained with each set of training data. In addition, the plurality of second machine defect determination learning models may include a total of at least 18 to a maximum of 60 second machine defect determination learning models for each set.

In addition, the optimal training method determination apparatus 100 trains the third learning model implemented with the third algorithm using the first to third features as training data, respectively, and uses the features of the vibration signal generated by the machine. A plurality of third machine defect determination learning models for determining whether a machine is defective may be generated. Here, the training data for training the third learning model may include a total of minimum 18 sets to a maximum of 60 sets of training data, and the third learning model may be trained with each set of training data. In addition, the plurality of third machine defect determination learning models may include a total of at least 18 to a maximum of 60 third machine defect determination learning models for each set.

The optimal training method determination apparatus 100 may output a first classification value for whether a total of at least 18 to a maximum of 60 machine defects from the first machine defect determination learning model, and a total from the second machine defect determination learning model It is possible to output a second classification value for whether there are a minimum of 18 to a maximum of 60 machine defects, and a third classification value for a total of at least 18 to a maximum of 60 machine defects can be output from the third machine defect determination learning model. can Accordingly, the apparatus 100 for determining the optimal training method may output a classification value for a total of at least 54 and at most 180 machine defects.

In operation S960 , the apparatus 100 for determining the optimal training method may verify the accuracy of the plurality of machine defect identification learning models by using the second data group among the original data sets. In this embodiment, the apparatus 100 for determining the optimal training method verifies the first accuracy of the plurality of first machine defect determination learning models by using the second data group among the original data sets, and the second among the original data sets Validating the second accuracy of the plurality of second machine defect determination learning models using the data group, and verifying the third accuracy of the plurality of third machine defect determination learning models using the second data group of the original data set can do.

Here, verifying the accuracy means using the classification value output from each of a plurality of machine fault determination learning models using the first data group and the simulated abnormal data group among the original data set, and the second data group of the original data set to compare the classification values output from each of the plurality of machine defect determination learning models to analyze the degree of agreement.

In step S970, the optimal training method determination apparatus 100 determines the optimal ratio of the first data group and the simulated abnormal data group in the original data set for each feature extraction algorithm based on the result of verifying the accuracy of the machine fault determination learning models. And, it is possible to determine an optimal learning model among a plurality of machine defect discrimination learning models.

In this embodiment, the optimal training method determination apparatus 100 is a training with the lowest ratio of the first data group among the training data sets that derive more than a predetermined threshold accuracy from the result of verifying the accuracy of the machine fault determination learning models. A ratio between the first data group and the simulated abnormal data group in the data set may be determined as an optimal ratio. The optimal training method determination apparatus 100 may determine any one machine defect determination learning model having the highest accuracy among the first machine defect determination learning model to the third machine defect determination learning model as the optimum learning model. The optimal training method determination apparatus 100 uses any one feature extraction algorithm that generates training data to be input to the machine fault determination learning model having the highest accuracy among the first feature extraction algorithm to the third feature extraction algorithm, and the optimal feature extraction algorithm. can be determined as

Effective feature extraction algorithms and artificial intelligence models are different for each type of machine, and the ratio between original data and simulated abnormal data may be different. and a feature extraction algorithm may be selected (determined). Here, the selection may include any one of the above-described first to sixth ratios, and any one of a plurality of machine defect determination learning models outputting a minimum of 54 to a maximum of 180 classification values. may include, and any one algorithm among a plurality of feature extraction algorithms for extracting a minimum of 18 to a maximum of 60 features may be selected. In the ratio of the original data to the simulated data, increasing the size of the training data by adding the simulated data and decreasing the accuracy of the training data due to the addition of the simulated data may be a trade-off relationship. According to the determination of the optimal training method of the present invention, it is possible to find the maximum simulated data ratio that does not decrease the accuracy of the training data while increasing the size of the training data (or does not drop more than a certain amount).

The above-described embodiment according to the present invention 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, and 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 code generated by a compiler, but also high-level language code 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 the order is explicitly stated or there is no 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 limited by the claims. it is 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: optimal training method determination device 100
101: processor 102: memory
110: storage unit 120: loading unit
130: first generating unit 140: second generating unit
150: extraction unit 160: third generation unit
170: Verification unit 180: Determination unit
190: control unit

Claims (19)

A method of searching for an optimal training scheme, performed by a processor of an optimal training scheme determining device, comprising:
loading an original data set comprising a labeled normal data set and a labeled abnormal data set in relation to a vibration signal generated by the machine;
generating a fake abnormal data group by applying at least a portion of the labeled abnormal data set to a generative adversarial network (GAN);
generating a plurality of training data sets while adding data of the simulated abnormal data group to a first data group among the original data sets and changing the ratio between the first data group and the simulated abnormal data group;
extracting features by inputting the plurality of training data sets into a plurality of different feature extraction algorithms;
A plurality of learning models implemented with different algorithms are trained using the result of extracting features by input to the plurality of different feature extraction algorithms as training data. generating a plurality of machine defect determination learning models implemented with different algorithms to determine ;
verifying the accuracy of the plurality of machine defect determination learning models by using a second data group among the original data sets; and
Based on the result of verifying the accuracy of the machine defect determination learning model, the optimal ratio of the first data group and the simulated abnormal data group in the original data set for each feature extraction algorithm, and the plurality of machine defect determination learning models Including the step of determining the optimal learning model,
The determining step is
The first data group and the simulated abnormal data in the training data set having the lowest ratio of the first data group among the training data sets for deriving more than a predetermined threshold accuracy from the results of verifying the accuracy of the machine fault determination learning models determining the proportion of the group;
determining any one machine defect determination learning model having the highest accuracy among the plurality of machine defect determination learning models; and
Comprising the step of determining any one feature extraction algorithm that generates training data to be input to the machine fault determination learning model having the highest accuracy among the plurality of feature extraction algorithms,
How to determine the optimal training method. The method of claim 1,
The loading step is
labeling a first vibration signal generated in a normal machine as a normal state; and
Labeling the second vibration signal generated in the abnormal machine as a broken state,
How to determine the optimal training method. 3. The method of claim 2,
The vibration signal is
A horizontal vibration signal generated in the machine, a vertical vibration signal generated in the machine, a depth vibration signal generated in the machine, and horizontal, vertical and depth directions generated in the machine Including a vibration signal in the diagonal direction as a reference,
How to determine the optimal training method. The method of claim 1,
The step of creating the fake abnormal data group comprises:
generating simulated anomalous data from random noise generated using a Gaussian distribution; and
identifying whether the simulated anomalous data is true or false through comparison of at least a portion of the simulated anomalous data and the labeled anomalous data set;
How to determine the optimal training method. 4. The method of claim 3,
The step of extracting the feature is
extracting, as a first feature, a data group of a frequency domain generated by applying a first feature extraction algorithm to the plurality of training data sets;
extracting, as a second feature, a statistical analysis calculated by applying a second feature extraction algorithm to the data group of the frequency domain; and
and extracting a reduced-dimensional data group generated by applying a third feature extraction algorithm to the plurality of training data sets as a third feature,
How to determine the optimal training method. 6. The method of claim 5,
The first feature extraction algorithm,
It includes a fast Fourier transform (FFT) algorithm that converts a time domain vibration signal into a frequency domain vibration signal,
The second feature extraction algorithm,
Mean value, median value, minimum value of each of the vibration signal in the horizontal direction, the vibration signal in the vertical direction, the vibration signal in the depth direction, and the vibration signal in the diagonal direction converted into the frequency domain An algorithm for calculating one or more of a value, a max value, a kurtosis value, a skewness value, a standard deviation value, and a peak-to-peak value; ,
The third feature extraction algorithm,
comprising a principal component analysis (PCA) algorithm that generates reduced-dimensional data;
How to determine the optimal training method. 6. The method of claim 5,
The step of generating the plurality of machine defect determination learning models,
A first learning model implemented with a first algorithm is trained using the first to third features as training data, respectively, and a plurality of first learning models for judging whether the machine is defective using the features of the vibration signal generated by the machine 1 generating a machine fault determination learning model;
A second learning model implemented as a second algorithm is trained using the first to third features as training data, respectively, and a plurality of second learning models that determine whether the machine is defective using the features of the vibration signal generated by the machine 2 generating a machine fault determination learning model; and
A third learning model implemented as a first algorithm is trained using the first to third features as training data, respectively, and a plurality of third learning models that determine whether the machine is defective using the features of the vibration signal generated by the machine 3 comprising the step of generating a machine fault determination learning model,
How to determine the optimal training method. 8. The method of claim 7,
The step of verifying the accuracy is
verifying first accuracy for the plurality of first machine defect determination learning models by using a second data group of the original data set;
verifying second accuracy of the plurality of second machine defect determination learning models by using a second data group among the original data sets; and
using a second data group of the original data set to verify a third accuracy for the plurality of third machine defect determination learning models,
How to determine the optimal training method. delete A computer-readable recording medium storing a computer program for executing the method of any one of claims 1 to 8 using a computer. As a device for searching for an optimal training method,
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, the processor loads an original data set comprising a labeled normal data set and a labeled abnormal data set in relation to a vibration signal generated in the machine;
generating a fake abnormal data group by applying at least a portion of the labeled abnormal data set to a generative adversarial network (GAN);
generating a plurality of training data sets by changing the ratio of the first data group to the simulated abnormal data group while adding data of the simulated abnormal data group to a first data group among the original data sets;
extracting features by inputting the plurality of training data sets into a plurality of different feature extraction algorithms;
A plurality of learning models implemented with different algorithms are trained using the result of extracting features by input to the plurality of different feature extraction algorithms as training data. Create a plurality of machine fault determination learning models implemented with different algorithms to determine
verifying the accuracy of the plurality of machine fault determination learning models by using a second data group among the original data sets;
Based on the result of verifying the accuracy of the machine defect determination learning model, the optimal ratio of the first data group and the simulated abnormal data group in the original data set for each feature extraction algorithm, and the plurality of machine defect determination learning models stores the code that causes to determine the optimal learning model,
The memory causes the processor to
When performing the determination, the first in the training data set in which the ratio of the first data group is the lowest among the training data sets for deriving more than a predetermined threshold accuracy from the result of verifying the accuracy of the machine fault determination learning models. determining the ratio of the data group to the simulated abnormal data group;
Determining any one machine defect determination learning model having the highest accuracy among the plurality of machine defect determination learning models,
Storing a code causing to determine one of the feature extraction algorithms that generated training data to be input to the machine fault determination learning model with the highest accuracy among the plurality of feature extraction algorithms,
Optimal training method judgment device. 12. The method of claim 11,
The memory causes the processor to
When loading the original data set, code that causes a first vibration signal generated by a normal machine to be labeled as a normal state and a second vibration signal generated from an abnormal machine to be labeled as a broken state. to save,
Optimal training method judgment device. 13. The method of claim 12,
The memory causes the processor to
The vibration signal includes a horizontal vibration signal generated in the machine, a vertical vibration signal generated in the machine, a depth direction vibration signal generated in the machine, and horizontal and vertical vibration signals generated in the machine. storing a code causing to include a vibration signal in a diagonal direction based on a direction and a depth direction,
Optimal training method judgment device. 12. The method of claim 11,
The memory causes the processor to
When generating the fake anomalous data group, simulated anomalous data is generated from random noise generated using a Gaussian distribution,
storing code that causes the comparison of at least a portion of the simulated anomalous data and the labeled anomalous data set to identify whether the simulated anomalous data is true or false;
Optimal training method judgment device. 14. The method of claim 13,
The memory causes the processor to
When extracting the feature, a data group of a frequency domain generated by applying a first feature extraction algorithm to the plurality of training data sets is extracted as a first feature,
extracting a statistical analysis calculated by applying a second feature extraction algorithm to the data group of the frequency domain as a second feature;
storing a code that causes a reduced-dimensional data group generated by applying a third feature extraction algorithm to the plurality of training data sets to be extracted as a third feature,
Optimal training method judgment device. 16. The method of claim 15,
The memory causes the processor to
The first feature extraction algorithm includes a fast Fourier transform (FFT) algorithm for converting a time domain vibration signal into a frequency domain vibration signal,
The second feature extraction algorithm includes an average value of each of the vibration signal in the horizontal direction, the vibration signal in the vertical direction, the vibration signal in the depth direction, and the vibration signal in the diagonal direction converted into the frequency domain. one of median value, min value, max value, kurtosis value, skewness value, standard deviation value, and peak-to-peak value an algorithm for calculating an anomaly;
wherein the third feature extraction algorithm stores code that causes it to include a principal component analysis (PCA) algorithm that generates reduced-dimensional data;
Optimal training method judgment device. 16. The method of claim 15,
The memory causes the processor to
When generating the plurality of machine defect discrimination learning models, the first learning model implemented by the first algorithm is trained using the first to third features as training data, respectively, and the characteristics of the vibration signal generated by the machine to generate a plurality of first machine defect determination learning models that determine whether the machine is defective,
A second learning model implemented as a second algorithm is trained using the first to third features as training data, respectively, and a plurality of second learning models that determine whether the machine is defective using the features of the vibration signal generated by the machine 2 Create a machine fault determination learning model,
A third learning model implemented as a first algorithm is trained using the first to third features as training data, respectively, and a plurality of third learning models that determine whether the machine is defective using the features of the vibration signal generated by the machine 3 to store the code that causes the machine fault determination learning model to be created,
Optimal training method judgment device. 18. The method of claim 17,
The memory causes the processor to
When verifying the accuracy, verifying the first accuracy for the plurality of first machine defect determination learning models by using a second data group among the original data set,
verifying the second accuracy of the plurality of second machine defect determination learning models by using a second data group among the original data sets;
storing code that causes to verify a third accuracy for the plurality of third machine fault determination learning models using a second data group of the original data set;
Optimal training method judgment device. delete

Images