KR20210048100A - Condition monitoring data generating apparatus and method using generative adversarial network - Google Patents

Abstract

Disclosed is an apparatus for condition monitoring data generation which uses machine learning-based artificial intelligence. An objective of the present invention is to generate reliable fault condition data. The apparatus for condition monitoring data generation comprises a computing system. The computing system includes a generative adversarial network and a processor. The processor generates first training input data for a discriminator neural network by using a generator neural network which receives an initialized first input. The processor transfers data actually generated in the machine equipment as second training input data for the discriminator neural network to the discriminator neural network, and uses a second loss function resulting from combining a local feature function, which verifies whether the data match physical features of the condition monitoring data and the machine equipment, with a first loss function for discriminating the first training input data from the second training input data to train the discriminator neural network and the generator neural network when the local feature function is defined.

Description

[CONDITION MONITORING DATA GENERATING APPARATUS AND METHOD USING GENERATIVE ADVERSARIAL NETWORK}

The present invention relates to the field of monitoring the state of machine equipment using artificial intelligence based on machine learning. Specifically, it relates to a method for more effectively monitoring the state of machine equipment and a method for handling the data set by using a machine learning-based artificial neural network including a generative adversarial neural network.

Among the machine learning application cases in the industry, the classification task is mainly used for diagnosis in the medical and engineering fields. Machine learning techniques, including deep learning, require a sufficient number of balanced, labeled data to learn this classification problem.

A large amount of data on normal and abnormal/abnormal states is required in order to accurately diagnose the operation state of machinery and parts, and to implement a technology that detects or predicts abnormal behavior with an artificial neural network. However, in general, normal state data and abnormal state data occur disproportionately, and it is very difficult to secure abnormal state data to a balanced degree.

A generative adversarial network (GAN) has recently been proposed as one of the breakthrough research achievements in the field of artificial neural networks. GAN is drawing attention in that it generates data that is very similar to actual data by processing non-existent data.

However, in order for GAN to be applied to each industrial field, it needs to be optimized for the purpose of that industrial field.

As a conventional technology using GAN, a conventional technology used for diagnosing abnormal conditions of mechanical equipment and parts has not been specified yet, and no case of applying GAN has been found in the industrial field.

As a prior art that does not use GAN, another attempt to diagnose the state of mechanical equipment and parts is a method of using accelerated life test (ALT) data. This method is to predict the average life of the system or the remaining life of parts by operating under operating conditions that are more severe than field operating conditions (actual operating conditions).

For the purpose of predicting the remaining life of the system using the accelerated life test data, as a technology for generating data using an artificial neural network, Korean Patent No. 10-1925480 "Operating load data for failure prediction based on accelerated life test data" Production method and apparatus" and the like. The prior art creates a data set from existing data using a plurality of regression models. The above prior art is a method of predicting the remaining life when operating under actual field operating conditions by applying a regression model to the accelerated life test data, and the deterioration pattern of the parts shown in the accelerated life test data is applied to the field operating conditions as it is. Since there is no way to verify whether or not the data is reliable, the generated data is placed only on the extension line of the accelerated life test data, so scalability is poor and the data imbalance problem remains.

The types of state data that can be obtained from a running mechanical device include a normal state and an abnormal/abnormal state in which various failures occur. However, since data is rarely acquired when a failure occurs, a testbed, which is a similar device that mimics actual machine equipment, is set up or manufactured to learn a machine learning model, and the failure is artificially approved or simulated. A method of acquiring fault status data can be used.

In the present invention, from the normal operation data of the actual mechanical equipment, the normal operation data of the test bed, and the abnormal operation data of the test bed, a latent space in which abnormal operation data of the actual mechanical equipment may exist is set, Data belonging to a latent space can be simulated based on an arithmetic rule.

In the present invention, it is an object of the present invention to generate reliable failure state data by applying data simulated in a latent space to a GAN along with verification of physical characteristics.

In the present invention, a modified GAN network is proposed to utilize the conventional generative hostile neural network (GAN) technology for the purpose of generating state monitoring data of machine equipment. The modified GAN network may include a verification process for physical characteristics, and may be trained to generate data similar to patterns of actual data.

The present invention was derived as a means for solving the problems of the prior art, and the state monitoring data generating apparatus according to an embodiment of the present invention may be implemented in the form of a computing system. The state monitoring data generating apparatus generates state monitoring data of machine equipment using machine learning-based artificial intelligence, and the computing system includes a generative adversarial neural network and a processor. Generative adversarial neural networks include generator neural networks and discriminator neural networks. The processor generates first training input data for the discriminator neural network using the generator neural network receiving the initialized first input. An example of the initialized first input may be random noise.

The processor transmits data actually generated in the machine equipment as second training input data for the discriminator neural network to the discriminator neural network, and verifies whether the physical characteristics of the machine equipment and the state monitoring data are met. When a feature function is defined, the discriminator neural network and the generator neural network are made using a second loss function in which the local feature function is synthesized to a first loss function for discriminating the first training input data from the second training input data. To train.

In this case, the mechanical equipment may include real mechanical equipment and a test bed imitating the real mechanical equipment.

At this time, the processor includes a first data domain corresponding to first normal state data obtained by a normal operation of the test bed, a second data domain corresponding to first abnormal state data obtained by an abnormal operation of the test bed, and the Potential corresponding to the second abnormal state data obtained by the abnormal operation of the actual mechanical equipment by arithmetic operation based on the association between the third data domain corresponding to the second steady state data obtained by the normal operation of the actual mechanical equipment. Second input data belonging to the fourth data domain, which is a data space, may be generated.

In this case, the processor transmits the second input data to the trained generator neural network as an input of the generator neural network of the generative hostile neural network, and the trained generator neural network and the trained discriminator neural network Output data determined through inference may be generated as state monitoring data of the actual machine equipment belonging to the fourth data domain.

In this case, the processor may generate output data determined through inference of the trained generator neural network and the trained discriminator neural network as the state monitoring data of the machine equipment. In this case, the output data may be used as input data for machine learning of another artificial neural network to monitor the state of the machine equipment.

In this case, the local feature function may be a function that quantifies a correlation with local feature data based on a physical feature of a target data domain to be generated as the state monitoring data of the mechanical equipment. For example, the local feature data may be data based on a gear mesh frequency (GMF) of vibration data generated from a gear included in the mechanical equipment. In addition, the second loss function may be defined as a weight summation between the first loss function and a negative value of the local feature function.

In this case, the processor may generate the initialized first input based on random noise and transmit it to the generator neural network as an input for training the generator neural network. In addition, the processor may generate the second input data as a random number belonging to the fourth data domain, and transmit the second input data as an input of the trained generator neural network as an input of the trained generator neural network. .

A method of generating state monitoring data according to an embodiment of the present invention generates state monitoring data of machine equipment using machine learning-based artificial intelligence, which is executed by a computing system. The state monitoring method of the present invention comprises the steps of: generating, by the processor, first training input data for the discriminator neural network by using the generator neural network receiving an initialized first input; The processor transmitting data actually generated in the machine equipment as second training input data for the discriminator neural network to the discriminator neural network, and the processor is configured to provide physical characteristics of the machine equipment and the state monitoring data. When a local feature function for verifying conformance is defined, a second loss function in which the local feature function is synthesized to a first loss function for discriminating the first training input data from the second training input data is used. And training the discriminator neural network and the generator neural network.

According to the present invention, the type of state data that can be obtained from a mechanical device in operation resolves asymmetry and imbalance between a normal state and an abnormal/abnormal state in which various failures occur, and data of abnormal/abnormal states that may occur in an actual mechanical device are stored. By simulating it is possible to generate well-balanced labeled data.

According to the present invention, it is possible to generate balanced data necessary for learning for fault diagnosis for a machine device by using machine learning, and thereby, it is possible to expand the application range of machine learning for fault diagnosis for a machine device.

According to the present invention, from the normal operation data of the actual mechanical equipment, the normal operation data of the test bed, and the abnormal operation data of the test bed, a latent space in which abnormal operation data of the actual mechanical equipment may exist is set. , It is possible to simulate data belonging to a latent space based on an arithmetic rule.

According to the present invention, it is possible to generate highly reliable failure state data by applying data simulated in a latent space to a GAN along with verification of physical characteristics.

According to the present invention, it is possible to implement a modified GAN network for utilizing the conventional generative hostile neural network (GAN) technology for the purpose of generating state monitoring data of machine equipment. The modified GAN network of the present invention may include a verification process for physical characteristics, and may be trained to generate data similar to a pattern of actual data.

According to the present invention, by using the abnormal state industrial data generation technique, it is possible to acquire the abnormal state industrial data at the actual system level, which is very scarce in the field. The abnormal state industrial data obtained in this way is used to develop accurate industry information prediction and management algorithms, through which 1) maximization of productivity through predictive diagnosis of the health of manufacturing facilities (near-zero downtime), and 2) products through outgoing product quality inspection Reliability improvement (near-zero defect), 3) facility energy efficiency maximization (near-zero waste) can be achieved.

According to the present invention, unobserved data is generated using a generative hostile neural network (GAN), and an abnormal state of an actual machine equipment can be learned in advance using the newly generated data, and the occurrence of a failure can be predicted. .

According to the present invention, a potential space of data can be reconstructed according to the user's intention in order to generate suitable industrial data using a generative hostile neural network (GAN).

1 is a diagram illustrating a computing system including a generative adversarial neural network (GAN) using artificial intelligence based on machine learning according to an embodiment of the present invention.
2 is a diagram illustrating a generative adversarial neural network (GAN) using artificial intelligence based on machine learning according to an embodiment of the present invention.
FIG. 3 is a diagram schematically illustrating data generated in an embodiment of the present invention using the GAN of FIG. 2 in a data space.

In addition to the above objects, other objects and features of the present invention will become apparent through the description of the embodiments with reference to the accompanying drawings.

A preferred embodiment of the present invention will be described in detail with reference to the accompanying drawings. In describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the subject matter of the present invention, a detailed description thereof will be omitted.

The present invention is a technology for generating data used for condition monitoring of machinery parts (bearings, gears, motors, etc.). The generated state monitoring data may include data such as infrared images, time series acceleration, gas concentration, humidity, flow rate, pressure, concentration ratio, torque, and position signals. That is, all information generated during operation of mechanical equipment or parts can be used as condition monitoring data, and vibration, noise, and changes in appearance of mechanical equipment or parts are all information indicating the condition.

Among the application cases of machine learning in the industry, the classification task is mainly used for diagnosis in the medical and engineering fields. Machine learning technologies, including deep learning, require a sufficient number of balanced, labeled data to learn these classification problems. Therefore, in order to accurately diagnose and predict the condition of machinery parts and make technology, a large amount of normal and abnormal state data is required. However, it is difficult to learn a diagnostic model because the abnormal state data is relatively insufficient compared to the steady state data. Accordingly, the present invention proposes a method of using generative adversarial networks (GANs) to generate scarce data, particularly abnormal state data.

The types of state data that can be obtained from a running mechanical device include various fault states in addition to normal operation. However, since it is very rare to acquire data when an actual failure occurs, generating failure data in order to diagnose or predict a failure of a mechanical device in the future is one of the very important themes in the technical field of the present invention.

In order to generate fault data, a method of obtaining fault state data by artificially applying or simulating a fault by creating a similar device (testbed) is mainly used. However, as in the preceding document KR 10-1925480 cited above, since the test bed is only a test bed and is not a mechanical device itself to be diagnosed/predicted, a method of obtaining virtual data of a mechanical device based on the data obtained from the test bed. There may be several types of data, but in order to apply the virtual data thus obtained, verification of the validity of the virtual data is required.

In the present invention, failure data of a mechanical device obtained by expanding previously generated data is solved through the application of a generative adversarial neural network to verify the verification process by physical laws and pattern analysis.

In the present invention, a method of generating failure state data is based on an arithmetic rule in a latent space of GANs. If the system that is easy to obtain normal and fault (or abnormal) data, that is, the test bed is called A, and the system where it is difficult to obtain failure data, that is, the actual equipment is called B, then A is 1) normal and 2) fault, B is 3) normal. , B's fault data can be generated by using the data of three states.

Specifically, the present invention consists of two steps including firstly learning a generator and secondly generating data. In the learning process of the generator neural network, the generator neural network learns a criterion for filtering out data that violates physical laws or is inappropriate.

1 is a diagram illustrating a computing system 100 including a generative hostile neural network (GAN) 120 using artificial intelligence based on machine learning according to an embodiment of the present invention.

Referring to FIG. 1, the computing system 100 includes a processor 110 and a generative adversarial neural network 120. The generative adversarial neural network 120 is stored in memory or storage and learns the weight of the edges between each internal node through interaction with the processor 110, and based on the learned weight. Perform inference on the new input data.

2 is a diagram illustrating a generative adversarial neural network (GAN) 120 using artificial intelligence based on machine learning according to an embodiment of the present invention.

Referring to FIG. 2, the generative hostile neural network 120 of FIG. 1 includes a generator/generator neural network 210, a discriminator network 220, and a discriminator neural network 220. The first loss function 240, the local correlation 230 to which the data generated from the local feature data 232 is applied, the local feature loss function 250 resulting from the local correlation 230, and the first loss function 240 And a second loss function 260 in which the local feature loss function 250 is synthesized.

The processor 110 controls the generator neural network 210 and the discriminator neural network 220 of the generative adversarial neural network 120 to allow the generative adversarial neural network 120 to generate state monitoring data. 120).

The generator neural network 210 receives the random noise 214 as an initialized first input and generates first training input data 216. The first training input data 216 is one of inputs for the discriminator neural network 220 and is a candidate for data to be generated in the future.

Training data (Training Set) 212 generated based on data actually generated in the machine equipment is transmitted to the discriminator neural network 220 as second training input data for the discriminator neural network 220.

In general, conventional GANs can elaborately generate data, but this process requires trial and error such as structural modification and adjustment of learning parameters. As in the present invention, when data of an unlearned domain is generated, data that does not match actual physical phenomena may be generated. Therefore, based on physical facts, the generative adversarial neural network 120 of the present invention defines “Local Feature Data” 232 that enables data to be learned, and thus the generator neural network 210 and the discriminator neural network ( 220) is reflected in the loss function. The local feature data 232 is composed of representative data that well reflects the physical features of the domain to be generated. From the local feature data 232, a local metric that reflects the physical feature of the data domain to be generated may be additionally derived. In the embodiment of the present invention, vibration data generated from a gear of a gearbox is taken as an example. The vibration generated by the gear generates a unique frequency called the gear mesh frequency (GMF), and the spectrum of this frequency occupies the largest portion of the entire spectrum. The local feature data 232 is data of local features that contain physical features specific to this domain. The second loss function 260 of the generative adversarial neural network 120 for learning is a loss function of conventional GANs (e.g., a combination of only the generator neural network 210 and the discriminator neural network 220) It consists of the sum of the negative values of the Local Correlation 230 of the generated data for the local feature data 232 to the first loss function 240. A weighted sum parameter is required in order to generate the second loss function 260 as one function by adding the two loss functions 240 and 250. Weights applied to each of the first loss function 240 and the local feature loss function 250 may be individually optimized. The weight may be determined in consideration of physical characteristics included in the local feature loss function 250.

When the discriminator neural network 220 included in the generative adversarial neural network expands by arithmetic operation using the first loss function 240, the training data (training set) 212 in which the pattern of the expanded data is a learning target That is, the generative adversarial neural network is controlled to be similar to the second training input data. In addition, the local feature loss function 150 and the second loss function 260 feed back whether the generated data 216 conforms to local features and physical laws, so that the generative adversarial neural network 120 can self-verify the generated data. It functions as a means of being.

3 is a diagram schematically illustrating data generated in an embodiment of the present invention using the GAN 120 of FIGS. 1 and 2 in a data space.

3, using a generative adversarial neural network 120 to expand data in data spaces 310 and 320, adjust the expanded data, and verify the adjusted data (verification based on physical characteristics) The process of doing is shown.

The generative hostile neural network 120 of the present invention can generate a random number as a candidate for new data in the latent space 320 based on the data actually observed in the observation space 310. have. At this time, the learned generative hostile neural network 120 may receive the generated random number and generate new data by the operation of the discriminator neural network 220 and the second loss function 260. In this case, the characteristics of data generated for each domain are unique in the latent space 320.

The mechanical equipment to be observed includes real mechanical equipment and a testbed that mimics the real mechanical equipment.

The data obtained in the observation space 310 is a first data domain corresponding to the first steady state data obtained by the normal operation of the test bed, and the first abnormal state data obtained by the abnormal operation of the test bed. It may be divided into a corresponding second data domain (testbed fault) and a third data domain (field normal) corresponding to second steady state data obtained by normal operation of the actual machine equipment.

In case the actual mechanical equipment deliberately operates abnormally, it does not occur frequently, and it is very difficult to accurately obtain the state information when the actual mechanical equipment operates abnormally. Therefore, it is difficult to obtain the data of the fourth data domain (Field fault) corresponding to the second abnormal state data obtained by the abnormal operation of the actual mechanical equipment, and in this regard, the data of the fourth data domain is in the observation space 310. It does not exist and is handled in the latent space 320.

As shown in FIG. 3, the first data domain, the second data domain, and the third data domain of the observation space 310 are input to the generative hostile neural network 120 as available data, and The hostile neural network 120 is trained/learned.

By learning of the generative hostile neural network 120, the second training input data 212 belonging to the first data domain, the second data domain, and the third data domain of the observation space 310 are in the latent space 320. It is adjusted to have an extended domain. In FIG. 3, for convenience of explanation, this process is described using the first potential data domain 322 in the latent space 320 of the first data domain. The generator neural network 210 of the generative hostile neural network 120 may widen the range of data that can be generated from the first potential data domain 322 to the first extended data domain 324 by using a random number. Since the generator neural network 210 is a technology that expands the range of the generated data from the basic data, it is expressed that the numerical range that the data can have on the latent space 320 is expanded. However, at this time, the data that can be generated by the local feature loss function 250 taking into account physical characteristics can be selectively generated only data conforming to the physical characteristics/local characteristics, and the similarity of the generated data to actual data can be further increased. have. In addition, a unique pattern of data is identified by the first loss function 240 of the discriminator neural network 220, and data having a unique pattern appearing in the first data domain is selectively generated as data that can be generated. Can be.

Likewise, for data in the second data domain, a potential domain of data that can be generated by the generative adversarial neural network 120 will be expanded on the latent space 320 by learning of the generative adversarial neural network 120. In this case as well, data reflecting unique patterns and physical/local features appearing in the second data domain will be selectively generated, and the same technical features are applied to the third data domain as well.

The processor 110 may generate new second input data belonging to the fourth data domain by an arithmetic operation based on the association between the first data domain, the second data domain, and the third data domain. A process of generating second input data belonging to the fourth data domain based on the association between the first data domain, the second data domain, and the third data domain may be implemented by various arithmetic operations. A method of obtaining data of the fourth data domain may be obtained using conventional machine learning/machine learning. For example, data of the fourth data domain may be obtained using supervised learning through acquisition of actual system simulation test bed data (first and second data domains), and actual system acquisition normal data (third data domain) Data belonging to the fourth data domain may be obtained using unsupervised learning using (data domain). However, the conventional supervised learning-based method has a limitation in that the algorithm is learned to be biased according to the characteristics of the test bed such as manufacturer, driving environment, and capacity. In addition, there is a concern that the conventional unsupervised learning method may misdiagnose a heterogeneous normal (a normal type different from the learned normal) as a failure.

The generative adversarial neural network 120 of the present invention can learn and expand various characteristics of data in a latent space 320 lower than the data dimension. Since the generative adversarial neural network 120 learns data smoother than conventional machine learning in the latent space 320, it supports arithmetic operation according to the characteristics of the data.

In addition, in the existing unsupervised learning, in the process of extracting features of real data and reconstructing the extracted features into data, the shape of the latent space is often random and the reconstructed data is far from reality. , The generative hostile neural network 120 of the present invention aims to learn to the extent that the discriminator neural network 220 cannot distinguish between fake data and real data generated by the generator neural network 210 Because it is designed, it can simulate data that is so similar that it is difficult to distinguish it from reality. The generative adversarial neural network 120 can generate very realistic data compared to a conventional Restricted Boltzmann Machine (RBM) or a denoising autoencoder (dAE). In addition, by including the local feature loss function 150 in consideration of the physical characteristics of the domain to which the data belongs in the generative adversarial neural network 120, the generated data conforms to the physical laws and can be used with high value for fault diagnosis. System fault data can be generated.

A well-formed latent space tends to align the features contained in the data well, and new data can be generated using an arithmetic operation of the coordinates of the latent space using this. One of the good examples that can be assumed through this invention is feature 1 (normal, fault), feature 2 (normal data and fault data artificially generated in the test bed, and phase data acquired from an actual industrial system). Test bed, industrial system) are each acquired and generated by acquiring the coordinates of the fault data of the industrial system in the latent space through arithmetic operation. Through this, fault data of an industrial system that is difficult to obtain can be generated and used for learning of a diagnostic model.

When the characteristics of the data generated for each domain are unique, and when it is known in which domain of the latent space 320 the data learned using this property occurs, new data is generated through arithmetic operations in the latent space 320 You can do it. Through arithmetic operation, the processor 110 may generate data of the fourth data domains 326 and 328 in the latent space 320. This arithmetic operation process may be executed inside the generative adversarial neural network 120 or by another generator neural network or generator logic inside the computing system 100.

As an example of such arithmetic operation, the invention of the prior art Korean Patent No. KR 10-1925480 "A method and apparatus for generating operating load data for failure prediction based on accelerated life test data" may be modified and used for the purpose of the present invention. . The prior literature may be modified and applied as part of the configuration of the present invention. However, as described above, the preceding literature focuses only on arithmetic and statistical relevance, so data expansion is impossible, and information on whether generated fake data is similar to real data, or whether actual data follows physical laws. It does not include it, and verification accordingly is also impossible. In contrast, the generative hostile neural network 120 of the present invention can be expanded within the latent space 320 of data, and uses a second loss function 160 in which two types of loss functions 140 and 150 are synthesized. By doing so, the generated data is similar enough to be difficult to distinguish from the actual data, and can be learned to follow the laws of physics.

Data of the fourth potential data domain 326 is generated by arithmetic operations, and after inputting the data of the fourth potential data domain 326 as second input data into the generative adversarial neural network 120, a generative adversarial neural network The data generated by 120 corresponds to the fourth extended domain 328. Data generated by the generative adversarial neural network 120 belongs to the fourth data domain and is generated as state monitoring data of actual machine equipment. Depending on the embodiment implementing the generative adversarial neural network 120, the output data of the generator neural network 210 based on the second input data may be generated as state monitoring data, and according to another embodiment, the generator neural network 210 ), the output data selected by inference of the discriminator neural network 220 may be generated as state monitoring data.

The processor 110 generates output data generated through inference of the trained generative adversarial neural network 120 as state monitoring data of the mechanical equipment, and the output data at this time is for monitoring the state of the mechanical equipment. It may be used as input data for machine learning of another artificial neural network (not shown).

The generative adversarial neural network 120 trained using the first data domain, the second data domain, and the third data domain as input data may generate data belonging to the extended data domain in the latent space 320. In this case, the discriminator neural network 220 may learn a discrimination criterion optimized for each data domain. Accordingly, even if the potential data domain is expanded in the latent space 320, the extended data domain 324 of the first data domain may be controlled so as not to invade another second data domain or a third data domain. Since the discriminant neural network 220 is determined by learning, the fourth data domain 328 is extended/generated from the data of the fourth potential data domain 326 because the fourth extended data domain 328 is different from the first data domain. It can be controlled not to invade the data domain, the second data domain, or the third data domain.

The local feature loss function 250 may be selected as a function for quantifying a correlation with local feature data based on a physical feature of a target data domain to be generated as state monitoring data of a machine equipment. In addition, the local feature loss function 250 may be defined in the form of a local metric in which physical characteristics appearing in each of the first data domain, the second data domain, the third data domain, and the fourth data domain are individually optimized. May be.

As a modified embodiment of the present invention, the processor 110 may generate the initialized first input based on random noise and transmit it to the generator neural network 210 as an input for training the generator neural network 210. In this case, the processor 110 may generate the second input data based on a random number belonging to the fourth data domain, and may transmit the second input data as an input of the trained generator neural network 210.

The apparatus and method for generating condition monitoring data using machine learning-based artificial intelligence of the present invention proposes a generative adversarial neural network 120 adapted to be suitable for generating condition monitoring data. In addition, the generative adversarial neural network 120 of the present invention can expand data in the latent space 320 and obtain failure data of actual mechanical equipment from unobserved or very insufficient data. Imbalance and asymmetry can be resolved. In addition, the generative adversarial neural network 120 of the present invention does not simply expand data by arithmetically combining it, but learns the data pattern of each data domain to selectively generate data that is difficult to distinguish from actual data. In this process, by additionally applying a loss function reflecting the physical characteristics appearing in each data domain, the generated data follows the physical laws, and data very similar to reality can be generated.

The state monitoring data generation method using machine learning-based artificial intelligence according to an embodiment of the present invention may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -A hardware device specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

However, the present invention is not limited or limited by the embodiments. The same reference numerals shown in each drawing indicate the same members. The length, height, size, width, etc. introduced in the embodiments and drawings of the present invention may be exaggerated to aid understanding.

As described above, in the present invention, specific matters such as specific components, etc., and limited embodiments and drawings have been described, but this is provided only to help a more general understanding of the present invention, and the present invention is not limited to the above embodiments. , If a person of ordinary skill in the field to which the present invention belongs, various modifications and variations are possible from these descriptions.

Therefore, the spirit of the present invention is limited to the described embodiments and should not be determined, and all things that are equivalent or equivalent to the claims as well as the claims to be described later fall within the scope of the present invention. .

100: computing system
110: processor
120: generative hostile neural network
210: generator neural network 220: discriminator neural network
240: first loss function 260: second loss function

Claims (12)

As a state monitoring data generating device that generates state monitoring data of machine equipment using machine learning-based artificial intelligence, the state monitoring data generating device includes a computing system, and the computing system is a generative adversarial network (GAN). Network) and a processor,
The generative adversarial neural network includes a generator neural network and a discriminator network,
The processor is
Generating first training input data for the discriminator neural network using the generator neural network receiving the initialized first input,
Transmitting data actually generated in the machine equipment to the discriminator neural network as second training input data for the discriminator neural network,
If a local feature function for verifying whether the mechanical equipment and the condition monitoring data conforms to the physical feature is defined, the local feature in a first loss function for discriminating the first training input data from the second training input data A state monitoring data generation apparatus for training the discriminator neural network and the generator neural network using a second loss function from which a function is synthesized. The method of claim 1,
The mechanical equipment includes real mechanical equipment and a test bed imitating the real mechanical equipment,
The processor is
A first data domain corresponding to first normal state data obtained by the normal operation of the test bed, a second data domain corresponding to first abnormal state data obtained by an abnormal operation of the test bed, and the actual machine equipment This is a potential data space corresponding to the second abnormal state data obtained by the abnormal operation of the actual machine equipment by arithmetic operations based on the association between the third data domains corresponding to the second steady state data obtained by the normal operation of 4 Create second input data belonging to the data domain,
Transferring the second input data to the trained generator neural network as input to the generator neural network of the generative hostile neural network,
State monitoring data for generating output data determined through inference of the trained generator neural network and the trained discriminator neural network as state monitoring data of the actual machine equipment belonging to the fourth data domain Generating device. The method of claim 1,
The processor generates output data determined through inference of the trained generator neural network and the trained discriminator neural network as the state monitoring data of the machine equipment,
The output data is used as input data for machine learning of another artificial neural network for monitoring the state of the machine equipment. The method of claim 1,
The local feature function is
A condition monitoring data generating device that is a function for quantifying a correlation with local feature data based on a physical feature of a target data domain to be generated as the condition monitoring data of the mechanical equipment. The method of claim 4,
The local feature data is data based on a gear mesh frequency (GMF) of vibration data generated from a gear included in the mechanical equipment. The method of claim 1,
The second loss function is defined as a weight summation between the first loss function and a negative value of the local feature function. The method of claim 2,
The processor generates the initialized first input based on random noise and transfers the initialized first input to the generator neural network as an input for training the generator neural network,
A state monitoring data generation device that generates the second input data as a random number belonging to the fourth data domain, and transmits the second input data as an input of the trained generator neural network to an input of the trained generator neural network . As a condition monitoring data generation method that generates condition monitoring data of machine equipment using machine learning-based artificial intelligence executed by a computing system,
The computing system includes a generative adversarial network (GAN) and a processor, and the generative adversarial network includes a generator neural network and a discriminator network,
The processor generating first training input data for the discriminator neural network using the generator neural network for receiving the initialized first input;
The processor transferring data actually generated in the machine equipment as second training input data for the discriminator neural network to the discriminator neural network, and
The processor, when a local feature function for verifying whether the mechanical equipment and the state monitoring data conforms to the physical feature is defined, a first loss function for discriminating the first training input data from the second training input data Training the discriminator neural network and the generator neural network using a second loss function from which the local feature function is synthesized;
State monitoring data generation method comprising a. The method of claim 8,
The mechanical equipment includes real mechanical equipment and a test bed imitating the real mechanical equipment,
The processor includes a first data domain corresponding to first normal state data obtained by a normal operation of the test bed, a second data domain corresponding to first abnormal state data obtained by an abnormal operation of the test bed, and the actual Potential data corresponding to the second abnormal state data obtained by the abnormal operation of the actual mechanical equipment by arithmetic operations based on the association between the third data domain corresponding to the second steady state data obtained by the normal operation of the mechanical equipment of Generating second input data belonging to a fourth data domain that is a space;
Transmitting, by the processor, second input data to the trained generator neural network as input to the generator neural network of the generative adversarial neural network; And
The processor generates output data determined through inference of the trained generator neural network and the trained discriminator neural network as state monitoring data of the actual machine equipment belonging to the fourth data domain. step;
State monitoring data generation method further comprising a. The method of claim 8,
Generating, by the processor, output data determined through inference of the trained generator neural network and the trained discriminator neural network as the state monitoring data of the machine equipment;
Including more,
The output data is used as input data for machine learning of another artificial neural network to monitor the state of the machine equipment. The method of claim 8,
The local feature function is
It is a function for quantifying the association with local feature data based on the physical feature of the target data domain to be generated as the state monitoring data of the machine equipment,
The local feature data is data based on a gear mesh frequency (GMF) of vibration data generated from a gear included in the mechanical equipment. The method of claim 8,
Generating the first training input data comprises:
The initialized first input is generated based on random noise and transmitted to the generator neural network as the first training input data for training the generator neural network,
Transmitting second input data to the trained generator neural network as an input to the generator neural network of the generative hostile neural network
A state monitoring data generation method for generating the second input data as a random number belonging to the fourth data domain, and transmitting the second input data as an input of the trained generator neural network to the input of the trained generator neural network .

Images