WO2023120888A1

WO2023120888A1 - Predictive maintenance method, for device, using gan algorithm

Info

Publication number: WO2023120888A1
Application number: PCT/KR2022/014427
Authority: WO
Inventors: 이영규
Original assignee: 주식회사 아이티공간
Priority date: 2021-12-23
Filing date: 2022-09-27
Publication date: 2023-06-29
Also published as: KR20230096364A

Abstract

The present invention relates to a predictive maintenance method, for a device, using a GAN algorithm, the method in which a decision module learns a large amount of normal image information, which is collected when the device is in a normal state, and a large amount of quasi-defect image information which is generated by a generation module on the basis of defect image information collected from the device before a malfunction occurs, thereby deciding whether real-time image information, which is collected from the device operating in real time, is a normal image or a defect image. If the real-time image information is decided to be a defect image, the method detects the abnormal state of the device and provides a guide to perform maintenance and replacement of the device at an appropriate time, thereby preventing huge losses due to a malfunction of the device.

Description

Method for predictive maintenance of devices using GAN algorithm

The present invention relates to a method for predictive maintenance of a device using a GAN algorithm, and more particularly, a determination module includes a large amount of normal image information collected when the device is in a normal state and a defective image collected from the device before a failure occurs in the generation module. Real-time image information collected from real-time driven devices is determined as a normal or defective image by learning a large amount of similar defective image information generated based on the information. It relates to a predictive maintenance method for devices using a GAN algorithm that can prevent huge losses due to device failures in advance by inducing maintenance and replacement of devices at an appropriate time.

In general, stable operation of various devices used for the automation process of facilities is very important.

For example, dozens or hundreds of devices are installed in facilities of a large-scale production plant to continuously produce products while interlocking with each other. situation may arise.

In this case, not only equipment repair costs due to downtime due to equipment failure, but also huge losses due to wasted operating costs and business effects while the equipment is stopped.

According to recent data from the Ministry of Employment and Labor and the Korea Occupational Safety and Health Administration, the number of casualties due to industrial safety accidents annually is estimated to be around 100,000, and when converted into costs, it is estimated that an annual loss of 18 trillion won is occurring.

As a way to avoid such unexpected downtime costs, it is urgent to introduce a predictive maintenance system.

The present invention has been proposed to solve various problems as described above, and its purpose is that the determination module is a large amount of normal image information collected when the device is in a normal state and a bad image collected from the device before a failure occurs in the generation module. Real-time image information collected from real-time driven devices is determined as a normal or defective image by learning a large amount of similar defective image information generated based on the information. It is to provide a predictive maintenance method for devices using a GAN algorithm that can prevent huge losses due to device failures in advance by inducing maintenance and replacement of devices at an appropriate time.

In addition, the discrimination module increases the learning effect of the discrimination module by relearning by receiving feedback on the success or failure of the discrimination result in the learning stage and the relearning stage, thereby securing excellent reliability for the discrimination result determined by the discrimination module. It is to provide a method for predictive maintenance of a device using a GAN algorithm.

The predictive maintenance method of a device using the GAN algorithm according to the present invention to achieve the above object is to perform a process in a normal operating state. A normal information collection step (S10) in which at least one or more are collected and converted into an image file, and when two or more normal waveforms are collected, each normal waveform is superimposed on each other and then converted into an image file and collected as a normal image; Before the occurrence, a process is performed while the device is running, and at least one defective waveform representing the change information of the amount of energy over time measured by the device is collected, converted into an image file, and collected. Once collected, a defect information collection step (S20) of overlapping the defective waveforms with each other and then converting them into image files and collecting them as defective images; and learning the defective image information collected in the defect information collection step (S20), Based on the learned bad image information, a creation module mass-produces similar bad images similar to the bad images (S30); and, the normal image information collected in the normal information collection step (S10) and the good work generation A learning step (S40) of learning the similar defective image information produced in step (S30) in the discrimination module; and, a process performed in a real-time driving state, showing the change information of the amount of energy over time measured by the device At least one real-time waveform is collected, converted into an image file, and collected. When two or more real-time waveforms are collected, each real-time waveform is superimposed on each other, converted into an image file, and collected as a real-time image, and the real-time image is collected in the learning step ( A determination step (S50) of determining whether the determination module through S40) is normal or defective; and, if the real-time image collected from the real-time device is determined to be defective through the determination module in the determination step (S50), the device is abnormal. It is characterized in that it includes; a detection step (S60) of detecting the state.

In addition, the learning step (S40) includes a first step (S41) of learning about a normal image based on a large amount of normal image information collected in the normal information collection step (S10) by the discrimination module; In (S41), the determination module that has learned about the normal image is provided with a large amount of similar defective images generated in the generation module together with the normal image, and the image determined as a normal image through the determination module is normal, and other In the second process (S42) of determining that the image is defective, the success or failure of the result determined by the discrimination module in the second process (S42) is fed back to the discrimination module to maximize the learning efficiency of the discrimination module. , and a third step (S43) of learning whether or not the determination module's determination result for similar defective images was successful by receiving feedback from the generating module.

In addition, the determination module re-learns by receiving feedback on the success or failure of the determination result of the real-time image determined in the determination step (S50), thereby inducing reliability of the determination result determined in the determination step (S50) to be improved. Characterized in that it further includes; learning step (S70).

In addition, the energy according to the time measured by the device is any one selected from the current consumed in driving the device, the vibration or noise generated when the device is driven, the frequency of power supplied to the device, and the temperature, humidity, and pressure of the device when driven. It is characterized by using one or a combination of two or more.

As described above, according to the method for predictive maintenance of a device using the GAN algorithm according to the present invention, the determination module includes a large amount of normal image information collected when the device is in a normal state and a bad image collected from the device before a failure occurs in the generation module. Real-time image information collected from real-time driven devices is determined as a normal or defective image by learning a large amount of similar defective image information generated based on the information. By inducing maintenance and replacement of devices at an appropriate time, there is an effect of preventing huge losses due to device failures in advance.

In addition, the discrimination module increases the learning effect of the discrimination module by relearning by receiving feedback on the success or failure of the discrimination result in the learning stage and the relearning stage, thereby securing excellent reliability for the discrimination result determined by the discrimination module. There is an effect.

1 is a block diagram of a method for predictive maintenance of a device using a GAN algorithm according to an embodiment of the present invention.

2 to 6 are diagrams for explaining a method for predictive maintenance of a device using the GAN algorithm shown in FIG. 1 .

A method for predictive maintenance of a device using a GAN algorithm according to a preferred embodiment of the present invention will be described in detail based on the accompanying drawings. Detailed descriptions of well-known functions and configurations that may unnecessarily obscure the subject matter of the present invention will be omitted.

1 to 6 show a method for predictive maintenance of a device using a GAN algorithm according to an embodiment of the present invention, and FIG. 1 is a block diagram of a method for predictive maintenance of a device using a GAN algorithm according to an embodiment of the present invention. , FIGS. 2 to 6 each show diagrams for explaining a method for predictive maintenance of a device using the GAN algorithm shown in FIG. 1 .

As shown in the figure, the predictive maintenance method 100 of a device using the GAN algorithm according to an embodiment of the present invention includes a normal information collection step (S10), a bad information collection step (S20), and a good product generation step ( S30), a learning step (S40), a determination step (S50), and a detection step (S60).

As shown in FIG. 1, the normal information collection step (S10) performs a process in a normal driving state, and collects at least one normal waveform representing the change information of the amount of energy over time measured by the device Conversion into image files is performed, but if two or more normal waveforms are collected, each normal waveform is overlapped with each other, converted into an image file, and collected as a normal image.

As shown in FIG. 2, in the present invention, for convenience of description, the energy measured by the device performing the process, the current consumed in driving the device, the vibration generated in the device during driving, and the temperature of the device during driving, respectively. It is measured to extract a total of three types of waveforms from the device, but energy waveforms are not limited to these types and numbers to be extracted. Of course, any one selected from vibration or noise generated, the frequency of power supplied to the device, and the temperature, humidity, and pressure of the device during driving, or a combination of two or more may be used.

That is, as shown in FIG. 3, in the normal information collection step (S10), a normal waveform of current consumed in driving from a normal device and a normal waveform of vibration and temperature are respectively measured and collected. In the way of taking a picture in the state of overlapping each other, it is converted into an image of a mixture of three types of waveforms and extracted and collected as a normal image.

If one type of energy waveform is measured and collected from a device, only the collected energy waveform is converted into an image and extracted and collected as a normal image.

As described above, the normal image information collected in the normal information collection step (S10) is an important basis for determining the state of the device in the discrimination module (S40) to be learned by the discrimination module in the learning step (S40) to be described later. .

As shown in FIG. 1, in the defect information collection step (S20), a process is performed in the operating state of the device before a failure occurs, indicating the change information of the amount of energy measured by the device over time. At least one waveform is collected, converted into an image file, and collected. When two or more defective waveforms are collected, each of the defective waveforms is overlapped with each other, converted into an image file, and collected as a defective image.

Here, as shown in FIG. 4, in the failure information collection step (S20), in the same manner as the normal information collection step (S10), the failure waveform of current consumed for driving from the device before failure and the failure waveform of vibration and temperature Of course, each of the collected defective waveforms is converted into an image in which three types of waveforms are mixed by taking pictures in a state of overlapping each other, and extracting and collecting the defective images.

Since the defective image is extracted from the device before failure, it may include an energy value that is abnormally changed, and thus may have a slightly different image from a normal image extracted and collected from a normal device.

As described above, the normal image information collected in the defective information collection step (S20) is learned by the determination module in the learning step (S40), and becomes an important basis for determining the state of the device in the determination step (S50).

As shown in FIG. 1, in the best work generating step (S30), the defective image information collected in the defective information collection step (S20) is learned, and the defective image information is generated in the generating module 20 based on the learned defective image information. This is a step of mass-producing similar defective images similar to the image.

In general, since the amount of information of a defective image that can be measured and collected from a device before a failure occurs is practically insufficient compared to the amount of information of a normal image that can be measured and collected from a normal device, as shown in FIG. 5, the generation module 20 In the learning step (S40) to be described later, efficient learning of the judgment module 10 is induced by mass production of similar bad image information.

Here, the generation module 20 can be implemented through a general artificial intelligence (Artificial Intelligence) program. The determination and

generation modules

10 and 20 are implemented with a GAN (Generative Adversarial Network) algorithm, but the implementation is limited to this algorithm. Of course not.

1 and 5, in the learning step (S40), the normal image information collected in the normal information collection step (S10) and the similar bad image information produced in the best work generation step (S30) are determined by a discriminating module. This is the learning step in (10), and the process is as follows.

First, the determination module 10 learns about normal images based on a large amount of normal image information collected in the normal information collection step (S10). Through this learning, the determination result of the determination module in the determination step to ensure reliability.

At this time, since the accuracy of the determination result of the determination module 10 can be improved as the information of the normal image collected in the normal information collection step (S10) is rich, it goes without saying that a large amount of normal image information is collected and provided. (S41)

Then, a large amount of similar defective images generated in the generation module 20 are provided to the determination module 10 that has learned about the normal image through the first process S41, and the determination module ( 10), the images determined as normal images are judged normal, and the other images are judged defective.

Through this process, the discrimination module 10 clearly classifies and learns normal images and defective images. (S42)

Then, in the second process (S42), since the success or failure of the result determined by the determination module 10 is fed back to the determination module 10, the determination module 10 determines success or failure of determination. While clearly recognizing and supplementing learning, the learning efficiency of the discrimination module 10 is maximized.

At the same time, the generation module 20 receives feedback and clearly recognizes and learns the success of the determination result of the determination module 10 for the similar defect image, and more precisely so that the generation module 10 can be deceived. An image is generated, and this process eventually helps efficient learning of the generation module 10, and thus becomes a basis for securing excellent reliability for the discrimination result of the generation module 10. (S43)

As shown in FIG. 1, the determination step (S50) performs a process in a real-time driving state, and collects at least one real-time waveform representing the change information of the amount of energy over time measured by the device to obtain an image file. When two or more real-time waveforms are collected, each real-time waveform is superimposed on each other, converted into an image file, collected as a real-time image, and the real-time image is passed through the learning step (S40). This is the step of determining whether it is normal or defective.

That is, as shown in FIG. 6, the determination module 10 determines whether real-time images repeatedly extracted and collected from devices in a real-time driving state are normal or defective, and continuously outputs the determination result. Since the module 10 has gone through the learning step (S40), excellent reliability of the discrimination result can be expected.

As shown in FIG. 1, the detection step (S60) detects the device as an abnormal state when the real-time image collected from the real-time device in the discrimination step (S50) is determined to be defective through the discrimination module 10. It is a step.

That is, as shown in FIG. 6, when the determination module 10 determines that the real-time image is normal, it detects the state of the real-time driven device as a stable state, and when the real-time image is determined to be defective, the real-time driven device is placed in an abnormal state. will be detected with

Therefore, when an abnormal state of a device is detected in real time through the determination module 10, the manager can induce stable inspection and management of the device, thereby reducing the enormous economic cost that can occur when the entire operation of the facility is stopped due to a sudden device failure. phosphorus loss can be prevented.

On the other hand, as shown in FIG. 1, the determination module 10 receives feedback on the success or failure of the determination result of the real-time image determined in the determination step (S50) and re-learns to discriminate in the determination step (S50). A re-learning step (S70) leading to improved reliability of the result; is further included.

That is, as shown in FIG. 6, the determination module 10 determines whether a real-time image extracted from a device driven in real time is normal or defective, and receives feedback on whether the determination result is successful or not through a process of re-learning the determination. The reliability of the determination of the module 10 is induced to be further improved.

In the predictive maintenance method 100 of a device using the GAN algorithm of the present invention consisting of the above process, the determination module 10 generates a large amount of normal image information collected in a normal state of the device and a failure occurs in the generation module 20 Real-time image information collected from real-time driven devices is determined as a normal or defective image by learning a large amount of pseudo-defective image information generated based on defective image information collected from devices prior to processing, and real-time image information is determined as a defective image. In this case, it is effective to prevent huge losses due to failure of the equipment in advance by detecting the equipment in an abnormal state and inducing maintenance and replacement of the equipment at an appropriate time.

In addition, the discrimination module 10 increases the learning effect of the discrimination module 10 by receiving feedback on the success or failure of the discrimination result in the learning step (S40) and the re-learning step (S70) and re-learning. There is an effect of securing excellent reliability for the discrimination result determined by the discrimination module 10 .

Of course, the predictive maintenance method 100 of a device using the GAN algorithm of the present invention can be implemented through a combination of various electronic devices and programs capable of measuring, collecting, discriminating, and detecting the energy waveform of the device.

The present invention has been described with reference to the embodiments shown in the accompanying drawings, but these are illustrative and not limited to the above-described embodiments, and those skilled in the art can make various modifications and equivalent embodiments therefrom. you will understand the point. In addition, of course, modifications by those skilled in the art are possible within a range that does not impair the spirit of the present invention. Therefore, the scope claimed in the present invention will not be determined within the scope of the detailed description, but will be limited by the claims described later and their technical spirit.

The present invention is applicable to the predictive maintenance industry.

Claims

In the predictive correction method of a device performing a repetitive process,

To perform a process in a normal operating state, collect at least one normal waveform representing the change in energy level over time measured by the device, convert it into an image file, and collect it. If two or more normal waveforms are collected, each normal waveform a normal information collection step of superimposing the waveforms and then converting them into image files and collecting them as normal images;

Before a failure occurs, a process is performed while the device is running, and at least one defective waveform representing the change information of energy over time measured by the device is collected, converted into an image file, and collected. If two or more are collected, a defect information collection step of overlapping each of the defect waveforms and then converting them into image files and collecting them as a defect image;

a creation step of learning defective image information collected in the defective information collection step, and mass-producing similar defective images similar to the defective image in a generation module based on the learned defective image information;

a learning step of learning the normal image information collected in the normal information collection step and the similar bad image information produced in the best product creation step in a discrimination module;

When performing a process in real-time operation, at least one real-time waveform representing the change in energy size over time measured by the device is collected and converted into an image file. If two or more real-time waveforms are collected, each real-time waveform A determination step of superimposing waveforms on each other, converting them into image files, collecting real-time images, and discriminating the real-time images as normal or defective in the determination module that has gone through the learning step; and

A method for predictive maintenance of a device using a GAN algorithm, comprising: a detecting step of detecting the device as an abnormal state when the real-time image collected from the real-time device in the determining step is determined to be defective through the determining module.
According to claim 1,

In the learning phase,

The determination module may include: a first step of learning a normal image based on a large amount of normal image information collected in the normal information collection step;

The determination module that has learned about the normal image through the first process provides a large amount of similar defective images generated in the generation module together with the normal image, and the image determined as a normal image through the determination module is normal. A second process of determining that images other than those are defective; and

In the second process, the success or failure of the result determined by the determination module is fed back to the determination module to maximize the learning efficiency of the determination module, and at the same time, the success or failure of the determination result of the determination module for similar defective images is determined. A predictive maintenance method for a device using a GAN algorithm, characterized in that it includes a; third process in which the generation module 20 receives feedback and learns.
According to claim 1,

The discrimination module further includes a re-learning step in which the reliability of the discrimination result determined in the discrimination step is improved by re-learning by receiving feedback on the success or failure of the discrimination result of the real-time image determined in the discrimination step. A method for predictive maintenance of a device using a GAN algorithm.
According to claim 2,

The discrimination module further includes a re-learning step in which the reliability of the discrimination result determined in the discrimination step is improved by re-learning by receiving feedback on the success or failure of the discrimination result of the real-time image determined in the discrimination step. A method for predictive maintenance of a device using a GAN algorithm.
According to claim 1,

The energy over time measured by the device is any one selected from the current consumed in driving the device, vibration or noise generated when the device is driven, the frequency of the power supplied to the device, and the temperature, humidity, and pressure of the device when driven. A method for predictive maintenance of a device using a GAN algorithm, characterized in that two or more are used in combination.