Method for extracting time value through ai supervised learning
The AI supervised learning method addresses the challenge of detecting time values from diverse waveform information by learning relationships between waveform data and offset values, enhancing predictive maintenance efficiency and versatility across various devices, reducing downtime and operational losses.
Patent Information
- Application Number
- PCT/KR2024/007437
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-06-02
- Filing Date
- 2024-05-31
- Publication Date
- 2024-12-05
AI Technical Summary
Existing predictive maintenance methods for devices require individually building programs to detect time values from various forms of waveform information, limiting their versatility and applicability to facilities with hundreds or thousands of devices performing diverse operations, leading to inefficiencies and increased downtime costs due to equipment failures.
A time extraction method through AI supervised learning that collects and analyzes waveform information from diverse devices, learning the relationship between waveform data, offset values, and time values, allowing for real-time detection and provision of accurate time values without the need for individual program building, enhancing predictive maintenance versatility.
Enables efficient and accurate real-time detection of time values from various waveform information types, reducing the need for custom programming and improving predictive maintenance across multiple devices, thereby minimizing downtime and operational losses.
Smart Images

Figure KR2024007437_05122024_PF_FP_ABST
Abstract
Description
Time extraction method using AI supervised learning
[0001] The present invention relates to a time extraction method through AI supervised learning, and more particularly, to a time extraction method through AI supervised learning, which collects waveform information in various forms from various types of devices and devices performing various operations, detects an accurate time value from the collected waveform information, and then provides a large amount of the detected time value and waveform information to AI so that the AI learns about the relationship between the waveform information and the time value, thereby easily obtaining the time value for the real-time waveform information from the learned AI when providing real-time waveform information measured and collected in real time from the device, thereby solving the practical problem of having to individually construct a program for detecting the time value according to the form of the waveform information in order to obtain the time value from the waveform information collected in various forms in the past, thereby ensuring excellent usability for a predictive maintenance method of the device.
[0002] Generally, stable operation is very important for equipment (motors, pumps, conveyors, compressors, etc.) used for automation processes in facilities.
[0003] For example, in the facilities of a large-scale transport factory, hundreds or thousands of devices are installed and operate in conjunction with each other to continuously transport the materials to be transported. If any one of the devices among the many devices breaks down, a catastrophic situation can occur where the entire operation of the facility comes to a halt.
[0004] At this time, downtime due to equipment failure inevitably leads to enormous losses not only in equipment repair costs but also in operating costs and business effects wasted while the equipment is down.
[0005] According to recent data from the Ministry of Employment and Labor and the Korea Occupational Safety and Health Agency, the total number of casualties from industrial accidents annually is estimated at around 100,000, which, when converted to cost, results in an annual loss of 18 trillion won.
[0006] To avoid such unexpected downtime costs, the present applicant has developed and proposed a predictive maintenance method for many types of devices (drive units) that performs predictive maintenance of the devices based on waveform information measured and collected from the devices.
[0007] Among the predictive maintenance methods for the above devices, there is a technology that performs predictive maintenance of the device based on a time value extracted from waveform information.
[0008] However, in order to accurately obtain time values from waveform information collected in various forms from devices due to the various types of devices or the various operations performed by the devices, there was a practical difficulty in that a system (algorithm) capable of detecting time values according to each waveform form had to be individually built as a program.
[0009] Therefore, the predictive maintenance method of the device developed and proposed by the applicant was difficult to apply to facilities consisting of hundreds or thousands of devices of various types and performing various operations, and thus there was a problem that the versatility of the predictive maintenance method was somewhat reduced.
[0010] The present invention has been proposed to solve the above-mentioned problems, and its purpose is to collect waveform information in various forms from various types of devices and devices performing various operations, detect an accurate time value from the collected waveform information, and then provide a large amount of the detected time value and waveform information to AI so that the relationship between the waveform information and the time value can be learned, and when real-time waveform information measured and collected in real time from the device is provided to AI, the time value for the real-time waveform information can be easily obtained from the learned AI, thereby solving the practical problem of having to individually build a program for detecting the time value according to the form of the waveform information in order to obtain the time value from the waveform information collected in various forms in the past, thereby providing a time extraction method through AI supervised learning that can secure excellent usability for a predictive maintenance method of the device.
[0011] In addition, the method of extracting time through AI supervised learning is provided so that the relationship between waveform information and the offset value set in the waveform information and the operation section and time value is further learned through AI, and when real-time waveform information and offset value measured and collected in real time from the device are provided to the AI side, the time value for the operation section is detected and provided based on the real-time waveform information and offset value information from the learned AI.
[0012] In order to achieve the above object, the present invention provides a time extraction method through AI supervised learning, which comprises: a waveform collection step (S10) for collecting waveform information such as a graph showing information on changes in energy size over time measured in an operating state of a device; a detection step (S20) for detecting and collecting a time value between a start point where the waveform information starts and an end point where the waveform information ends from the waveform information collected in the waveform collection step (S10); a supervised learning step (S30) for providing the waveform information collected in the waveform collection step (S10) and the time value detected from the waveform information in the detection step (S20) to the AI side to induce the AI side to learn the relationship between the waveform information and the time value; and an acquisition step (S40) for collecting real-time waveform information showing information on changes in energy size over time measured in a device in a real-time operating state, and providing the real-time waveform information to the AI side to obtain a time value for the real-time waveform information from the AI side.
[0013] In addition, the energy measured and collected from the device is characterized in that one of the following is selected and used: current consumed in operating the device, vibration generated when operating the device, noise generated when operating the device, frequency of the power supply to the device, temperature, humidity, and pressure of the device when operating the device.
[0014] In addition, in the detection step (S20), an offset value is set, and a point where the energy value of the waveform information collected in the waveform collection step (S10) exceeds the offset value is set as a starting point, and a point where the energy value goes down below the offset value is set as an end point, and a section from the starting point to the end point is set as an operation section in which the actual operation of the device is performed, and a time value of the set operation section is detected and collected, and in the supervised learning step (S30), the time value detected for the operation section of the waveform information in the detection step (S20) is provided to the AI side together with the waveform information collected in the waveform collection step (S10), thereby inducing the AI side to learn the relationship between the waveform information and the time value for the operation section of the waveform information, and the acquisition step (S40) is characterized in that it collects real-time waveform information representing information on the change in energy size over time measured in a real-time driving state, and provides an offset value for the real-time waveform information together with the real-time waveform information to the AI side, thereby obtaining a time value for the operation section of the real-time waveform information from the AI side.
[0015] As described above, according to the time extraction method through AI supervised learning according to the present invention, waveform information in various forms is collected from various types of devices and devices performing various operations, and an accurate time value is detected from the collected waveform information, and then the detected time value and waveform information are provided in large quantities to the AI side to learn about the relationship between the waveform information and the time value, so that when real-time waveform information measured and collected in real time from the device is provided to the AI side, the time value for the real-time waveform information is easily obtained from the learned AI, thereby solving the practical problem of having to individually build a program for detecting the time value according to the form of the waveform information in order to obtain the time value from the waveform information collected in various forms in the past, thereby ensuring excellent usability for the predictive maintenance method of the device.
[0016] In addition, by further learning the relationship between the waveform information and the offset value set in the waveform information and the operation section and time value through AI, the real-time waveform information and offset value measured and collected in real time from the device are provided to the AI side, and the time value for the operation section is detected and provided based on the real-time waveform information and offset value information from the learned AI.
[0017] Figure 1 is a block diagram of a time extraction method through AI supervised learning according to an embodiment of the present invention.
[0018] Figures 2 to 9 are diagrams for explaining the time extraction method through AI supervised learning illustrated in Figure 1.
[0019] A time extraction method using AI supervised learning according to a preferred embodiment of the present invention is described in detail with reference to the attached drawings. Detailed descriptions of known functions and configurations that may unnecessarily obscure the gist of the present invention are omitted.
[0020] FIGS. 1 to 9 illustrate a time extraction method through AI supervised learning according to an embodiment of the present invention. FIG. 1 is a block diagram of a time extraction method through AI supervised learning according to an embodiment of the present invention, and FIGS. 2 to 9 are drawings for explaining the time extraction method through AI supervised learning illustrated in FIG. 1, respectively.
[0021] As illustrated in Fig. 1, the time extraction method (100) through AI supervised learning according to an embodiment of the present invention includes a waveform collection step (S10), a detection step (S20), a supervised learning step (S30), and an acquisition step (S40).
[0022] The above waveform collection step (S10) is a step for collecting waveform information such as a graph showing information on changes in energy size over time measured in the operating state of the device.
[0023] Here, in the waveform collection step (S10), the energy measured and collected from the device may be selected and used from among the current consumed in operating the device, vibration generated when operating the device, noise generated when operating the device, frequency of the power supply to the device, temperature, humidity, and pressure of the device when operating the device. In the time extraction method (100) through AI supervised learning of the present invention, the current consumed in operating the device is selected and used as the energy measured from the device, but the selection is not limited to this current.
[0024] That is, when information on changes in the size (value) of current consumed by a device performing an operation in the waveform collection step (S10) is collected over time, a current (energy) waveform such as a graph is collected as waveform information. If waveform information is collected from various types of devices or from devices performing various operations, it is natural that various forms of waveform information can be collected from the devices.
[0025] For example, if the device is limited to a motor, and the motor performs various purposes and operations, for example, if the motor is used in a drilling device or an elevator, if information on changes in the amount of current consumed to drive the motor is collected over time, various types of waveform information, such as that shown in Fig. 2, can be collected.
[0026] The above detection step (S20) is a step for detecting and collecting the time value between the start point where the waveform information begins and the end point where the waveform information ends in the waveform information collected in the above waveform collection step (S10).
[0027] That is, as illustrated in FIG. 3, in the detection step (S20), the time values between the start and end points of various types of waveform information collected in the waveform collection step (S10) are detected and collected, and at this time, the time values for the waveform information collected are preferably detected and collected as learning materials (data) for AI to learn in the supervised learning step (S30) described later, so that the time values for a large amount of waveform information are efficiently collected for AI learning.
[0028] Here, a system (algorithm) for detecting the exact time value of the waveform information collected in the waveform collection step (S10) is implemented as a program so that the detection step (S20) detects and collects the exact time value from the waveform information.
[0029] Therefore, the learning objective of the AI that learns based on the time value information for the waveform information detected and collected in the above detection step (S20) is accurately set.
[0030] If the time value of the waveform information is not accurately detected and provided in the detection step (S20), it is difficult to perform the AI learning purpose of obtaining the time value from the waveform information through the supervised learning step (S30) preferably, and this may lower the reliability of the time value obtained from the real-time waveform information through the AI side learned in the acquisition step (S40) described later. Therefore, the detection step (S20) must detect and provide an accurate time value from the waveform information.
[0031] In summary, the above detection step (S20) must accurately detect the time value for each waveform information collected in the waveform collection step (S10) and provide the time value for the detected waveform information as learning material for AI learning.
[0032] The above map learning step (S30) is a step that provides the waveform information collected in the waveform collection step (S10) and the time value detected from the waveform information in the detection step (S20) to the AI (Artificial Intelligence) side, thereby inducing the AI side to learn the relationship between the waveform information and the time value.
[0033] That is, as illustrated in FIG. 4, the AI clearly learns the relationship between waveform information and time values through a large amount of learning data on various types of waveform information and the time values of the waveform information. At this time, the AI performs learning in a supervised learning manner through deep learning, but of course, it is not limited to performing learning in this manner.
[0034] Here, the AI can effectively understand and learn the relationship between waveform information and time values as the provided learning materials (waveform information and time values) are abundant. Therefore, it is desirable to collect, detect, and provide time values together with a large amount of waveform information in the waveform collection step (S10) and the detection step (S20).
[0035] The above acquisition step (S40) is a step of collecting real-time waveform information representing information on changes in energy size over time measured from a device in a real-time operating state, and providing the real-time waveform information to the AI side to obtain a time value for the real-time waveform information from the AI side.
[0036] That is, as illustrated in FIG. 5, the AI, which has learned the relationship between waveform information and time values through the above-described waveform collection step (S10), detection step (S20), and supervised learning step (S30), automatically detects and provides the time value for the real-time waveform information when real-time waveform information measured from a device in real-time operation is provided, thereby enabling the time value for the real-time waveform information to be obtained very easily through the AI.
[0037] In summary, AI learned through the above process can very easily detect and provide time values for each waveform information even when various types of waveform information are collected and provided from various types of devices and devices performing various operations.
[0038] For example, when real-time waveform information collected from the real-time operating status of a motor used in a drilling device or real-time waveform information collected from the real-time operating status of a motor used in an elevator is provided to the AI, the AI detects and provides time values for each of the real-time waveform information of the motor, thereby enabling time values for various real-time waveform information to be obtained very easily.
[0039] Therefore, by training AI through the above process and enabling the trained AI to easily detect and acquire time values from various types of waveform information, the practical difficulty of having to individually build a system (algorithm) for detecting time values according to the type of waveform information in order to acquire time values from waveform information collected in various forms due to the type of device or various operations performed in the past can be resolved.
[0040] That is, the applicant has developed and proposed various types of predictive maintenance methods that measure and collect time values from waveform information (current waveform) collected from devices and perform predictive maintenance of the devices based on the collected time values. These predictive maintenance methods can be very effectively applied to facilities consisting of hundreds or thousands of devices through AI learned as described above.
[0041] 『"Precise predictive maintenance method of a driving part (Registration No. 10-2103145)", "Predictive maintenance method of a device through defining a constant speed for time (Registration No. 10-2477712)", "Predictive maintenance method of a device through defining a constant speed for time (Registration No. 10-2477714)", "Predictive maintenance method of a device through a control output signal (Registration No. 10-2316493)", "Predictive maintenance method of a device through multiple control output signals (Registration No. 10-2316483)"』
[0042] Meanwhile, in the above detection step (S20), an offset value is set,
[0043] In the waveform collection step (S10), the point where the energy value of the waveform information collected exceeds the offset value is set as the starting point, and the point where the energy value falls below the offset value is set as the ending point, and the section from the starting point to the ending point is set as the operating section where the actual operation of the device is performed, and the time value of the set operating section is detected and collected.
[0044] In order to achieve the purpose of the present invention, waveform information must be collected from a device in an operating state, and the time value for the waveform information must be accurately detected. In a device that repeats operation and rest, if the device is completely stopped during rest, the start point and the end point where the waveform information starts can be clearly detected, and the time value for the waveform information of the operating state can be accurately detected from the detected start point to the end point by making the device the operating state, that is, the operating section in which the device actually operates, but as illustrated in FIG. 6, if the device is not completely stopped during rest, it is difficult to clearly detect the start point and the end point where the waveform information starts due to the standby current flowing in the device, making it difficult to detect an accurate time value from the waveform information.
[0045] Accordingly, as shown in FIG. 7, by setting an offset value having a predetermined size, even when the device is not completely stopped when at rest, the point where the current value in the waveform information of the device exceeds the offset value is set as the starting point, and the point where the current value goes down below the offset value is set as the ending point, and the section from the starting point to the ending point is forcibly set as the operating state of the device, that is, the operating section (operating section) in which the device is actually operated, and the time value is detected from the waveform information of the set operating section, thereby allowing an accurate time value to be detected from the waveform information collected from the device, so that the technology of the present invention can be easily applied to various types of devices.
[0046] Then, in the map learning step (S30), the waveform information collected in the waveform collection step (S10) and the time value detected for the operation section of the waveform information in the detection step (S20) are provided to the AI side, thereby inducing the AI side to learn the relationship between the waveform information and the time value for the operation section of the waveform information.
[0047] That is, as illustrated in FIG. 8, the AI clearly learns the relationship between waveform information, offset values, operation intervals, and time values. In the waveform collection step (S10) and the detection step (S20), it is desirable to provide information on offset values, operation intervals, and time values together with a large amount of waveform information to induce efficient learning of the AI.
[0048] After that, the acquisition step (S40) collects real-time waveform information representing information on changes in energy size over time measured in a real-time operating state, and provides the AI side with an offset value for the real-time waveform information together with the real-time waveform information so as to acquire a time value for the operating section of the real-time waveform information from the AI side.
[0049] That is, as shown in Fig. 9, the AI that has learned the relationship between waveform information, offset value, operating section, and time value through the learning process described above automatically detects the time value for the operating section of the real-time waveform information and the offset value measured from the device in real-time operation when provided with the real-time waveform information and the offset value, thereby enabling the AI to very easily obtain the time value for the real-time waveform information.
[0050] In summary, even in the information waveform of a device that does not come to a complete stop during rest, the time value can be automatically and accurately obtained from the AI learned as described above.
[0051] The time extraction method (100) through AI supervised learning of the present invention, which is composed of the steps described above, collects waveform information in various forms from various types of devices and devices performing various operations, detects an accurate time value from the collected waveform information, and then provides a large amount of the detected time value and waveform information to the AI side to learn about the relationship between the waveform information and the time value, so that when real-time waveform information measured and collected in real time from the device is provided to the AI side, the time value for the real-time waveform information is easily obtained from the learned AI, thereby solving the practical problem of having to individually build a program for detecting the time value according to the form of the waveform information in order to obtain the time value from the waveform information collected in various forms in the past, thereby ensuring excellent usability for the predictive maintenance method of the device.
[0052] In addition, by further learning the relationship between the waveform information and the offset value set in the waveform information and the operation section and time value through AI, the real-time waveform information and offset value measured and collected in real time from the device are provided to the AI side, and the time value for the operation section is detected and provided based on the real-time waveform information and offset value information from the learned AI.
[0053] While the present invention has been described with reference to the embodiments illustrated in the accompanying drawings, these are merely illustrative and are not limited to the above-described embodiments. Those skilled in the art will readily appreciate that various modifications and equivalent embodiments are possible. Furthermore, it should be understood that modifications are possible by those skilled in the art without departing from the spirit and scope of the present invention. Therefore, the scope of the rights claimed in the present invention is not limited to the detailed description set forth herein, but rather by the claims set forth below and their technical spirit.
[0054] The present invention can be used in the predictive maintenance industry of equipment.
Claims
1. A waveform collection step for collecting waveform information such as a graph showing information on changes in energy size over time measured in the operating state of the device; A detection step for detecting and collecting the time value between the start point where the waveform information starts and the end point where the waveform information ends from the waveform information collected in the above waveform collection step; A supervised learning step that provides the waveform information collected in the waveform collection step and the time value detected from the waveform information in the detection step to the AI (Artificial Intelligence) side so that the AI side learns the relationship between the waveform information and the time value; and A time extraction method through AI supervised learning, characterized by comprising an acquisition step (S40) of collecting real-time waveform information representing information on changes in energy size over time measured from a device in a real-time operating state, and providing the real-time waveform information to the AI side to obtain a time value for the real-time waveform information from the AI side.
2. In paragraph 1, The energy measured and collected from the device is: A time extraction method through AI supervised learning characterized in that any one of the following is selected and used: current consumed in operating the device, vibration generated during device operation, noise generated during device operation, frequency of the device power supply, temperature, humidity, and pressure of the device during device operation.
3. In paragraph 1 or 2, In the above detection step, an offset value is set, The energy value of the waveform information collected in the waveform collection step is set as the starting point when it exceeds the offset value, and the end point when it falls below the offset value, and the section from the starting point to the end point is set as the operation section in which the actual operation of the device is performed, and the time value of the set operation section is detected and collected. In the above-mentioned map learning step, the time values detected for the operation section of the waveform information in the detection step are provided to the AI side together with the waveform information collected in the waveform collection step, thereby inducing the AI side to learn the relationship between the waveform information and the time values for the operation section of the waveform information. The above acquisition step is a time extraction method through AI supervised learning, characterized in that it collects real-time waveform information representing information on changes in energy size over time measured in a real-time operating state, and provides an offset value for the real-time waveform information together with the real-time waveform information to the AI side, thereby obtaining a time value for the operating section of the real-time waveform information from the AI side.
Citation Information
Patent Citations
Analysis of Electro-polarization Characteristics of Cells by Regulation of Dielectrophoretic Force
KR1020200026169A
System and Methods for Matching Cloth Seller via On-line
KR1020220053396A
Apparatus for cleaning sensor
KR1020240046969A
A method of detecting a text for detecting a scene text and an apparatus for performing the same
KR1020240113259A
Process abnormality detection device and method using analysis of control section temperature signal
WO2022030831A1
