KR20220059837A - Method, server and computer program for learning predictive maintenance model to monitor machine status - Google Patents

Abstract

The present invention relates to a method, server, and computer program for training a predictive maintenance model to monitor a machine status to increase accuracy of a predictive maintenance model. According to the present invention, the method comprises the following steps: collecting and pre-processing at least one of sound data, vibration data, and temperature data of a machine; generating a predictive maintenance model on the basis of at least one of the preprocessed sound data, vibration data, and temperature data; simulating occurrence of an abnormality in a mechanical state by applying a decision logic to at least one of the collected sound data, vibration data, and temperature data; and extracting a discount factor on the basis of a simulation result and performing learning of the predictive maintenance model through reinforcement learning on the basis of the discount factor.

Description

METHOD, SERVER AND COMPUTER PROGRAM FOR LEARNING PREDICTIVE MAINTENANCE MODEL TO MONITOR MACHINE STATUS

The present invention relates to a predictive maintenance model learning method, a server and a computer program for monitoring a machine condition.

Predictive maintenance technology is a technology that can predict the safety of factory automation and the probability of machine failure in advance. .

Predictive maintenance technology creates a data-based monitoring algorithm to detect abnormal signs of a machine and repair its condition through a machine failure prediction model.

The conventional predictive maintenance technology was performed through machine voice analysis through abnormal noise detection of the machine.

However, the conventional data-based machine voice analysis is performed based on a model learned using artificially collected data rather than actual data collected at the site. In this case, using the ground truth technique, The accuracy of the learned model, that is, a model based on real data, was significantly lower.

In addition, in the conventional case, even if a model is built based on the ground truth technique, in general, environmental factors of the machine site are not reflected.

Korea Patent Publication No. 10-2051226 (Registered on November 26, 2019) Korean Patent Publication No. 10-2020-0014129 (published on February 10, 2020)

The present invention is to solve the problems of the prior art, and by updating the predictive machine maintenance model through re-learning, it is possible to detect the abnormality of the mechanical device by reflecting the environmental factors of the machine site. would like to provide

However, the technical problems to be achieved by the present embodiment are not limited to the technical problems described above, and other technical problems may exist.

As a means for achieving the above-described technical problem, an embodiment of the present invention is a predictive maintenance model learning method for monitoring a machine condition, at least one of sound data, vibration data, and temperature data of the machine is collected and pre-processed to do; generating a predictive maintenance model based on at least one of the preprocessed sound data, vibration data, and temperature data; simulating occurrence of an abnormality in the machine state by applying a decision logic to at least one of the collected sound data, vibration data, and temperature data; and deriving a discount factor based on the simulation result, and performing learning of the predictive maintenance model through reinforcement learning based on the discount factor, a predictive maintenance model learning method can provide

Another embodiment of the present invention provides a predictive maintenance model learning server for monitoring a machine state, comprising: a data pre-processing unit for pre-processing by collecting at least one of sound data, vibration data, and temperature data of a machine; a predictive maintenance model unit for generating a predictive maintenance model based on at least one of the preprocessed sound data, vibration data, and temperature data; a simulation unit for simulating occurrence of an abnormality in the mechanical state by applying a decision logic to at least one of the collected sound data, vibration data, and temperature data; and a learning unit for deriving a discount factor based on the simulation result and learning the predictive maintenance model through reinforcement learning based on the discount factor, a predictive maintenance model learning server can provide

Another embodiment of the present invention provides a computer program stored in a computer-readable medium including a sequence of instructions for learning a predictive maintenance model for monitoring a machine condition, wherein the computer program is executed by a computing device, Collect and preprocess at least one of sound data, vibration data, and temperature data of a machine, generate a predictive maintenance model based on at least one of the preprocessed sound data, vibration data, and temperature data, and sound data collected from the machine , by applying a decision logic to at least one of vibration data and temperature data to simulate the occurrence of an abnormality in the machine state, derive a discount factor based on the simulation result, and reinforcement learning based on the discount coefficient Learning), it is possible to provide a computer program stored in a computer-readable recording medium, including a sequence of instructions to proceed with the learning of the predictive maintenance model.

The above-described problem solving means are merely exemplary, and should not be construed as limiting the present invention. In addition to the exemplary embodiments described above, there may be additional embodiments described in the drawings and detailed description.

According to any one of the above-described problem solving means of the present invention, through reinforcement learning using a discount coefficient that appropriately adjusts the ratio of sound data, vibration data, and temperature data of the machine based on the judgment logic for monitoring the machine state By training the model, it is possible to provide a predictive maintenance model training method that can improve the accuracy of the model.

1 is a configuration diagram of a predictive maintenance model learning server according to an embodiment of the present invention.
FIG. 2 is an exemplary diagram for explaining a logic for determining machine sound data preprocessed according to an embodiment of the present invention.
3 is an exemplary view for explaining the logic of determination of machine vibration data preprocessed according to an embodiment of the present invention.
4 is an exemplary diagram for explaining the logic of determining the preprocessed machine temperature data according to an embodiment of the present invention.
5 is an exemplary view for explaining a PF curve according to an embodiment of the present invention.
6 is a flowchart of a predictive maintenance model learning method according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement them. However, the present invention may be embodied in several different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part is "connected" with another part, this includes not only the case of being "directly connected" but also the case of being "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated, and one or more other features However, it is to be understood that the existence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded in advance.

In this specification, a "part" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. In addition, one unit may be implemented using two or more hardware, and two or more units may be implemented by one hardware.

Some of the operations or functions described as being performed by the terminal or device in this specification may be instead performed by a server connected to the terminal or device. Similarly, some of the operations or functions described as being performed by the server may also be performed in a terminal or device connected to the server.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a configuration diagram of a predictive maintenance model learning server according to an embodiment of the present invention. Referring to FIG. 1 , the predictive maintenance model learning server 100 may include a data preprocessing unit 110 , a predictive maintenance model unit 120 , a simulation unit 130 , and a learning unit 140 .

In addition, the data preprocessor 110 may include a sound data unit 111 , a vibration data unit 112 , and a temperature data unit 113 , and the simulation unit 130 includes a sound determination unit 131 and a vibration determination unit. It may include a unit 132 and a temperature determination unit 133 . However, the above components 110 to 140 are merely illustrative of components that can be controlled by the predictive maintenance model learning server 100 .

Each component of the predictive maintenance model learning server 100 of FIG. 1 is generally connected through a network. For example, as shown in FIG. 1 , the data preprocessing unit 110 , the predictive maintenance model unit 120 , the simulation unit 130 , and the learning unit 140 may be connected simultaneously or at intervals of time.

A network refers to a connection structure in which information can be exchanged between each node, such as terminals and servers, and includes a local area network (LAN), a wide area network (WAN), and the Internet (WWW: World). Wide Web), wired and wireless data communication networks, telephone networks, wired and wireless television networks, and the like. Examples of wireless data communication networks include 3G, 4G, 5G, 3rd Generation Partnership Project (3GPP), Long Term Evolution (LTE), World Interoperability for Microwave Access (WIMAX), Wi-Fi, Bluetooth communication, infrared communication, ultrasound Communication, Visible Light Communication (VLC), LiFi, and the like are included, but are not limited thereto.

The data preprocessor 110 according to an embodiment of the present invention may collect and preprocess at least one of sound data, vibration data, and temperature data of a machine. The data preprocessor 110 may include a sound data unit 111 , a vibration data unit 112 , and a temperature data unit 113 .

For example, the data preprocessor 110 may preprocess each of sound data, vibration data, and temperature data to obtain sound data of 3000 Hz or higher, vibration data of 1 kHz or higher, and temperature data of the last 48 hours.

For example, the data preprocessor 110 may build an initial predictive maintenance model through 4 days of learning based on the collected data, and may transmit the collected data to a simulation unit to be described later.

The sound data unit 111 may convert sound data of the machine into a frequency domain. For example, the sound data unit 111 may transform the collected sound data into a frequency domain using Fast Fourier Transform (FFT).

The sound data unit 111 may remove a sound of 3000 Hz or less, which is highly likely to be noise in the frequency domain.

The vibration data unit 112 may convert the vibration data of the machine into an energy value and remove noise by applying a moving average filter to the energy value. For example, the vibration data unit 112 may collect vibration data of three axes including the x-axis, the y-axis, and the z-axis from the vibration sensor of the machine. The vibration data unit 112 may, for example, convert the vibration data of three axes into energy values through Equation 1 below to check the amount of change and determine the trend.

<Equation 1>

Also, the vibration data unit 112 may apply a moving average filter such as a Simple Moving Average (SMA) or an Exponential Moving Average (EMA) to remove noise from the vibration data.

The temperature data unit 113 may accumulate machine temperature data for a predetermined period of time. Since a sudden change does not appear in the case of temperature data, the temperature data unit 113 may utilize, for example, 48 hours of accumulated temperature data.

The temperature data unit 113 may check the intersection of each trend by, for example, applying a 10-minute moving average filter and an hourly moving average filter to the accumulated temperature data.

The predictive maintenance model unit 120 according to an embodiment of the present invention may generate a predictive maintenance model based on at least one of preprocessed sound data, vibration data, and temperature data.

For example, the predictive maintenance model may be a model based on at least one of a Recurrent Neural Network (RNN) and a Long Short-Term Memory model (LSTM). The predictive maintenance model may perform self-learning based on at least one of sound data, vibration data, and temperature data preprocessed initially.

The simulation unit 130 according to an embodiment of the present invention may simulate occurrence of an abnormality in a mechanical state by applying a decision logic to at least one of the collected sound data, vibration data, and temperature data.

For example, the simulation unit 130 may detect the occurrence of an abnormality in the machine state by performing first to fourth conditions of determination logic, which will be described later, based on the collected data. As will be described later, the predictive maintenance model may be updated based on the simulation result.

The simulation unit 130 may include a sound determination unit 131 , a vibration determination unit 132 , and a temperature determination unit 133 .

2 is an exemplary diagram for explaining application of a judgment logic to machine sound data preprocessed according to an embodiment of the present invention.

Referring to FIG. 2A , the sound determination unit 131 may extract peak values 211 and 212 having the greatest magnitude change in the frequency domain 210 above or below a preset reference frequency range.

Referring to (b) of FIG. 2 , the sound determination unit 131 sets the frequency characteristic range 230 based on the extracted peak values 211 and 212 to determine the upper limit of the set frequency characteristic range 230 . When it corresponds to the first condition to deviate or the second condition to deviate from the lower limit of the set frequency characteristic range 230, the occurrence of an abnormality in the machine state may be detected.

For example, the sound determination unit 131 may set the frequency characteristic range 230 based on 16000 Hz in the frequency domain 210 of 44100 Hz. The sound determination unit 131 can set the upper limit of the frequency characteristic range 230 by extracting the peak value 211 with the largest change in magnitude at 16000 Hz or higher, and the peak value 212 with the largest change in magnitude at 16000 Hz or less. It is possible to set the lower limit of the frequency feature range 230 by extraction.

Referring to FIG. 2A , the extracted peak values 211 and 212 are identified as features 221 and 222 of the corresponding sound when the frequency domain 210 is subjected to fast Fourier transform processing 220 .

The sound determination unit 131 may detect a case out of the set frequency characteristic range 230 and detect abnormal occurrence of a machine state. For example, the sound determination unit 131, according to the first condition, is out of the upper limit 211 of the set frequency characteristic range 230, that is, out of the upper limit 211 of 16000 Hz by detecting the case of the machine state abnormal occurrence can be detected. For another example, the sound determination unit 131, according to the second condition, is out of the lower limit 212 of the set frequency feature range 230, that is, when it is out of the lower limit 212 of 16000 Hz to detect the machine It is possible to detect the occurrence of an abnormal state.

3 is an exemplary view for explaining application of a judgment logic of pre-processed machine vibration data according to an embodiment of the present invention. Referring to FIG. 3 , the vibration determination unit 132 calculates a threshold value (Threshold, (a)) and hysteresis (Hysteresis, (b), (c)) of the converted energy value, and hysteresis for a predetermined time period. When the third condition in which the state outside the range of (b) and (c)) is satisfied, the occurrence of an abnormality in the machine state may be detected.

For example, the vibration determination unit 132 may set the threshold value a based on a preset standard. Here, the preset standard may be the ISO 10816 standard.

In addition, the vibration determination unit 132, for example, through the following Equations 2-1, 2-2 and 2-3, positive hysteresis (b) and negative hysteresis (b) in the converted energy value of the vibration data ( c) can be calculated separately.

<Equation 2-1>

<Equation 2-2>

<Equation 2-3>

Referring to Equations 2-1, 2-2, and 2-3, the vibration determination unit 132 may calculate the hysteresis ((b), (c)) based on the difference between the same input and output. For example, the vibration determination unit 132 may calculate the average of deviations of two peaks during the initial 10 seconds of the vibration data to calculate the hysteresis compared to the maximum peak value.

The vibration determination unit 132 includes not only the threshold value but also the hysteresis ((b), (c)) in the reference range for detecting the occurrence of an abnormality in the mechanical state, thereby providing a temporary value due to noise or other causes included in the vibration data. It is possible to prevent erroneous detection by

For example, the vibration determination unit 132, according to the third condition, escapes the negative hysteresis (c) in the vibration data converted to the energy value (310) and maintains the threshold value (a) or more for 10 seconds or more ( 321), it is possible to detect the occurrence of an abnormality in the machine state. As another example, the vibration determination unit 132, according to the third condition, out of the positive hysteresis (b) and maintaining the threshold value or more for 10 seconds or more (320) can detect the occurrence of an abnormality in the machine state. there is.

4 is an exemplary diagram for explaining application of a decision logic of preprocessed machine temperature data according to an embodiment of the present invention.

4 , the temperature determination unit 133 applies a short-term moving average filter 411 and a long-term moving average filter 412 to the accumulated temperature data, respectively, and a short-term moving average filter 411 and a long-term moving average filter When 412 corresponds to the fourth condition intersecting 410 , it is possible to detect the occurrence of an abnormality in the machine state.

In general, due to the characteristics of the moving average filter, the moving average of the long-term moving average filter 412 changes relatively slowly, and the moving average of the short-term moving average filter 411 changes relatively rapidly.

Therefore, according to the fourth condition, the temperature determination unit 133 is configured to, according to the fourth condition, when the short-term moving average 411 abruptly changes and crosses the long-term moving average 412 in the temperature data that does not show an abrupt change (410), the machine By judging that a large change has occurred in the temperature of the

The learning unit 140 according to an embodiment of the present invention may learn the predictive maintenance model through reinforcement learning using a discount factor based on the simulation result. Here, the learning of the predictive maintenance model may be re-learning through reinforcement learning. In Reinforcement Learning, the Max value is assumed to be 0.5, because the closer the discount factor is applied to 1, the greater the change in the model. Therefore, when calculating the discount factor, each condition is converted to a ratio of 0.1 for sound and heat and 0.15 for vibration and noise. apply to

For example, the learning unit 140 may calculate a discount coefficient based on at least one of sound data, vibration data, and temperature data of the machine, and may perform reinforcement learning of the predictive maintenance model based on the calculated discount coefficient. Through this, the learning unit 140 may build a more systematic predictive maintenance model reflecting environmental factors of the machine site.

For example, the determination model (q _π ) of reinforcement learning may be expressed by Equation 3 below.

<Equation 3>

Here, S means the current state, A means the expected result, and G _t means the optimized action value.

In reinforcement learning, there is a concept of a reward for a result, and by applying a discount coefficient to this, it is possible to determine how much reward is reflected. For example, by applying Equation 4 below, the determination model (q _π ) of reinforcement learning may be expressed as Equation 5 below.

<Equation 4>

<Equation 5>

Equation 5 is a decision model (q _π ) to which R _t+1 is applied by the result (A _t+1 ) for the next state (S _t+1 ). Based on this, the probability for the next state (S _t+1 ) can be expressed, for example, by Equation 6 below.

<Equation 6>

Accordingly, the final model may be expressed, for example, by Equation 7 below.

<Equation 7>

In reinforcement learning based on Equations 3 to 7, the factor that has the greatest influence on learning is the discount coefficient ( γ ). The discount coefficient γ is the rate of application of compensation to the result, as described above. In predictive maintenance, when an arbitrary value is substituted for the discount coefficient ( γ ), there is a problem in that various environmental factors for the machine site cannot be systematically reflected.

Therefore, the predictive maintenance model learning server 100 according to the present invention calculates a discount coefficient based on the simulation results for the sound data, vibration data, and temperature data of the machine, so that it can organically respond to various environmental factors of the machine site do.

According to an embodiment of the present invention, a discount coefficient may be calculated by giving weights to the first to fourth conditions of the decision logic, respectively.

According to an embodiment, the accuracy of the predictive maintenance model may be further improved by re-learning the predictive maintenance model using reinforcement learning and at least one of a recurrent neural network and an LSTM model.

5 is an exemplary diagram for explaining a P-F curve according to an embodiment of the present invention. Referring to FIG. 5 , in general, according to the P-F curve, abnormal signs of mechanical conditions may appear in the order of high-frequency sound (a), vibration (b), low-frequency sound (c), and temperature (d). In this case, the fewer the generated characteristics, the more early warning characteristics. In the present invention, the discount coefficient of reinforcement learning is calculated by utilizing these characteristics.

According to the P-F curve, there are a total of five points in the section from the point where a mechanical failure may occur to the occurrence of a mechanical failure, but in this predictive maintenance model learning server 100, the decision logic The occurrence of an abnormality in the machine state may be detected based on the first to fourth conditions. For example, the learning unit 140 may utilize, for example, Equation 8 below.

<Equation 8>

Here, ir ₁ to ir ₄ are the first to fourth conditions of the decision logic, and a constant multiplied by each of ir ₁ to ir ₄ may be a weight for each condition.

Due to the characteristics of the P-F curve, a precursor to failure occurs in the facility in the order of high-frequency sound, vibration, oil, noise, and heat. Since the present invention does not correspond to the oil measurement part, the weight of oil is distributed to vibration and noise to 0.2, 0.3, Weights of 0.3 and 0.2 are given.

However, if the sum of the weights of each condition is set to 1, the results of the existing learning prior to re-learning may not be sufficiently reflected. In the present invention, the sum of the weights of each condition is set to 0.5 to maintain the results of the existing learning. do.

That is, the above 0.1 and 0.15 mean the weight ratio corresponding to the condition, and the more conditions satisfied, the closer to 0.5. Accordingly, reward is variably applied to the model.

That is, the learning unit 140 can build a predictive maintenance model reflecting various environmental factors of the machine site by calculating the ratio of the discount coefficient based on the distributed weights for the first to fourth conditions of the judgment logic, It is possible to more accurately detect the occurrence of an abnormality in the machine state.

6 is a flowchart of a predictive maintenance model learning method according to an embodiment of the present invention. The predictive maintenance model learning method illustrated in FIG. 6 includes steps processed in time series according to the embodiments illustrated in FIGS. 1 to 5 . Therefore, even if omitted below, it is also applied to the method of learning the predictive maintenance model in the predictive maintenance model learning server 100 according to the embodiment shown in FIGS. 1 to 5 .

In step S610, the predictive maintenance model learning method may collect and pre-process at least one of sound data, vibration data, and temperature data of the machine.

In step S620, the predictive maintenance model learning method may generate a predictive maintenance model based on at least one of preprocessed sound data, vibration data, and temperature data.

In the predictive maintenance model learning method in step S630, at least one of sound data, vibration data, and temperature data collected from the machine is input to the generated predictive maintenance model, and a decision logic for monitoring the machine state may be applied.

In step S640 , the predictive maintenance model learning method may proceed to learn the predictive maintenance model through reinforcement learning using a discount coefficient based on a result of application of the determination logic.

In the above description, steps S610 to S640 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present invention. In addition, some steps may be omitted as needed, and the order between the steps may be switched.

The method for learning the predictive maintenance model in the predictive maintenance model learning server 100 described through FIGS. 1 to 6 includes a computer program stored in a computer-readable recording medium executed by a computer or instructions executable by the computer. It may also be implemented in the form of a recording medium. In addition, the method for learning the predictive maintenance model in the predictive maintenance model server described with reference to FIGS. 1 to 6 may be implemented in the form of a computer program stored in a computer-readable recording medium that is executed by a computer.

A computer-readable recording medium may be any available medium that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. In addition, the computer-readable recording medium may include a computer storage medium. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The description of the present invention described above is for illustration, and those of ordinary skill in the art to which the present invention pertains can understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a dispersed form, and likewise components described as distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the following claims rather than the above detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

100: predictive maintenance model training server
110: data preprocessor
120: predictive maintenance model part
130: simulation unit
140: study unit

Claims (20)

A method for learning a predictive maintenance model for monitoring a machine condition, the method comprising:
collecting and pre-processing at least one of sound data, vibration data, and temperature data of the machine;
generating a predictive maintenance model based on at least one of the preprocessed sound data, vibration data, and temperature data;
simulating occurrence of an abnormality in the machine state by applying a decision logic to at least one of the collected sound data, vibration data, and temperature data; and
Deriving a discount factor based on the simulation result, and performing learning of the predictive maintenance model through reinforcement learning based on the discount factor
Including, a predictive maintenance model learning method.
The method of claim 1,
The pre-processing step is
Pre-processing the sound data of the machine in the frequency domain, the predictive maintenance model learning method.
3. The method of claim 2,
The simulation step is
extracting a peak value having the largest magnitude change in the sound data above or below a preset reference frequency range; and
setting a feature range of the frequency based on the extracted peak value
Including, a predictive maintenance model learning method.
4. The method of claim 3,
The simulation step is
Detecting the occurrence of an abnormality in the machine state when it corresponds to a first condition deviating from the upper limit of the characteristic range of the set frequency, or corresponds to a second condition deviating from the lower limit of the characteristic range of the set frequency
Further comprising, a predictive maintenance model learning method.
The method of claim 1,
The pre-processing step is
Converting the vibration data of the machine to an energy value and removing noise by applying a Moving Average Filter to the energy value
Including, a predictive maintenance model learning method.
6. The method of claim 5,
The simulation step is
calculating a threshold and a hysteresis of the converted energy value; and
Detecting occurrence of an abnormality in the machine state when a third condition in which the state outside the hysteresis range is maintained for a predetermined period of time
Further comprising, a predictive maintenance model learning method.
The method of claim 1,
The pre-processing step is
accumulating the temperature data of the machine for a predetermined period of time
Further comprising, a predictive maintenance model learning method.
8. The method of claim 7,
The simulation step is
applying a short-term moving average filter and a long-term moving average filter to the accumulated temperature data, respectively; and
detecting the occurrence of an abnormality in the machine state when a fourth condition in which the short-term moving average filter and the long-term moving average filter intersect is satisfied;
Further comprising, a predictive maintenance model learning method.
The method of claim 1,
The step of learning the predictive maintenance model is,
calculating the discount coefficient by giving each of the weights to at least one condition of the determination logic
Including, a predictive maintenance model learning method.
The method of claim 1,
wherein the predictive maintenance model is based on at least one of a Recurrent Neural Network (RNN) and a Long Short-Term Memory model (LSTM).
In the predictive maintenance model learning server for monitoring machine status,
a data preprocessing unit that collects and preprocesses at least one of sound data, vibration data, and temperature data of the machine;
a predictive maintenance model unit for generating a predictive maintenance model based on at least one of the preprocessed sound data, vibration data, and temperature data;
a simulation unit for simulating occurrence of an abnormality in the mechanical state by applying a decision logic to at least one of the collected sound data, vibration data, and temperature data; and
a learning unit deriving a discount factor based on the simulation result and learning the predictive maintenance model through reinforcement learning based on the discount factor;
Including, predictive maintenance model learning server.
12. The method of claim 11,
The data preprocessor,
A sound data unit for pre-processing the sound data of the machine in the frequency domain
Including, predictive maintenance model learning server.
13. The method of claim 12,
The simulation unit,
Extracting the peak value with the largest magnitude change in the sound data above or below a preset reference frequency range, setting the characteristic range of the frequency based on the extracted peak value, and determining the upper limit of the characteristic range of the set frequency A sound determination unit that detects the occurrence of an abnormality in the machine state when it corresponds to the first condition to deviate or corresponds to the second condition deviates from the lower limit of the characteristic range of the set frequency
Including, predictive maintenance model learning server.
12. The method of claim 11,
The data preprocessor,
A vibration data unit that converts the vibration data of the machine into an energy value and removes noise by applying a Moving Average Filter to the energy value
Including, predictive maintenance model learning server.
15. The method of claim 14,
The simulation unit,
Threshold and hysteresis of the converted energy value are calculated, and when the third condition in which the state outside the range of hysteresis is maintained for a predetermined period of time is met, abnormal occurrence of the machine state is detected Vibration detection unit
Including, predictive maintenance model learning server.
12. The method of claim 11,
The data preprocessor,
Temperature data unit for accumulating the temperature data of the machine for a predetermined period of time
Including, predictive maintenance model learning server.
17. The method of claim 16,
The simulation unit,
When a short-term moving average filter and a long-term moving average filter are respectively applied to the accumulated temperature data, and the short-term moving average filter and the long-term moving average filter intersect the fourth condition, the occurrence of an abnormality in the machine state is detected temperature determination unit
Including, predictive maintenance model learning server.
12. The method of claim 11,
The learning unit,
The predictive maintenance model learning server, which calculates the discount coefficient by giving a weight to each of the at least one condition of the determination logic.
12. The method of claim 11,
The predictive maintenance model is based on at least one of a Recurrent Neural Network (RNN) and a Long Short-Term Memory model (LSTM) model.
A computer program stored in a computer-readable medium comprising a sequence of instructions for learning a predictive maintenance model for monitoring a machine condition, the computer program comprising:
When the computer program is executed by a computing device,
Collect and preprocess at least one of sound data, vibration data and temperature data of the machine;
generate a predictive maintenance model based on at least one of the preprocessed sound data, vibration data, and temperature data;
applying a decision logic to at least one of the collected sound data, vibration data, and temperature data to simulate the occurrence of an abnormality in the machine state,
A computer readable sequence comprising a sequence of instructions to derive a discount factor based on the simulation result, and to proceed with the learning of the predictive maintenance model through reinforcement learning using reinforcement learning based on the discount factor A computer program stored on a recording medium.

Images