KR102471427B1 - Predictive diagnosis system for mechanical and electrical failures through motor noise and thermal distribution analysis - Google Patents

Abstract

The present invention relates to a system for predicting and diagnosing mechanical and electrical failures through motor noise and heat distribution analysis, which automatically searches filter coefficients from the noise characteristics of field motors using a genetic algorithm (GA), applies the noise data from removing noise around the motor to the machine learning engine to diagnose the mechanical failure using the same, converts low-resolution heat data collected from field motors into high-resolution heat data using a heat distribution image learning engine, and then applies the heat data to the machine learning engine to diagnose electrical failures.

Description

Mechanical and electrical fault prediction and diagnosis system through motor noise and heat distribution analysis

The present invention relates to a system for predicting and diagnosing mechanical and electrical failures through motor noise and thermal distribution analysis, and more particularly, by collecting and analyzing noise data from machines, devices, and manufacturing equipment, predicting and diagnosing mechanical failures, and temperature data It relates to a system for predicting and diagnosing mechanical and electrical failures through motor noise and heat distribution analysis that collects and analyzes them to predict and diagnose electrical failures.

As production facilities have been automated to improve productivity, the reliability and safety of production facilities have become more important, and the need for maintenance technology for production facilities is increasing day by day.

In the detection and analysis of couplings in maintenance technology, a method for diagnosing failures using vibration data has been generally used. In the case of an induction motor, fault diagnosis using vibration data is a very powerful method for detecting and diagnosing most faults that may occur in an induction motor. However, for fault diagnosis using vibration data to have reliability, a large number of sensors are required and professional experience is required for accurate diagnosis.

When diagnosing failures of an induction motor using vibration data, not only does the difference in accuracy depend on the number and location of sensors installed, but also temperature data is added to detect electrical defects because vibration can only detect mechanical defects. However, it is a contact type and has a large number of sensors, so installation and management are still difficult.

Therefore, it is a trend to study non-contact methods that measure the noise generated from the induction motor using a microphone instead of vibration and measure the thermal image generated from the induction motor using a sensor. However, it is difficult to secure the reliability of the noise data because not only the noise generated by the induction motor itself but also the ambient noise is measured, and a high-resolution thermal imaging camera must be used to accurately measure the thermal image, but there is a problem of installation cost.

Republic of Korea Patent No. 10-2174722

The present invention is to solve the above problems of the prior art, and FFT data extracted by an adaptive filter algorithm that removes noise around an on-site induction motor and an image interpolation algorithm of a low-resolution sensor that is 1/10 the price of a conventional thermal imaging camera. (Fast Fourier Transform) To provide a mechanical and electrical failure prediction and diagnosis system through analysis of motor noise and heat distribution that can realize the convenience of application and accuracy of failure detection by artificial intelligence machine learning method made of spectogram images. There is a purpose.

Specifically, the filter coefficient of the noise characteristics of the motor in the field is automatically searched using a genetic algorithm (GA), and the noise data obtained by removing the noise around the motor is applied to the machine learning engine to diagnose mechanical failure. , Prediction and diagnosis of mechanical and electrical failures through motor noise and thermal distribution analysis, which converts low-resolution thermal data collected from field motors into high-resolution thermal data using a thermal distribution image learning engine and applies it to a machine learning engine to diagnose electrical failures It is an object of the invention to provide a system.

In addition, the technical problems to be achieved by the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the description below. You will be able to.

In order to achieve the above object, the system for predicting and diagnosing mechanical and electrical failures through motor noise and thermal distribution analysis of the present invention includes a data collection unit for collecting noise data input from a motor in the field; A filter coefficient search unit that automatically searches for filter coefficients from noise characteristics of field motors using a genetic algorithm; an adaptive filter unit for removing ambient noise from the noise data collected by the data collection unit by applying the filter coefficient automatically found by the filter coefficient search unit; a mechanical failure prediction unit that predicts a mechanical failure of a field motor by applying the noise data from which ambient noise has been removed through the adaptive filter unit to a first machine learning engine to determine whether the data is normal or abnormal; a thermal data collection unit that collects thermal data generated from an on-site motor using a thermal imaging camera or a thermal sensor; a thermal data converter converting the low-resolution thermal data collected by the thermal data collection unit into high-resolution thermal data using a thermal distribution image learning engine that learns a low-resolution thermal image and predicts a high-resolution thermal image; and an electrical failure prediction unit that applies the high-resolution thermal data converted by the thermal data conversion unit to a second machine learning engine to determine whether the data is normal data or abnormal data and predicts an electrical failure of the on-site motor.

In addition, an initial filter storage unit for storing a plurality of initial filter coefficients; and a filter combination generating unit that stores new filter coefficients generated from the initial filter coefficients using a genetic algorithm, wherein the genetic algorithm selects the initial filter coefficients stored in the initial filter storage unit until a stop condition is satisfied. After repeatedly performing selection, crossover, and mutation operations, a plurality of new filter coefficients are generated and stored in the filter combination generation unit.

The stop condition is characterized in that a condition for confirming a filter coefficient exhibiting a frequency response characteristic similar to that of a normal signal by applying a new filter coefficient obtained through selection, crossing, and mutation to a filter simulation equation.

The filter coefficient search unit simulates a frequency response curve by applying a plurality of LPFs (Low Pass Filters) and HPFs (High Pass Filters) stored in the filter combination generation unit to a filter simulation formula, and as a result, an LPF showing a result most similar to a normal signal and HPF are automatically searched.

The adaptive filter unit applies the noise data input from the motor in the field to the filter equations in the up, down, left, and right directions to which the filter coefficients LPF and HPF automatically found by the filter coefficient search unit are applied, respectively, and obtains four result values (Y _HPH, Y _LPH, Y _HPV, Y _LPV ), and using it to create four filtering result values (f _HPH, f _LPH, f _HPV, f _LPV ), and then calculating the average value, It is characterized by generating a final result value from which noise is removed.

In addition, it further includes a thermal distribution image storage unit for storing thermal distribution images for various types of motors according to motor types, capacities, and installation environments in a thermal distribution DB, and the thermal distribution image learning engine learns various thermal distribution images stored in the thermal distribution DB. And, the thermal data conversion unit is characterized in that it predicts a high-resolution image by applying the low-resolution thermal image collected by the thermal data collection unit to the thermal distribution image learning engine.

According to the system for predicting and diagnosing mechanical and electrical failures through motor noise and thermal distribution analysis according to the present invention, by using a genetic algorithm, filter coefficients for the noise characteristics of an on-site induction motor are automatically found, and noise around the induction motor is removed. Mechanical failure of an induction motor can be predicted by machine learning of noise data. In addition, after converting the low-resolution thermal image data generated by the on-site induction motor into high-resolution thermal image data, it is possible to predict electrical failure of the induction motor through machine learning.

A typical noise elimination filter has a problem in which a high or low pass is cut off due to a flattening phenomenon. It is possible to implement an adaptive filter by detecting and automatically setting it, and automatically changing the 3-Tab coefficients of the high-pass and low-pass filters according to changes in ambient noise even during normal operation.

1 is a block diagram of a system for predicting and diagnosing mechanical and electrical failures through motor noise and thermal distribution analysis according to the present invention.
Figure 2 is a flow chart showing a method for predicting mechanical failure in the mechanical and electrical failure prediction and diagnosis system through motor noise and thermal distribution analysis according to the present invention
3 is a 3-Tap filter coefficient used in the system for predicting and diagnosing mechanical and electrical failures through motor noise and thermal distribution analysis according to the present invention
4 is a frequency response curve of the 3-Tap filter coefficients of FIG. 2
5 is an exemplary diagram of outlier data detected by the system according to the present invention;
6 is an exemplary diagram for distinguishing abnormal symptom data in the system according to the present invention.
7 is a conceptual diagram showing how a similar learning model is created using data collected from similar devices according to the present invention and then transferred and used in a supervised learning model.
8 is a flowchart showing a method for predicting electrical failure in the system for predicting and diagnosing mechanical and electrical failures through motor noise and thermal distribution analysis according to the present invention.
9 is an exemplary view of a high-resolution image obtained by improving a low-resolution image in a system according to the present invention.

Advantages and features of the present invention, and methods for achieving them will become clear with reference to the detailed embodiments described later in conjunction with the accompanying drawings.

However, the present invention is not limited to the embodiments disclosed below and will be implemented in various different forms.

This embodiment in this specification is provided to complete the disclosure of the present invention, and to completely inform those skilled in the art of the scope of the invention to which the present invention belongs.

And the invention is only defined by the scope of the claims.

Thus, in some embodiments, well-known components, well-known operations and well-known techniques have not been described in detail in order to avoid obscuring the interpretation of the present invention.

In addition, like reference numerals designate like components throughout the specification, and terms used (referred to) in this specification are for describing embodiments and are not intended to limit the present invention.

In this specification, the singular also includes the plural unless specifically stated in the phrase, and components and operations referred to as 'comprising (or including)' do not exclude the presence or addition of one or more other components and operations. .

Unless otherwise defined, all terms (including technical and scientific terms) used in this specification may be used in a meaning commonly understood by those of ordinary skill in the art to which the present invention belongs.

In addition, terms defined in commonly used dictionaries are not interpreted ideally or excessively unless they are defined.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

1, 2 and 8, the mechanical and electrical failure prediction and diagnosis system 100 through motor noise and thermal distribution analysis diagnoses and predicts mechanical failures using noise data input from the motor M in the field, Predict electrical failures using thermal images of on-site motors. Here, on-site motors include those used in various environments such as buildings and instrumentation systems.

The mechanical and electrical failure prediction diagnosis system 100 according to the present invention is collected during a trial operation period (t0) and a normal operation period (t1) after the end of the trial operation period in order to solve the problem of insufficient noise data collected from motors in the field. It is possible to secure big data by applying different processing and learning methods for the noise data to be used, and to improve the noise anomaly detection performance of the motor.

The mechanical and electrical failure prediction and diagnosis system 100 according to the present invention includes a data collection unit 110, a filter coefficient search unit 130, an adaptive filter unit 140, a mechanical failure prediction unit 150, and a thermal data collection unit. 200, thermal data conversion unit 210, thermal distribution image storage unit 220 and electrical failure prediction unit 230, including data amplification unit 120, similar data storage unit 160, learning model transfer unit 170, a similar device search unit 180, and a discrimination data incremental learning unit 190 may be further included.

The data collection unit 110 collects noise data input from the motor M in the field during the trial operation period t0 and the normal operation period t1 after the trial operation period ends. The trial operation period (t0) may be about 2 to 3 weeks, but is not necessarily limited thereto, and the normal operation period (t1) is continuously measuring noise data through the operation of the actual on-site motor after the trial operation period (t0) is over. refers to the period during which it is collected.

When noise data is input from the motor M in the field through the data collection unit 110 during the trial operation period t0 (S110), the data amplification unit 120 converts the input noise data during the trial operation period t0. Amplify (S120).

The data amplification unit 120 amplifies the collected data to detect anomalies based on the big data technique because the trial operation period t0 is generally short and the amount of data collected during this period is small.

In the case of noise data targeted in the present invention, considering a situation in which a single class mainly consists of only normal data, the collected noise data is amplified using a Variational Auto Encoder (VAE) algorithm.

VAE (Variational Auto Encoder) is a type of auto encoder that copies input to output, and uses average and standard deviation to randomly sample input code values in the latent space and decode them again. By doing so, it performs an amplification function through data-like replication.

The filter coefficient search unit 130 automatically retrieves filter coefficients from noise characteristics of the field motor using a genetic algorithm (S130).

Genetic Algorithm (GA) is an algorithm proposed by John Holland in 1975 by imitating natural evolutionary phenomena such as crossover, mutation, and survival of the fittest. It is widely used in machine learning, etc. In this algorithm, a chromosome means an individual (corresponding to a noise reduction filter combination in the present invention) that solves a given problem, and its suitability is evaluated through a fitness function. The basic operating process of this algorithm is to repeat genetic operations such as selection, crossover, and mutation until a desired solution is found to create a new individual after evaluating the suitability of all individuals constituting the group.

The adaptive filter unit 140 removes ambient noise from the noise data amplified by the data amplification unit 120 by applying the filter coefficient automatically found by the filter coefficient search unit 130 (S140).

The mechanical failure prediction unit 150 applies the noise data from which ambient noise has been removed through the adaptive filter unit 140 to the first machine learning engine to determine whether it is normal data or abnormal data to predict mechanical failure of the on-site motor. do. Specifically, the mechanical failure predictor 150 includes a test operation period anomaly detection unit 151 and a normal operation period anomaly symptom prediction unit 152 . During the pilot operation period, the anomaly detection unit 151 detects anomalies by applying an unsupervised learning-based algorithm to the noise data from which the adaptive filter unit 140 applies filter coefficients to remove ambient noise. (S150), it is determined whether the corresponding data is normal data or abnormal data (S180).

Referring to FIG. 5 , abnormal data means data such as outliers. The left side shows the case of globally outliers, and the right side shows the case of regional outliers. Noise data is data to be removed in the preprocessing process, so it is different from outliers such as outliers.

Referring to FIG. 6, an example of abnormal symptom data is shown, and as shown in a box plot, a histogram, and a scatter plot from the left, it is far from the data belonging to the normal group. Able to know.

The normal operating period anomaly prediction unit 152 receives the data determined as normal or abnormal from the anomaly symptom detection unit 151 during the trial operation period and applies it to a supervised learning-based algorithm for learning (S170), The data collected during the normal operating period (t1) in the data collection unit 110 is applied to the learned supervised learning-based algorithm (S170) to determine whether it is normal data or abnormal data (S180).

If, for various reasons during the normal operating period (t1), if the noise data to be continuously input does not exist because the motor in the field is idle, stopped, or stopped (S160), the noise anomaly detection process of the motor is terminated.

The filter coefficient search unit 130 includes an initial filter storage unit 131 that stores a plurality of initial filter coefficients and a filter combination generation unit 132 that stores new filter coefficients generated from the initial filter coefficients using a genetic algorithm. do.

The genetic algorithm (GA) repeatedly performs selection, crossover, and mutation operations until the stop condition is satisfied for the initial filter coefficients stored in the initial filter storage unit 131, and then a plurality of new Filter coefficients are generated and stored in the filter combination generating unit 132 . The filter coefficients of FIG. 3 represent a plurality of filter coefficients stored in the filter combination generating unit 132 . Initial filter coefficients are specified by the user. Users can specify LPF or HPF, or both LPF and HPF. Selection, intersection, and mutation operations are repeatedly performed using a plurality of (n) filter coefficients designated by the user to generate a plurality of (k) filter coefficients that satisfy the stopping condition. The best solution among the generated k filter coefficients is selected as the answer. 3 is an example in which k = 7 filter coefficients satisfying the stop condition are generated.

Here, the stopping condition refers to checking filter coefficients exhibiting frequency response characteristics similar to those of a normal signal by applying new filter coefficients obtained through selection, crossing, and mutation to the filter simulation formula.

Equation 1 below is an example in which the filter coefficients [1 2 1] and [-1 2 -1] of FIG. 3 are applied to the filter simulation equation.

[Formula 1]

By simulating this, frequency response characteristics as shown in FIG. 4 can be confirmed.

4 shows frequency response curves of 3-Tap filter coefficients. Figure 4 (a) shows the frequency response curve of the LPF, (b) shows the frequency response curve of the HPF.

The filter coefficient search unit 130 applies a plurality of LPFs (Low Pass Filters) and HPFs (High Pass Filters) stored in the filter combination generator 132 to the filter simulation formula to simulate the frequency response curve. As a result, the normal signal and The LPF and HPF with the most similar results are automatically searched.

According to the present invention, the filter coefficient search unit 130 applies the 7 LPFs and HPFs stored in the filter combination generator 132 to the filter simulation Equation 1 to simulate the frequency response curve as shown in FIG. It can be seen that the frequency response curve of the filter coefficient [1 1] and the [1 2 1] filter coefficient among the filter coefficients generated by the present invention are the most similar. It can be seen that the frequency response curves of the [-1 1] and [-1 2 -1] filter coefficients are most similar to the HPF.

The adaptive filter unit 140 applies the filter coefficients (LPF [1 2 1], HPF [-1 2 -1]) automatically found by the filter coefficient search unit 130 through the frequency response curve simulation to data data. Ambient noise is removed from noise data amplified by the amplifier 120 .

Specifically, the adaptive filter unit 140 converts the noise data input from the motors in the field into filter equations in the up, down, left, and right directions to which the filter coefficients LPF and HPF automatically found by the filter coefficient search unit 130 are applied. 4 result values (Y _HPH, Y _LPH, Y _HPV, Y _LPV ) of Equation 1 are generated by applying each, and the four result values (f _HPH, f _LPH, f _HPV, f _LPV ) filtered using this are generated. By obtaining an average value after generation, a final result value obtained by removing ambient noise from the noise data is generated. The formula for generating the final result value is shown in Equation 2 below.

[Equation 2]

The similar data storage unit 160 collects and stores data from external institutions or stores data collected from similar devices published on the Internet in a similar data DB.

Referring to FIGS. 1 and 2 , the similar data storage unit 160 collects similar data from similar domains (external public institutions or the Internet) and stores them in a similar data DB.

Considering the situation in which the data collected during the pilot operation period (t0) is small and the lack of learning data at the beginning even during the normal operation period (t1), it is difficult to determine whether or not an anomaly is detected in the continuously collected data, and is disclosed on the Internet. Utilizing data collected from similar devices, or data on similar devices possessed by external public institutions, a learning model is created and transferred to be used as a model for supervised learning classification. Do (S170).

The present invention diagnoses and predicts mechanical failure using noise data input from the motor M in the field. Field motors refer to motors used in various environments such as buildings and instrumentation systems.

The above similar devices correspond to devices such as air conditioners, heat exchangers, pumps, and elevators in the case of buildings, and devices such as water purification plants, pressurization plants, and reservoirs in the case of instrumentation systems.

The learning model transfer unit 170 uses the data stored in the similar data DB through the above process to create a similar learning model, transfer learning, and use it as the supervised learning model (S190). With transfer learning, it is possible to learn from a similar domain and apply the trained model to the current supervised learning.

Looking briefly at the concept of transfer learning, when training a large-scale deep learning model, a slow learning rate occurs when learning from scratch. In this case, when there is a previously trained similar model, importing and reusing the lower layer of this model not only speeds up the learning speed, but also reduces the training set required for learning.

The learning model transfer unit 170 transfers and reuses a portion of the lower layer of the similar learning model, and creates a supervised learning model by learning the final classifier layer. .

Referring to FIG. 7, assuming that a motor in the field is used to drive a freight elevator in a building, a similar learning model is created by using noise data for a freight elevator among data stored in a similar data DB, Hidden Layer 1 and Hidden Layer 2 are transferred to be reused in the supervised learning model, and the supervised learning model is created by additionally learning only the final classifier layer, Hidden Layer 3. Since Hidden Layer 1 and Hidden Layer 2 are used after transition, the weight is fixed. If there is no noise data for freight elevators among the data stored in the similar data DB, noise data for the most similar small freight elevators or passenger elevators may be used. Even assuming that the on-site motor is used to drive the water purification plant of the instrumentation system, a similar learning model is created and transferred by using the noise data for the water purification plant among the data stored in the similar data DB to be reused in the supervised learning model, and only the final classifier layer is used. By additionally learning, a supervised learning model is created.

In this way, by performing transfer learning, the speed of learning can be increased or the performance of the final model can be improved compared to the case of learning anew from the beginning.

At this time, the similar data storage unit 160 classifies the collected data according to the type and weight of the device and stores them in the similar data DB. It is classified by type of device (air conditioner, heat exchanger, pump, elevator, etc. of a building, and water purification plant, pressurization plant, reservoir, etc. of instrumentation system) and weight (800kg, 1 ton, 1.5 ton, etc.) and stored in similar data DB. .

The learning model transition unit 170 is a situation where the data collected during the pilot operation period (t0) is small and abnormal symptoms of data that are continuously collected due to the lack of learning data at the beginning even during the normal operation period (t1) after the end of the pilot operation period Considering the situation where it is difficult to determine whether or not detection is detected, a device similar in type and weight to the motor (M) at the site is searched in the similar data DB, and a similar learning model is created and transferred to a supervised learning model using the data of the searched device. let it be used

Assuming that the motor (M) at the site is a motor used in a building air conditioner, noise data for the air conditioner is searched in the similar data DB, and then the data with the closest weight is found, a similar learning model is created, and it is transferred and used. Assuming that the motor (M) in the field is a motor used in the reservoir of the instrumentation system, noise data for the reservoir is searched in the similar data DB, and then the data with the closest weight is found, a similar learning model is created, and it is transferred and used. .

The similar device search unit 180 compares the similar data DB according to a predetermined criterion to search for data of a device quantitatively similar in type and weight to the motor at the site. The preset criteria are as follows.

For example, if the motor at the site is a motor used in an elevator in a building, the noise data for the elevator is first searched in the similar data DB, and then the weight is searched. If there is noise data about the elevator and the weight differs by less than 15% from the site motor (M), the data is selected and used.

In other words, it is most important that the type of device (elevator, air conditioner, water purification plant, reservoir, etc.) used by the motor (M) is similar. The weight of the device is also an important consideration.

The learning model transition unit 170 utilizes device data retrieved by the similar device search unit 180 to generate and transfer a similar learning model to be used as a supervised learning model. At this time, the similar device search unit 180 scores the degree of similarity for the type and weight of the motor M in the field and the device stored in the similar data DB, and converts the similarity learning model to the supervised learning model according to the similarity score. The number of can be set differently. Here, the similarity score is based on the score of the device most similar in kind and weight to the motor M in the field.

For example, referring to FIG. 7, if the similarity is high enough to have a similarity score of 70 points or more, a supervised learning model is generated by transferring and using Hidden Layer 1 and Hidden Layer 2, and additionally learning only Hidden Layer 3, which is the final classifier layer. However, if the similarity score is about 60 points, only Hidden Layer 1 is transferred and used, and a supervised learning model can be created by additionally learning Hidden Layer 2 and Hidden Layer 3. In FIG. 7, the supervised learning model has three layers as an example, and it may have a much larger number of layers than this, and the number of layers to be transferred and used may vary according to the similarity score.

The discrimination data incremental learning unit 190 applies the data collected during the normal operating period (t1) in the data collection unit 110 to a supervised learning-based algorithm (S170) to determine whether it is normal data or abnormal data ( S180) This is provided again to be used as training data for the supervised learning-based algorithm. By predicting whether the input data is normal or abnormal, and providing the classified normal and abnormal data to be used again as training data for the supervised learning model, quantitative and qualitative reinforcement of the learning data continues to improve the quality of new data in the data collection situation. Discrimination (or prediction) performance can be improved.

Referring to FIGS. 8 and 9 , a method of predicting an electrical failure will be described.

The thermal data collection unit 200 collects thermal data generated from a motor in the field using a thermal imaging camera or a thermal sensor (S810).

The thermal data conversion unit 210 converts the low-resolution thermal data collected by the thermal data collection unit 200 into high-resolution thermal data using a thermal distribution image learning engine that learns the low-resolution thermal image and predicts a high-resolution thermal image ( S820).

The electrical failure prediction unit 230 applies the high-resolution thermal data converted by the thermal data conversion unit 210 to the second machine learning engine to determine whether the data is normal or abnormal, and predicts electrical failure of the on-site motor (S830). ).

The heat distribution image storage unit 220 stores heat distribution images of various types of motors according to motor types, capacities, and installation environments in a heat distribution DB.

The thermal distribution image learning engine learns various thermal distribution images stored in the thermal distribution DB, and the thermal data conversion unit 210 applies the low-resolution thermal image collected by the thermal data collection unit 200 to the thermal distribution image learning engine to obtain a high-resolution image. predict

The low-resolution thermal image sensor has a lower resolution than the handheld type, and is about 80x60. As the price is 1/10, it is suitable for induction motor thermal distribution analysis. However, a preprocessing process is required for accurate analysis. The present invention implements a function of changing a low-resolution image to a high-resolution one through AI learned from an induction motor image captured by a high-resolution thermal imaging camera. Referring to FIG. 9 , an example in which the thermal data conversion unit 210 converts a low-resolution induction motor image into a high-resolution image using a thermal distribution image learning engine is shown.

The AI upscaling method is a method in which a deep learning model predicts a high-resolution image by looking at a low-resolution image. To this end, it is necessary to obtain various thermal distribution images of similar induction motors and make AI learn.

The electrical failure prediction unit 230 performs a function of improving a low-resolution image into a high-resolution image by using SRCNN. SRCNN (Super-Resolution Convolutional Neural Network) is a method of improving image quality through a convolutional neural network after growing an image using bicubic, an interpolation method.

The present invention builds a thermal distribution image database for each induction motor type/capacity/installation position to improve speed and accuracy by learning differently depending on the application target to enable induction motor shape prediction and optimization.

The present invention diagnoses mechanical failure and electrical failure of an induction motor through machine learning after converting the improved image of noise data and thermal image data filtered through a preprocessing process into an FFT spectogram.

Although the present invention has been described in detail with reference to preferred embodiments so far, those skilled in the art to which the present invention pertains can implement the present invention in other specific forms without changing the technical spirit or essential features, The embodiments are to be understood as illustrative in all respects and not limiting.

In addition, the scope of the present invention is specified by the claims to be described later rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts are included in the scope of the present invention. should be interpreted as

100...Mechanical and electrical failure prediction and diagnosis system through motor noise and heat distribution analysis
110 ... data collection unit
120 ... data amplification unit
130 ... filter coefficient search unit
131 ... initial filter reservoir
132 ... filter combination generating unit
140 ... adaptive filter unit
150 ... mechanical failure prediction unit
151... Anomaly detection unit during pilot operation period
152 ... normal operating period abnormal symptom detection unit
160 ... similar data storage unit
170 ... learning model transfer part
180 ... similar device search unit
190 Discrimination data incremental learning unit
200 ... thermal data collection unit
210 ... thermal data conversion unit
220 ... thermal distribution image storage unit
230 ... electrical failure prediction unit

Claims (6)

Data collection unit for collecting noise data input from the motor in the field;
A filter coefficient search unit that automatically searches for filter coefficients from noise characteristics of field motors using a genetic algorithm;
an adaptive filter unit for removing ambient noise from the noise data collected by the data collection unit by applying the filter coefficient automatically found by the filter coefficient search unit;
a mechanical failure prediction unit that predicts a mechanical failure of a field motor by applying the noise data from which ambient noise has been removed through the adaptive filter unit to a first machine learning engine to determine whether the data is normal or abnormal;
a thermal data collection unit that collects thermal data generated from an on-site motor using a thermal imaging camera or a thermal sensor;
a thermal data converter converting the low-resolution thermal data collected by the thermal data collection unit into high-resolution thermal data using a thermal distribution image learning engine that learns a low-resolution thermal image and predicts a high-resolution thermal image; and
An electrical failure prediction unit that applies the high-resolution thermal data converted by the thermal data conversion unit to a second machine learning engine to determine whether it is normal data or abnormal data and predicts an electrical failure of the on-site motor;
The adaptive filter unit applies the noise data input from the motor in the field to the filter equations in the up, down, left, and right directions to which the filter coefficients LPF and HPF automatically found by the filter coefficient search unit are applied, respectively, and obtains four result values (Y _HPH, Y _LPH, Y _HPV, Y _LPV ), and using it to create four filtering result values (f _HPH, f _LPH, f _HPV, f _LPV ), and then calculating the average value, A system for predicting and diagnosing mechanical and electrical failures through motor noise and heat distribution analysis, characterized in that the final result value is generated by removing the noise.
According to claim 1,
an initial filter storage unit that stores a plurality of initial filter coefficients; and
Further comprising a filter combination generation unit for storing new filter coefficients generated from initial filter coefficients using a genetic algorithm;
The genetic algorithm repeatedly performs selection, crossover, and mutation operations on the initial filter coefficients stored in the initial filter storage until a stop condition is satisfied, and then generates a plurality of new filter coefficients, , Mechanical and electrical failure prediction and diagnosis system through motor noise and heat distribution analysis, characterized in that stored in the filter combination generation unit.
According to claim 2,
The stop condition is a condition for confirming a filter coefficient exhibiting a frequency response characteristic similar to that of a normal signal by applying a new filter coefficient obtained through selection, crossover, and mutation to the filter simulation formula. and an electrical failure prediction and diagnosis system.
In the third,
The filter coefficient search unit simulates a frequency response curve by applying a plurality of LPFs (Low Pass Filters) and HPFs (High Pass Filters) stored in the filter combination generation unit to a filter simulation formula, and as a result, an LPF showing a result most similar to a normal signal A system for predicting and diagnosing mechanical and electrical failures through motor noise and heat distribution analysis, characterized by automatically searching for and HPF.
delete In the first
A heat distribution image storage unit for storing heat distribution images of various types of motors according to motor types, capacities, and installation environments in a heat distribution DB;
The thermal distribution image learning engine learns various thermal distribution images stored in a thermal distribution DB, and the thermal data converter applies the low-resolution thermal image collected by the thermal data collection unit to the thermal distribution image learning engine to predict a high-resolution image. Mechanical and electrical failure prediction and diagnosis system through thermal distribution analysis.

Images