KR102473124B1 - Error detection apparatus and method of motor - Google Patents

Abstract

An apparatus for detecting a failure of a motor according to a preferred embodiment of the present invention includes a resampling unit that performs resampling on a phase current signal of the motor and converts it into a resampling signal, a conversion unit that preprocesses the resampling signal and converts it into a preprocessed signal, and the preprocessed signal. It includes a learning unit that learns failure characteristic data through and a determination unit that determines whether the motor is out of order using the failure characteristic data.

Description

Motor failure detection device and method {ERROR DETECTION APPARATUS AND METHOD OF MOTOR}

The present invention relates to an apparatus and method for detecting a failure of a motor.

An electric motor is a device that converts electrical energy into mechanical energy and is used in various industrial fields including fans, robots, electric vehicles, and pumps. In the industrial field, AC motors using AC power are mainly used and are widely used because they have advantages of relatively high efficiency and long life compared to DC motors.

Fault diagnosis is essential to ensure smooth operation and safety of motors. In particular, if a failure of some motors occurs in an industrial line, it may cause a huge economic loss by causing the entire line to be stopped. Accordingly, developing a reliable and effective motor failure diagnosis technique is considered an important issue.

In general, for fault diagnosis of electric motors, fault diagnosis algorithms based on industrial robots and electric vehicles operating at constant speed are used.

As a conventional fault diagnosis algorithm, a Motor Signature Current Analysis (MCSA) method is applied that analyzes the stator current of the motor to extract fault frequency-related characteristics and then diagnoses the fault of the motor using the extracted characteristic frequency. .

Conventional fault diagnosis algorithms are applicable only to constant velocity and equal load signals that can be converted into a frequency domain. This is because it is impossible to obtain a fault characteristic frequency through a frequency analysis method because the frequency of the current also changes according to the variable rotational speed due to the variable speed operation of the motor. In addition, information on motor parameters (eg, number of poles, bearing information, etc.) is required to obtain the failure characteristic frequency. In addition, as the magnitude and frequency of the current of the motor continuously change, there are many cases in which similar failure characteristic frequencies occur despite different failures, so there is a problem in accurately diagnosing failures.

To solve this problem, a fault diagnosis method based on deep learning has recently been developed. Deep learning is a model created by stacking several neural networks in mathematical form, and performs the function of automatically learning the characteristics of data and classifying the learned data.

The deep learning-based fault diagnosis method has high diagnostic performance and is widely used for fault diagnosis, but a large amount of data is required to learn a model, and the model is excessively limited to the learned data, so there is a possibility that the generalization performance is not good. However, there is a disadvantage that the interpretation of diagnosis results is difficult.

Republic of Korea Patent Registration No. 10-2189269

Accordingly, the present invention has been devised in view of the above circumstances, and it is possible to diagnose faults using the magnitude and fault characteristic frequency of the phase current signal of the motor, which varies according to the operating conditions of the motor, and faults using a convolution-based deep learning learning algorithm. An object of the present invention is to provide an apparatus for detecting a failure of a motor to which a current data pre-processing method capable of improving classification and generalization performance is applied.

An apparatus for detecting a failure of a motor according to a preferred embodiment of the present invention for achieving the above object includes: a resampling unit that performs resampling on a phase current signal of the motor and converts it into a resampling signal; a converter for preprocessing the resampling signal and converting the resampling signal into a preprocessing signal; a learning unit learning failure characteristic data through the preprocessing signal; and a determination unit determining whether the motor is out of order by using the failure characteristic data.

The resampling unit may convert the phase current signal into an analysis signal, estimate an instantaneous angle from the analysis signal, and separate the analysis signal for each period based on a change point of the instantaneous angle.

The resampling unit may generate the resampling signal by resampling the analysis signal separated by period through a prearranged interpolation method into signals of equal intervals and connecting the signals of equal intervals.

The transform unit may perform a log scale fast Fourier transform on the resampling signal.

The transform unit may generate the preprocessed signal by performing scaling on a magnitude of a signal according to the fast Fourier transform.

The scaling unit may have a value of 0 to 1.

The learning unit may learn the preprocessed signal using a previously prepared convolution-based deep learning learning algorithm.

The determination unit may determine an axis imbalance or a winding deterioration failure of the motor using the failure characteristic data.

The determination unit may analyze the failure determination result of the motor using a previously prepared CAM (Class Activation Map) based analysis algorithm.

A method for detecting a failure of a motor according to a preferred embodiment of the present invention for achieving the above object includes a resampling step of performing resampling on a phase current signal of the motor and converting it into a resampling signal; a preprocessing step of preprocessing the resampling signal and converting the resampling signal into a preprocessing signal; A learning step of learning failure characteristic data through the preprocessing signal; and a failure determination step of determining whether or not the motor has failed using the failure characteristic data.

The resampling step may include converting the phase current signal into an analysis signal; estimating an instantaneous time from the analysis signal; Separating the analysis signal by period based on the change point of the instantaneous time; Resampling the analysis signal separated by period into a signal of the same interval through a pre-arranged interpolation method; and generating the resampling signal by connecting the signals at equal intervals.

The pre-processing step may include: a fast Fourier transform step of performing a log-scale fast Fourier transform on the resampling signal; and a scaling step of performing scaling on the magnitude of the signal according to the fast Fourier transform.

In the learning step, the preprocessed signal may be learned using a previously prepared convolution-based deep learning learning algorithm.

In the failure determination step, an axis imbalance or a winding deterioration failure of the motor may be determined using the failure characteristic data.

In the failure determination step, a failure determination result of the motor may be analyzed using a previously prepared CAM (Class Activation Map) based analysis algorithm.

According to the motor failure detection apparatus and method according to a preferred embodiment of the present invention, it is possible to diagnose failures using the magnitude and failure characteristic frequency of a motor current signal that varies depending on the operating conditions of the motor, and a convolution-based deep learning learning algorithm. A current data pre-processing method that can enhance fault classification and generalization performance using the applied current data has an effect of enabling real-time fault diagnosis for motors under variable speed operation and various load conditions.

In addition, since the characteristic frequency of the motor is automatically extracted through the convolution-based deep learning algorithm, prior knowledge related to motor parameters is not required, and fault detection is performed using relatively low frequency domain information (e.g., 4*pole*60Hz). Because diagnosis is made, a high-performance current measurement sensor is not required, which has the effect of reducing cost.

In addition, the decision result of the convolution-based deep learning learning algorithm can be interpreted through a post-processing algorithm, so that an expert judges the consistency of the algorithm.

In addition, there is an effect that can be applied to dynamic motor systems (eg, electric vehicles and industrial robots) in which failure diagnosis with vibration signals is difficult.

In addition, there is an effect of identifying the signal failure mode without disassembling the motor in which the failure occurred.

1 is a block diagram of an apparatus for detecting a failure of a motor according to a preferred embodiment of the present invention.
2 is a diagram showing an example of a speed profile of an electric motor.
FIG. 3 is a diagram showing signal processing results for phase currents of a motor operating at a constant speed according to the speed profile of FIG. 2 .
FIG. 4 is a diagram showing signal processing results for phase currents of an electric motor operating with variable speed according to the speed profile of FIG. 2 .
5 is a diagram showing the structure of a convolutional neural network of a learning unit according to a preferred embodiment of the present invention.
6 is a diagram showing an analysis result of analyzing failure characteristic data related to an internal winding deterioration failure of a motor.
7 is a diagram showing analysis results for failure characteristic data for which resampling is not performed in a matrix form.
8 is a diagram showing analysis results for failure characteristic data for which resampling has been performed in a matrix form.
9 is a flow chart of a motor failure detection method according to a preferred embodiment of the present invention.
FIG. 10 is a diagram for explaining the resampling step of FIG. 9 in detail.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. First, in adding reference numerals to components of each drawing, it should be noted that the same components have the same numerals as much as possible even if they are displayed on different drawings. In addition, although preferred embodiments of the present invention will be described below, the technical idea of the present invention is not limited or limited thereto and can be modified and implemented in various ways by those skilled in the art.

1 is a block diagram of an apparatus for detecting a failure of a motor according to a preferred embodiment of the present invention.

Referring to FIG. 1 , an apparatus 100 for detecting a failure of a motor according to a preferred embodiment of the present invention obtains a phase current signal of the motor, performs resampling on the phase current signal, and performs fast Fourier transform and scaling on the resampling signal. It is characterized by performing, learning failure characteristic data through fast Fourier transform and scaled preprocessing signals, and determining whether the motor is out of order by considering the failure characteristic data.

The motor failure detection apparatus 100 includes a receiving unit 110, a resampling unit 120, a conversion unit 130, a learning unit 140, and a determination unit 150.

The receiving unit 110 may receive the phase current signal of the motor from the sensing device 200 . Here, the sensing device 200 may sense phase current signals flowing in windings of the motor. The motor may be a surface type permanent magnet synchronous motor, but is not limited thereto. The sensing device 200 may be a non-contact type sensing device that is easy to install. In one embodiment, the sensing device 200 may measure the phase current current by attaching an AC current probe to the three-phase current line of the surface-type permanent magnet motor.

The resampling unit 120 may resample the phase current signal of the motor and convert it into a resampling signal. To this end, the resampling unit 120 may convert the phase current signal of the motor into an analytic signal using a pre-prepared Hilbert transform. The resampling unit 120 may estimate the instantaneous time from the analysis signal. The resampling unit 120 may detect a change point of the estimated instantaneous angle. The resampling unit 120 may detect whether the change point is equal to or greater than a predetermined boundary line. The resampling unit 120 may separate the analysis signal for each cycle based on the corresponding change point when the corresponding change point is equal to or greater than a predetermined boundary line. The resampling unit 120 may resample each analysis signal separated by period into signals at equal intervals through cubic spline interpolation. The resampling unit 120 may generate a resampling signal by connecting signals having equal intervals to each other. A detailed process and resampling process of the resampling unit 120 can be confirmed through FIG. 10 .

The conversion unit 130 may decompose the resampling signal by performing real-time extraction in a sliding window method in order to use it as input data of a convolution-based deep learning algorithm. The transform unit 130 may convert the signal-decomposed resampling signal into a fast Fourier transform signal using fast Fourier transform (FFT). In this case, the transform unit 130 may set a scale of the fast Fourier transform signal based on a log unit. To this end, the conversion unit 130 may select a relatively low sampling frequency region. Here, the selected frequency range may be approximately 4 times or less of the electrical angular frequency. The transform unit 130 may perform min-max scaling of the fast Fourier transform signal. In one embodiment, the minimum-maximum scaling is for setting a minimum and maximum value of failure characteristic data to be described later, and may be set to a value between approximately 0 and 1. Through this, the size of the failure characteristic frequency region having a relatively small intensity compared to the operation component of the motor can be increased. Then, the scaled fast Fourier transform signal is defined as a preprocessing signal.

The learning unit 140 may learn failure characteristic data through a preprocessing signal input to a convolution-based deep learning algorithm. Parameter information used in the convolution-based deep learning learning algorithm may appear as shown in Table 1 below.

layer name Parameter information number of parameters Convld 1 kernel size= 8, channel : 64 576 BatchNormalization 1 moment=0.01 128 Convld 2 kernel size= 5, channel : 128 41088 BatchNormalization 2 moment=0.01 256 Convld 3 kernel size= 3, channel : 64 24640 BatchNormalization 3 moment=0.01 128 Dense Node number : 3 195 all 67011

In Table 1, the parameters of the final Dense layer are variable according to the number of failure modes to be diagnosed. In one embodiment, the convolution-based deep learning learning algorithm according to the present invention may set a parameter to 3 to classify two failure modes and a normal situation. Here, two failure modes may include axis misalignment and winding deterioration (inter-turn short).

The determination unit 150 may determine whether the motor is out of order by using the failure characteristic data. The determination unit 150 may provide reliability of the failure determination result by interpreting the failure determination result through a CAM (Class Activation Map) based analysis algorithm. In one embodiment, failure characteristic data related to the deterioration failure of the internal winding of the motor can be confirmed through FIG. 6 .

Hereinafter, effects obtained through resampling of the phase current of the motor will be described.

2 is a diagram showing an example of a speed profile of an electric motor.

Referring to FIG. 2 , an example of RPM (Revolution Per Minute) according to the speed profile of the motor can be confirmed. When the motor operates with such a speed profile, the frequency of the phase current of the motor may be changed by a PWM (Pulse Width Modulation) inverter.

FIG. 3 is a diagram showing signal processing results for phase currents of a motor operating at a constant speed according to the speed profile of FIG. 2 . The scale unit of FIG. 3 may be logarithmic.

Figure 3 (b) is a diagram showing a general fast Fourier transform result for the phase current of the motor operating at a constant speed according to the speed profile of FIG. That is, if the fast Fourier transform is performed in the frequency domain without additional signal processing for the phase current of the motor, as shown in FIG. Here, the fault frequency is composed of a function related to the rotational speed of the motor. When the time axis of the phase current is changed to the angle axis through fast Fourier transform, the fault frequency and normal frequency of the phase current can be aligned even in various speed changes of the motor.

FIG. 3(c) is a diagram showing a result of fast Fourier transforming the resampling signal after resampling the phase current of the motor operating at constant speed according to the speed profile of FIG. 2 . The resampling unit 120 may resample the phase current of the motor to change phase current data, which is conventionally configured as a function of time and current, into a function of rotation angle and current. At this time, frequency components are gathered according to the rotational speed of the motor, and it is possible to determine the failure of the motor through the corresponding components.

FIG. 4 is a diagram showing signal processing results for phase currents of an electric motor operating with variable speed according to the speed profile of FIG. 2 .

Figure 4 (a) is a diagram showing a general fast Fourier transform result for the phase current of the motor operating with variable speed according to the speed profile of FIG.

As shown in (a) of FIG. 4, when the magnitude is expressed on a general scale after performing fast Fourier transform on the phase current of the motor, since the operating component is the main component, only the operating component has a very large value and other harmonics However, the sideband is relatively close to 0, resulting in poor learning performance during deep learning.

FIG. 4(b) is a diagram showing a result of fast Fourier transforming the resampling signal after resampling the phase current of the motor operating with variable speed according to the speed profile of FIG. 2 . The scale unit of FIG. 4(b) may be log.

As shown in (b) of FIG. 4 , when the fast Fourier transform signal is scaled in a logarithmic unit, peripheral frequency and harmonic components that are smaller in magnitude than the operating component can be emphasized. Then, for normalization, the magnitude of the fast Fourier transform signal may be re-transformed between 0 and 1 based on the maximum and minimum values of all data.

In order to use the preprocessing signal for fault diagnosis, it is necessary to check the size and presence of the fault characteristic frequency. However, there is a problem in that it is difficult to determine where the location of the failure characteristic frequency is and how large it is to be determined as failure unless a management expert skilled in the corresponding motor is present. In addition, when the load is changed, the size of the FFT component changes, and it is almost impossible to diagnose by grasping the size of all fault components according to the load change.

Accordingly, in order to solve this problem, the motor failure detection apparatus 100 according to a preferred embodiment of the present invention uses a convolution-based deep learning learning model that automatically extracts characteristics of each data and proceeds with learning.

5 is a diagram showing the structure of a convolutional neural network of a learning unit according to a preferred embodiment of the present invention.

Referring to FIG. 5 , the convolution-based deep learning algorithm can learn the frequency characteristics of phase current through three one-dimensional convolution layers (Conv 1d), and can solve the bottleneck of parameters through a global average pooling layer (GAP). and finally classify the fault frequency related to the fault of the motor through FC (Fully Connected Layer). Through this, failure characteristic data including classified failure frequencies may be generated.

6 is a diagram showing an analysis result of analyzing failure characteristic data related to an internal winding deterioration failure of a motor.

Referring to FIG. 6 , the CAM-based analysis algorithm according to a preferred embodiment of the present invention is a 1D CAM algorithm using an existing CAM technique and is configured to analyze the frequency domain of failure characteristic data. Through this, the CAM-based interpretation algorithm can display the frequency domain identified as the focus in the failure characteristic data in red through the heat map, and indicate the part that is hardly utilized in the failure characteristic data in blue through the heat map. Through this heat map, the user can interpret the judgment result of the CAM-based analysis algorithm, and the reliability of the judgment result in the industrial field can be increased by securing interpretability.

7 is a diagram showing analysis results for failure characteristic data for which resampling is not performed in a matrix form.

Referring to FIG. 7 , the confusion matrix may indicate the results of verification of the normal condition of the motor and two failure modes (winding deterioration and shaft imbalance). At this time, as learning progresses on the preprocessed signal for which resampling is not performed, the normal state of the matrix data appears at about 50%, the inter-turn short state appears at about 64%, and the axis imbalance ( It can be seen that the classification performance in which the misalignment state is approximately 80%. That is, it can be confirmed that the learning and classification performance is degraded.

8 is a diagram showing analysis results for failure characteristic data for which resampling has been performed in a matrix form.

Referring to FIG. 8 , the confusion matrix may indicate the result of verification of the normal condition of the motor and two failure modes (winding deterioration and shaft imbalance). At this time, as learning progresses on the resampled preprocessing signal, classification performance exceeding 95% can be confirmed in all of the normal, misaligned, and inter-turn short states of the matrix data. . That is, features for each fault frequency are properly extracted, and high learning and verification performance can be confirmed.

9 is a flow chart of a motor failure detection method according to a preferred embodiment of the present invention.

Referring to FIG. 9 , in the motor failure detection method according to a preferred embodiment of the present invention, phase current signal acquisition step (S910), resampling step (S920), fast Fourier transform step (S930), scaling step (S940), learning A step S950 and a failure determination step S960 may be included.

In step S910 of acquiring the phase current signal, the receiving unit 110 may acquire the phase current signal of the motor. Here, the phase current signal of the motor may be measured by a separate sensing device 200 .

In the resampling step (S920), the resampling unit 120 may resample the phase current signal of the motor and convert it into a resampling signal. The resampling step (S920) will be described later in detail with reference to FIG. 10 .

In the fast Fourier transform step (S930), the transform unit 130 may perform fast Fourier transform on the resampling signal. The transform unit 130 may decompose the resampling signal in a sliding window method prior to fast Fourier transform. The transform unit 130 may perform fast Fourier transform on the signal decomposed resampling signal in a decibel (DB) scale.

In the scaling step ( S940 ), the transform unit 130 may perform minimum-maximum scaling on the size of a relatively low frequency domain of the fast Fourier transform signal. The min-max scaling can be set to approximately a logarithmic value between 0 and 1.

In the learning step (S950), the learning unit 140 may learn the scaled preprocessed signal using a convolution-based deep learning learning algorithm prepared in advance. The learning unit 140 may learn failure characteristic data through a preprocessing signal.

In the failure determination step ( S960 ), the determination unit 150 may determine whether the motor has failed using the failure characteristic data. The determination unit 150 may analyze the failure determination result using a pre-prepared CAM-based interpretation algorithm.

FIG. 10 is a diagram for explaining the resampling step of FIG. 9 in detail.

Referring to FIG. 10 , in the resampling step ( S920 ) of FIG. 9 , first, the resampling unit 120 may convert the original phase current signal into an analysis signal by using the Hilbert transform. The resampling unit 120 may estimate the instantaneous angle Φ(t) from the analysis signal. The resampling unit 120 may separate or decompose the original current signal (interpretation signal) for each period (f0, f1) based on the change point of the instantaneous angle. The resampling unit 120 may resample each analysis signal separated by period into signals at equal intervals through cubic spline interpolation. The resampling unit 120 may generate a resampling signal by connecting signals having equal intervals.

The above description is merely an example of the technical idea of the present invention, and those skilled in the art can make various modifications, changes, and substitutions without departing from the essential characteristics of the present invention. will be. Therefore, the embodiments disclosed in the present invention and the accompanying drawings are not intended to limit the technical idea of the present invention, but to explain, and the scope of the technical idea of the present invention is not limited by these embodiments and the accompanying drawings. .

Steps and/or actions in accordance with the present invention may occur concurrently in different embodiments, in different orders, or in parallel, or for different epochs, etc., as would be understood by one skilled in the art. can

In some embodiments, some or all of the steps and/or operations may include instructions, programs, interactive data structures, clients and/or servers stored on one or more non-transitory computer-readable media. At least some of them may be implemented or performed using one or more processors that do. The one or more non-transitory computer-readable media may illustratively be software, firmware, hardware, and/or any combination thereof. Additionally, the functions of a “module” discussed herein may be implemented in software, firmware, hardware, and/or any combination thereof.

100: failure detection device
110: receiver
120: resampling unit
130: conversion unit
140: learning unit
150: judgment unit
200: sensing device

Claims (15)

a resampling unit that performs resampling on the phase current signal of the motor and converts it into a resampling signal;
a converter for preprocessing the resampling signal and converting the resampling signal into a preprocessing signal;
a learning unit learning failure characteristic data through the preprocessing signal; and
a determination unit determining whether the motor is out of order using the failure characteristic data;
including,
The resampling unit converts the phase current signal into an analysis signal using a previously prepared Hilbert transform, and estimates an instantaneous angle from the analysis signal. Motor failure detection device. According to claim 1,
The resampling unit,
Motor failure detection device, characterized in that for separating the analysis signal by period based on the change point of the instantaneous angle. According to claim 2,
The resampling unit,
Motor fault detection device, characterized in that for generating the resampling signal by resampling the analysis signal separated by period through a pre-arranged interpolation method into signals of equal intervals, and connecting the signals of equal intervals. According to claim 1,
The conversion unit,
Motor failure detection device, characterized in that for performing a log scale fast Fourier transform on the resampling signal. According to claim 4,
The conversion unit,
Motor failure detection device, characterized in that for generating the pre-processing signal by performing scaling on the magnitude of the signal according to the fast Fourier transform. According to claim 5,
The scaling unit is a fault detection device for a motor, characterized in that it has a value of 0 to 1. According to claim 1,
The learning unit,
Motor failure detection device, characterized in that for learning the pre-processed signal using a pre-arranged convolution-based deep learning learning algorithm. According to claim 1,
The judge,
Motor failure detection device, characterized in that for determining the shaft imbalance or winding deterioration failure of the motor using the failure characteristic data. According to claim 8,
The judge,
A failure detection device for a motor, characterized in that for analyzing the failure determination result for the motor using a pre-prepared CAM (Class Activation Map) based analysis algorithm. A resampling step of performing resampling on the phase current signal of the motor and converting it into a resampling signal;
a preprocessing step of preprocessing the resampling signal and converting the resampling signal into a preprocessing signal;
A learning step of learning failure characteristic data through the preprocessing signal; and
a failure determination step of determining whether the motor has a failure using the failure characteristic data;
including,
The resampling step is
converting the phase current signal into an analysis signal using a previously prepared Hilbert transform; and
estimating an instantaneous time from the analysis signal;
Failure detection method of the motor comprising a. According to claim 10,
The resampling step is
Separating the analysis signal by period based on the change point of the instantaneous time;
Resampling the analysis signal separated by period into a signal of the same interval through a pre-arranged interpolation method; and
generating the resampling signal by concatenating the equally spaced signals;
Failure detection method of the motor, characterized in that it further comprises. According to claim 10,
In the preprocessing step,
a fast Fourier transform step of performing log scale fast Fourier transform on the resampling signal; and
a scaling step of scaling a magnitude of a signal according to the fast Fourier transform;
Failure detection method of the motor comprising a. According to claim 10,
In the learning phase,
A method for detecting a failure of a motor, characterized in that for learning the preprocessing signal using a pre-arranged convolution-based deep learning learning algorithm. According to claim 10,
The failure determination step,
A fault detection method of a motor, characterized in that for determining an axis imbalance or a winding deterioration fault of the motor using the fault characteristic data. According to claim 10,
The failure determination step,
A method for detecting a failure of a motor, characterized in that for analyzing the failure determination result of the motor using a pre-prepared CAM (Class Activation Map) based analysis algorithm.

Images