ISA Transactions 81 (2018) 210–221

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Research article

Implementation of an embedded system for real-time detection of rotor bar T

failures in induction motors
Yilmaz Guvena, Selcuk Atisb,∗
a
Electronic and Automation Department, Vocational School of Technical Sciences, Kırklareli University, Karahıdır Mah. Harmanlık Mevkii 39100, Merkez/Kırklareli,
Turkey
b
Electrical and Energy Department, Vocational School of Technical Sciences, Marmara University, Marmara Üniversitesi Teknik Bilimler MYO, Göztepe Kampüsü, 34722
Kadıköy/İstanbul, Turkey

A R T I C LE I N FO A B S T R A C T

Keywords: In this work, the implementation of an embedded system for real-time detection of rotor bar failures in induction
Induction motor motor has been realized. The device is a prototype measurement device which can detect broken rotor bars on
Broken rotor bar the field without any additional setup or third-party software. This study has focused to derive a new method
Signal processing from previous studies on diagnosing of rotor failures and developed a microcontroller based embedded mea-
Embedded system
surement device. A new method which is combining Fast Fourier Transform and Grey Relational Analyses to get
Grey relational analysis
Failure diagnosis
best diagnosis results has been developed for an embedded system within a compact and low-cost measurement
device. Although, there are some computer-based techniques and motor drivers that can detect internal failures
of induction motor, there is no such a measurement device that can do all process by itself.

1. Introduction Detection of broken rotor bar and bearing faults can be performed by
using motor current signature analysis (MSCA). This method should be
Three-phase induction motors which are also known as asynchro- used within further analysing techniques such as Independent Compo-
nous motors are being used in many different areas because of their nent Analysis (ICA) [4], Gabor Analysis [5], Wavelet Packet Decom-
simple and maintenance-free structure. However, this simple structure position and Artificial Neural Networks (ANN) [6]. These techniques
makes them more sensitive against internal failures. These failures can require third party software, expensive data acquisition cards and
be separated into two main categories; electrical and mechanical fail- computer environment.
ures. Electrical failures can occur at stator windings, rotor or connec- The number of the broken bar can be calculated by using mathe-
tion parts. Broken rotor bars and end-rings can be formed during matical models for fault detection through the combination of current
manufacturing process because of none proper die-casting or they can and voltage signatures. Various parameters of motor such as load,
occur afterwards. In both cases, it is hard to detect these failures before torque and speed must be measured simultaneously. This technique also
they cause serious problems. Rotor failures cause both mechanical and requires complex controller board and inverters [7]. In some cases, it is
electrical problems in asynchronous motors. Besides mechanical vi- hard to determine whether there is broken bar or not because of light
bration and humming, asymmetry of rotor also affects electrical para- load or other operating condition. Different techniques like zero-setting
meters such as stator currents and coil voltages. The additional har- protection should be preferred for detection regardless of the motor
monic components appear in both start-up and steady-state currents. In status or MSCA results can be analysed within transform demodulation
other words, motor becomes low efficient and it fails eventually. algorithm in the early stage of incoming faults [8,9].
There are many different approaches for detecting broken rotor Even though there are many different methods to detect broken
bars. Dias and Chabu have used a Hall effect sensor for spectral analysis rotor bars, severity of the faults is still under investigation. Using syn-
of the magnetic flux density at very low slip in large induction motors chro-squeezing transform during start-up can improve time-frequency
[1]. Stator voltage and current can be also measured for analysing the resolution with high noise immunity [10]. Non-uniform resampling of
side band frequencies. This analysis can be done by using various motor current can separate fundamental harmonics from fault related
spectral analyses [2]. Special data acquisition cards (DAQ) can be used ones [11]. The study of comprehensive evaluation of intelligent clas-
for multiple signature processing-based fault detection techniques [3]. sifiers has shown that accuracy of fault identification methods can

∗
Corresponding author.
E-mail addresses: [email protected] (Y. Guven), [email protected] (S. Atis).

https://doi.org/10.1016/j.isatra.2018.08.004
Received 18 July 2017; Received in revised form 11 February 2018; Accepted 4 August 2018
Available online 09 August 2018
0019-0578/ © 2018 ISA. Published by Elsevier Ltd. All rights reserved.
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

Table 1 in the air gap. This magnetic field moves opposite to rotor depending on
The range of the sideband frequencies in terms of fault types. slip. This reversed magnetic field induces positive or negative se-
Fault type Frequency range quenced voltage with respect to stator frequency as in equation (1)
[16,17].
Air-gap Eccentricity 0–900 Hz
Broken Rotor Bar 0–400 Hz fs = f (1 ± 2s ) (1)
Bearing Failure 0–2000 Hz
Interturn Short Circuit 0–1000 Hz
Where s is the motor slip and f is the fundamental frequency of
electrical source, fs is the frequency of voltage which is induced by
reversed magnetic field [18]. This equation is derived from the ob-
Table 2 servation of the space harmonics as shown in equation (2). These low
Number of the calculation in 1024-point FFT. frequency harmonic components create torque and speed ripple.
Eventually, these ripples create higher harmonic components in the
FFT Algorithm Float Integer
stator current. This fault signals may also contain components caused
Adds Multiplications Adds Multiplications by static and dynamic eccentricities as in equation (3). Broken rotor
bars can be detected by tracing these components [19].
Radix-2 30720 20480 19450 2084
Radix-4 28960 15336 12902 3071 1 − s⎞
Split radix 25488 10016 12448 1024 fB = ⎡k ⎜⎛ ⎟ ± s
⎤f
⎢ ⎥
⎣ ⎝ p ⎠ ⎦ (2)

1 − s⎞
reach one hundred percent by combining different analysing methods fE = f ⎡ (kR ± nd ) ⎛⎜ ⎟ ± nk
⎤
⎢ ⎥
but it is still obvious that need for an embedded solution cannot be ⎣ ⎝ p ⎠ ⎦ (3)
ignored [12]. Besides, broken rotor bars can also be detected by using
Where f is the fundamental supply frequency, s is the motor slip, p
different techniques such as sound and vibration signals [13,14].
is the number of pole, R is the number of rotor bar, nd = 0 for static
As it can be seen, previously studies have shown that detecting and
eccentricity and nd = 1,2,3 … for dynamic, nk is the number of supply
diagnosing of these failures requires complex experimental setups and
harmonic, k is an integer k = 1,2,3 ….
additional computer software. There are some computer based devices
Harmonic components of faulty motors can be classified according
[15] and induction motor drivers which can observe many parameters
to their frequencies. Table 1 shows sideband frequencies in term of fault
of the motor for fault detection. However, these devices require com-
types.
puter and third-party software. They are far from being a measurement
We have used Motor Current Signature Analysis (MCSA) which is
device. Motor drivers are expensive and complex, it not feasible to use
sensing an electrical signal that contains current components which are
them for fault diagnosis. Therefore, we have designed an effective de-
direct product of unique rotating flux. This magnetic flux is affected by
vice that can be used in any circumstances in the field. This study has
broken rotor bars or end-rings [20,21]. The current signal in one phases
focused to derive a new method from previous studies on diagnosing of
of the healthy motor can be expressed as in equation (4). Correspond-
rotor failures and developed a microcontroller based embedded mea-
ingly, the magnetic flux in the stator is written in equation (5) and the
surement device. A new method which is combining FFT and GRA to
electrical torque produced by the interaction between the current and
get best diagnosis results has been developed for an embedded system
flux is shown in equation (6).
within a compact and low-cost measurement device.
ia = 2 I cos(2πft − α1) (4)

2. Materials & methods ϕa = 2 ϕ cos(2πft − α ϕ) (5)

T = 3PϕI sin(α1 − α ϕ) (6)

Previous studies have shown that rotor failures affect stator cur-
rent's amplitude and its wave form. Asymmetry of rotor circuit cause Where I and ϕ denote the root mean squared (RMS) value of supply
negative sequenced voltage over rotor bars and reversed magnetic field current and linkage flux, α1 and αϕ are phase of the current and flux, f is

Fig. 1. Block diagram of the experimental setup to develop the device.

211
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

Fig. 2. Synchronous generator driven by servo asynchronous motor.

Fig. 3. DTC Motor as variable load coupled with test motor.

frequency of fF = 2sf and phase angle is αF . Thus, the current of a

single phase with broken bar can be expressed as in equation (7).
if = 2 IF cos(2πfF t − αF ) (7)
Then, we have used a special Fast Fourier Transform algorithm to
transform time based current information to frequency based one. A
straightforward calculation of N point FFT requires N 2 operation which
contains complex multiplications followed by complex additions
[22,23]. FFT algorithm can be adjusted by changing signal flow or data
mapping sequence. If N = 4 k, sequences can be reorganized into N/4
arrays so it is called Radix-4. Different Radix algorithms can be used
together for precision and speed. Radix-2 and Radix-4 can be blended in
a new type of FFT algorithm which works faster and uses less space. It is
called Split-Radix algorithm. In this algorithm, unnecessary calcula-
tions have been removed by combining the various DFT outputs by
pairs of additions and subtractions.
Optimization of these FFT algorithms can be done by changing the
memory mapping. Standard FFT algorithm uses static function. This
increases accuracy and reduces calculation time. Dynamic functions
Fig. 4. Driver panel.
provide faster calculations. On the other hand, pointers can be used for
dynamic calculation instead of static variables. While Radix-2 and
Radix-4 use nearly the same amount of memory, Radix-4 needs less
the fundamental frequency of electrical supply and P is the number of calculation step as shown in Table 2. We have used an optimized Split-
pole pairs. If there is a broken rotor bar, there will be an additional Radix algorithm with dynamic variable and function declaration. This
current component due to interaction between the magnetic fields of has reduced the number of calculations and improved the speed of al-
stator and rotor. This additional current is a sinusoidal wave at gorithm.

212
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

FFT results are long data array and they are hard to interpret. For
that reason, we have used a Grey-Relation Analysis algorithm to detect
any little changes on FFT results. GRA can be used for modelling,
prediction and decision-making in the systems that contain incomplete
or unsure information. GRA provides effective solution to multivariate
problems. It can be simply described as analysis of the relation between
system's parameters and its performance. This way, unknown in-
formation can be interpreted by using these relations [24].
First step of GRA is the calculation of relation between input se-
quences. This can be done by using three different approach; higher-
better (8), smaller-better (9) and nominal-better (10). The best ap-
proach is determined by the range and order of the data sequence [25].

(Xi (k ) − min Xi (k ))
Yi (k ) =
(max Xi (k ) − min Xi (k )) (8)
Fig. 5. 3 kW Asynchronous motor.
(max Xi (k ) − Xi (k ))
Yi (k ) =
(max Xi (k ) − min Xi (k )) (9)
Table 3
Motor label values. Xi (k ) − Nv
Yi (k ) = 1 −
Parameter Value max[max Xi (k ) − Nv; Nv − min Xi (k )] (10)

Voltage 230/400 V Where Xi (k ) is the original input value, max Xi (k ) is the maximum
Current 11.6/6.7 A
value of the sequence, min Xi (k ) is the minimum value of the sequence,
Frequency 50 Hz
Cos φ 0.76
Yi (k ) is the preprocessed data, Nv is the desired normalize value and
Speed 1435 rpm k = 1,2,3 …. any integer numbers.
Efficiency %85.5 Second step of GRA is the calculation of grey relation grades for the
pre-processed data sequence. These grades determine the relational
degree between sequences (11). Finally, average value of these coeffi-
cients for n sample determines the grey relational rank (12).

Fig. 6. General diagram of the device.

213
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

Fig. 7. Final assembly of the device.

Δ min+ς Δ max calibrate the device. We have also tested the device with main power
Δi (k ) =
X0 (k ) − Xi (k ) + ς Δ max (11) grid after the calibration. The isolated environment has been created by
using an independent power source which was combination of a syn-
n
1 chronous generator driven by a servo asynchronous motor. A direct
Υ (Xi ) =
n
∑ Δi (k ) torque controlled (DTC) asynchronous motor has been used as an ad-
K =1 (12)
justable load (Fig. 1). Synchronous generator is being driven by a servo
By this way, multiple input data can be converted into a single grey asynchronous motor (Fig. 2) at constant speed for stable frequency and
relation grade that shows the correlation between different sequences. voltage output.
This method can be used for various problems such as calculating the A 4 kW asynchronous motor was used as a load because the side-
initial values of weights in an artificial neural network [26,27], band harmonics can be detected clearly when motor is fully loaded.
building power systems strategies in distribution systems [28] and DTC technique provides adjustable speed and torque for this motor.
hybrid forecasting modelling for wind power [29]. This turns a simple asynchronous motor into perfect dynamic load
We have blended three different method into one to create effective within regenerative working mode (Fig. 3). Driver panel combines a
new method. This method has enough precision to detect rotor failures servo synchronous motor driver and a DTC driver (Fig. 4).
while it can be run over an embedded system. This allowed us to build
an effective measurement device which can be used by anyone in every 3 kW asynchronous motors have been used for experimental studies
situation without complex set-up and theoretical knowledge. We had (Fig. 5). The label values of the motor are shown on Table 3. This
chance to test the device in a factory while motor was operational motor has 32 slots on its stator and 24 slots on its rotor with 75%
without disconnecting it from power grid. This measurement device has rotor-stator slots ratio.
been designed to be used in field without disconnecting motor from
power grid and its load. Previous studies have focus on light loaded or We have used a closed loop Hall Effect current transducer for the
transient regime but these motors works with full loaded in steady re- measurement. The current signal has all standard harmonics and some
gime in the field. This device is designed for field use apart from interference signals. An anti-aliasing filter has been used to eliminate
complex experimental set-up in the laboratory. Previous studies show those interference signals and spikes. The corner frequencies of ana-
that sideband harmonics are too low to detect with no load. Because of logue filter have been adjusted by using PWM signal from micro-
this, we design this device to be used for loaded motors in the field. controller. This provide an adjustable hybrid filter for precision. We
On the other hand, using of a GRA algorithm to interpret FFT result have preferred to use a microcontroller (AT91SAM3X8E) with ARM
is a new method. GRA has found many application areas in recent years. Cortex-M3 microprocessor. ARM Cortex M series are high performance
FFT results are mostly being handled by a third-party application such 32-bit microprocessors. Although, they are not considered for DSP, they
as MATLAB and LabView. This software is expensive and requires ad- have special multiplication command set. This way, they can achieve
ditional computer power. Therefore, we have developed an effective 32 and 64-bit calculation easily. The general block diagram of the de-
diagnosis algorithm within a low-cost embedded system. vice can be seen in Fig. 6 and final assembly of the device in Fig. 7.
This device uses a hybrid filter as anti-aliasing filter. This filter has
3. Experimental setup and device an adjustable analogue filter along with a band bass digital filter. The
corner frequency of analogue filter is depending on stop band fre-
At the beginning, we have needed an isolated environment for de- quencies of the digital filter. This method provides precision on low
velopment of the device. We didn't want any interference before we frequencies. On the other hand, FFT results are being handled by a GRA

214
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

Fig. 8. Fault Diagnosis Algorithm flowchart.

algorithm so any little change on signal can be detected instantly. The high resolution to determine side band frequencies clearly. For this
real novelty of this device that this embedded system can do all the reason, sampling frequency of ADC is set to 500 Hz while FFT takes
analysis by itself. 1024 samples point. This provides 0.488 Hz resolution for each FFT
The fault signal can be detected by focusing on the right frequency point. ARM Cortex M3 core can run this algorithm in 4283 μsec with its
range. We wanted to design a real time measurement device so we 84 MHz clock using 24712 bytes in memory. We have tried to change
reduced the resolution by purpose since the device measures AC grid so the resolution by changing sampling time. We have managed to lower it
high resolution is really not required. However, the FFT analysis needs to 0,1 Hz but it affects the speed of the measurement negatively. We

215
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

Table 4 These results are not reliable because they change in every second.
FFT result. They are not linear either. For that reason, we have used a grey-relation
n Load K1 K2 K3 K4 K5 K6 K7 algorithm to interpret these results. Actually, the device runs this al-
gorithm with a much longer sequence. Therefore, the actual process
(2s-1) Lower Base (2s + 1) Upper Avg. Broken calculates with more accuracy.
Table 4 shows the FFT result of three different motors with one
1 0 49.8 8.6 363 50.78 1.4 2.35 0
2 0.25 49.8 8.06 366 50.78 1.4 2.35 0 broken, two broken and three broken bars at four different loads. K1
3 0.5 49.8 8.94 384 50.78 2.2 2.39 0 represents lower sideband frequencies and K2 represents its amplitudes;
4 1 49.8 10.7 464 50.78 2.2 2.56 0 K3 represents amplitudes of main motor current; K4 represents upper
5 0 49.8 8.4 363 50.78 1 2.35 1 sideband frequencies and K5 represents its amplitudes; K6 shows
6 0.25 49.06 1 368 51.26 2 2.35 1
average value of FFT samples; K7 indicates the number of the broken
7 0.5 48.3 3 388 52.01 2.2 2.4 1
8 1 46.56 2 465 53.46 3 2.51 1 rotor bar. Repetitive values such as load percent have been ignored
9 0 49.8 8.06 359 50.78 1 2.36 2 from GRA calculations because they have no effect on GRA results.
10 0.25 49.07 2.82 363 51.26 3 2.36 2 FFT results of fault motors have been compared with the healthy
11 0.5 48.33 4 387 52.01 4.1 2.41 2
motor to obtain their correlation for GRA. This result can be obtained
12 1 46.87 2.23 466 53.46 2.23 2.59 2
13 0 49.8 8.6 359 50.78 1 2.36 3 by using any three-method stated below. The higher-better is more
14 0.25 49.07 2.82 363 51.26 3 2.36 3 suitable for our sequence since we are looking for the difference. This
15 0.5 48.33 4 387 52.01 4.12 2.43 3 method should be preferred when there is a little difference on input
16 1 46.63 4.47 476 53.95 6 2.63 3 parameters. Since we have preferred the higher-better approach, 1 re-
presents the healthy state of rotor and 0 represents faulty rotor.
Variation of difference indicates the level of the damage. Table 5 shows
Table 5
the pre-process of grey relational grades. The grades can be ranked
The difference of the healthy and faulty state measurement.
according the fault type. Each parameter affects the grade directly so
Δ1 Δ2 Δ3 Δ4 Δ5 Δ6 Δ7 one grey relational result emerges from multiple measurement value.
Table 6 shows the grey relational grades according to each coefficient.
0 0.2 0 0 0.4 0 1
0.74 7.06 2 0.48 0.6 0 1 GRA result of the measurements can be seen on Fig. 9. Grades are
1.5 5.94 4 1.23 0 0.01 1 varying between 1 and 0 for motors with broken bar(s). We have
3.24 8.7 1 2.68 0.8 0.05 1 ranked them into five different groups that indicate the number of the
0 0.54 4 0 0.4 0.01 2 broken bar. The range and amount of the change in grades define
0.73 5.24 3 0.48 1.6 0.01 2
1.47 4.94 3 1.23 1.9 0.02 2
whether rotor is healthy or not. This process repeats itself in the algo-
2.93 8.47 2 2.68 0.03 0.03 2 rithm and traces the changes over the FFT results. The results look
0 0 4 0 0.4 0.01 3 uncertain about the severity of the damage however they can be in-
0.73 5.24 3 0.48 1.6 0.01 3 terpreted by comparing results of different load level. Furthermore, an
1.47 4.94 3 1.23 1.92 0.04 3
ANN can be trained with these data to improve accuracy.
3.17 6.23 12 3.17 3.8 0.07 3
Computer based diagnosis techniques requires data acquisition.
These techniques also require an additional 3rd party computer soft-
have calibrated the device to get best results for its aim and scope. ware. The collected data must be transferred to a computer before
Device software consists of four parts; ADC sampler, FFT algorithm, analyses. All these processes take long time that cannot be consider as
Grey-Relational Analysis and graphic LCD interface. All data can be real-time diagnosis. In our system, measurement of the signal, Fast
logged while operation. This data can be send to a PC via USB for Fourier Transform and Grey-Relational Analysis take only 4 ms. Our
further analysis. Graphic LCD shows the signal spectrum of the stator method does not require an additional hardware such as DAQ or PC
current. Amplitudes and frequencies of each spectrum elements can computer so there is no delay caused by operation system and third-
also be monitored. Final result of GRA is shown after the analysis. party software. Comparison of previous studies can be seen on Table 7.
Flowchart of the fault diagnosis algorithm can be seen in Fig. 8.
GRA algorithm handles frequencies and amplitudes of each FFT 4. Results and discussion
point to evaluate the condition of rotor bars. We have demonstrated a
brief part of the measurements to show how GRA works on FFT results. The signal of the healthy motor (Fig. 10(a)) contains all standard
harmonics such as 3rd, 5th and 7th. Fig. 10 (b) shows the signal after all

Table 6
Grey Relation Coefficients (ξ) and grades (Ʃƴ).
ξ1 ξ2 ξ3 ξ4 ξ5 ξ6 ξ7 Ʃƴ

1 0.977528 1 1 0.904762 1 1 0.903184

0.81407 0.55203 0.85714 0.868493 0.863636 1 1 0.850768
0.683544 0.594262 0.75 0.720455 1 0.875 1 0.803323
0,5 0.5 0.92308 0.54188 0.826087 0.583333 1 0.69634
1 0.941558 0.75 1 0.904762 0.875 0.8 0.895903
0.816121 0.624103 0.8 0.868493 0.703704 0.875 0.8 0.783917
0.687898 0.63783 0.8 0.720455 0.666667 0.777778 0.8 0.727232
0.525122 0.506698 0.85714 0.54188 0.992167 0.7 0.8 0.653287
1 1 0.75 1 0.904762 0.875 0.666667 0.885204
0.816121 0.624103 0.8 0.868493 0.703704 0.875 0.666667 0.76487
0.687898 0.63783 0.8 0.720455 0.664336 0.636364 0.666667 0.68765
0.50546 0.582719 0.5 0.5 0.5 0.5 0.666667 0.536407

216
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

Fig. 9. GRA grades and ranking.

Table 7
Comparison of previous methods and devices.
Method PC Software Hardware Device Real Time

Spectral Analysis [1] Required MATLAB Tektronix TDS2014 Digital Oscilloscope Experimental No
OriginPro Setup
Gabor Analysis [5] Required MATLAB Yokogawa DL750 Experimental No
Digital Oscilloscope Setup
Synchro-squeezing transform [10] Required MATLAB ADS7841E Experimental No
Data Acquisition Card Setup
Non-Uniform Time Resampling [11] Required MATLAB ADS8341 Experimental No
Data Acquisition Card Setup
Self-Powered Instrument [15] Required PC interface PIC16F877A Wireless No
Device
Our Device No Embedded ARM Cortex M3 Embedded Yes
Code Device

Fig. 10. Healthy motor stator current signal. (Horizontal: 8 ms/div Vertical: Fig. 11. FFT result of healthy motor current.
2 V/div).
The LCD displays the spectrum of the stator current and value of the
these harmonics have been eliminated. When healthy motor's stator side band frequencies if they are available. Here, display shows only the
current has been analysed after harmonic filtering, FFT result shows line frequency which is 50 Hz.
that there is no side band frequency at all (Fig. 11). This is accepted. When there is a broken bar on the rotor, stator current signal starts
The test result can be displayed through device graphic LCD (Fig. 12). to decay because of the side band frequencies. This distortion can be

217
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

Fig. 12. Test Device screens for healthy

motor.

inertia of the rotor and type of the failures such as number of the broken
bars.
The side band frequencies are varying between 45 and 50 Hz on the
lower band; 50–55 Hz on the upper band. Amplitudes of these signals
are very low relative to main stator current so FFT parameters and
spectrum interval should be optimized sensitively. Measurement and
calibration tests of the device have shown that the number of the
broken bar do not affect the sideband frequency but affects its ampli-
tude.
The final experiments have been conducted by using three motors
with simulated broken bars as seen in Fig. 15. Measurements have been
started with no load and ended when motors were fully loaded. The rate
of changes has been analysed by diagnosis algorithm. Grey Relational
Grades (GRG) depend on motor load percent and number of the broken
bar can be seen on Fig. 16.
GRA result shows that the changes on motor parameters are strictly
Fig. 13. Faulty motor stator signal. (Horizontal: 4 ms/div Vertical: 800 mV/
depend on rotor condition. It has been seen that these changes are
div).
getting bigger based on number of the broken rotor bars and percent of
load. Our device can be used for monitoring and tracing of these
noticed at the peaks of the sinusoidal signal (Fig. 13). changes related to rotor bar failures. We have scaled the broken bar
Since we measure the frequencies of the main and sideband har- between one and three but this scale can be extended for different
monic, slip of the motor can be calculated with fs = f (1 ± 2s ) . We have motors and fault types. The device shows the result of one broken bar
also monitored the motor speed via driver to check whether measure- (Fig. 17 (a)), two broken bars (Fig. 17(b)) and three broken bars
ment result of the device is correct or not. Broken rotor bars cause (Fig. 17(c)).
symmetric side band frequencies in stator current (Fig. 14). These side We have used an isolated environment for development of the de-
band frequencies are related to slip and they move away from main line vice. Later on, we have conducted our measurements on main A.C. grid.
frequency proportional to load and inertia. The sideband components Side band harmonics are changing related to failure type. Tables 8–10
may overlap the main frequency when the motor is unloaded or lightly shows different type of rotor failures. We have simulated the broken bar
loaded. The motor should be fully loaded to get best results. On the by cutting rotor bars. Each failure creates a unique signal signature so
other hand, amplitudes of these sideband frequencies are related to the we measure different harmonics value for each time.

Fig. 14. FFT results of faulty motor with different loads.

218
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

Fig. 15. Simulated broken rotor bars.

Fig. 16. Grey relational grades of faulty motors.

Cosφ : Power factor

• We have measured different type of fault at least ten times at dif-
ferent temperatures

nr : Rotor speed • Motors have been powered by synchronous generator and AC grid
separately
s : Rotor slip value
• Results have been compared with healthy motors for each case
THDI : Total harmonic distortion
• Three different brand of induction motor company have been tested
IA
(current)
: Motor phase current • Two different researchers have been performed all measurements
individually
f : Supply frequency
f(2s-1) : Lower band component
The results of all these measurements were identical. Deflection
f(2s + 1) : Upper band component
between each measurement was less than %2 and we have performed
IF : Sideband component amplitude
many different measurements to make sure the diagnosis algorithm
GRG : Grey relational grade
works perfect.

We have also tested reliability of the device by repeating experi- 5. Conclusions

ments with different motors and different condition. The work range of
this device is limited by current range of motor. Device can measure up In this work, a fault diagnosis device to detect broken rotor bars in
to 25A, however a current transformer can be used to improve this induction motors has been realized. All previous theoretical and ex-
range. We could not mention all of our experiments here but we have perimental studies have been reviewed to find a reliable new method. A
done the following test to ensure the reliability of the device: microcontroller based embedded system has been developed for this

219
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

Fig. 17. Faulty motor test results (Device screen captures).

Table 8
Measurements results of the motor with one broken bar.
Load Cosφ nr (rpm) s THDI(%) IA(A) f(Hz) f(2s-1) f(2s + 1) IF(mA) GRG

No 0.1 1495 0.0033 5 5.12 50.2 – – – 0,903184

¼ 0.31 1485 0.01 4.5 5.25 50.2 49.12 51.26 8 0,850768
½ 0.45 1474 0.0173 4 5.6 50.2 48.33 52.1 11 0,803323
Full 0.7 1447 0.0353 3 6.7 50.2 46.85 53.46 15 0,69634

Table 9
Measurements results of the motor with two broken bars.
Load Cosφ nr (rpm) s THDI(%) IA(A) f(Hz) f(2s-1) f(2s + 1) IF(mA) GRG

No 0.1 1495 0.0033 5 5.15 50.2 – – – 0,895903

¼ 0.31 1485 0.01 4.5 5.25 50.2 49.23 51.33 11 0,783917
½ 0.45 1474 0.0173 4 5.7 50.2 48.35 52.27 16 0,727232
Full 0.7 1446 0.036 3.1 6.71 50.2 46.86 53.63 20 0,653287

Table 10
Test Measurements results of the motor with three broken bars.
Load Cosφ nr (rpm) s THDI(%) IA(A) f(Hz) f(2s-1) f(2s + 1) IF(mA) GRG

No 0.1 1495 0.0033 5.1 5.2 50.2 – – – 0,885204

¼ 0.32 1485 0.01 4.5 5.25 50.2 49.08 51.26 17 0,76487
½ 0.46 1474 0.0173 4.2 5.6 50.2 48.33 52.01 21 0,68765
Full 0.71 1445 0.036 3.2 6.72 50.2 46.63 53.67 25 0,536407

purpose. This device is inexpensive and it contains few parts as possible. faults. GRA is a perfect method to calculate initial weights for ANN.
It is also important that although there are many studies about broken Therefore, future works will be interested in building a database to
rotor bars and end-rings, there is no such an embedded portable device train an ANN to improve the accuracy of the method.
that can do all analyses by itself. This device is a prototype measure-
ment device. It provides a quick and easy method for detection of
broken rotor bars in induction motors. Rotor failures can be detected in Acknowledgements
the field without disconnecting or disassembling the motor.
Future work will be focused on improvement of fault diagnosis al- This research has been supported by Marmara University,
gorithm by using an Artificial Neural Network (ANN). This ANN will be Directorate of Scientific Research Projects. Project Number: FEN-C-
train by using measurement data from different motors with different DRP-121214-0392.

220
Y. Guven, S. Atis ISA Transactions 81 (2018) 210–221

References 2017;83:568–89. https://doi.org/10.1016/j.ymssp.2016.06.032.

[15] Cabanas MF, Pedrayes F, Rojas CH, Melero MG, Norniella JG, Orcajo GA, et al.
A new portable, self-powered, and wireless instrument for the early detection of
[1] Dias CG, Chabu IE. Spectral analysis using a Hall effect sensor for diagnosing broken broken rotor bars in induction motors. IEEE Trans Ind Electron 2011;58:4917–30.
bars in large induction motors. IEEE Trans Instrum Meas 2014;63:2890–902. https://doi.org/10.1109/Tie.2011.2109340.
https://doi.org/10.1109/Tim.2014.2328184. [16] Benbouzid ME. A review of induction motors signature analysis as a medium for
[2] Didier G, Ternisien E, Caspary O, Razik H. Fault detection of broken rotor bars in faults detection. IEEE Trans Ind Electron 2000;47:984–93. https://doi.org/10.
induction motor using a global fault index. IEEE Trans Ind Appl 2006;42:79–88. 1109/41.873206.
https://doi.org/10.1109/Tia.2005.861368. [17] Mehrjou MR, Mariun N, Marhaban MH, Misron N. Rotor fault condition monitoring
[3] Ayhan B, Chow MY, Song MH. Multiple signature processing-based fault detection techniques for squirrel-cage induction machine-A review. Mech Syst Signal Process
schemes for broken rotor bar in induction motors. IEEE Trans Energy Convers 2011;25:2827–48. https://doi.org/10.1016/j.ymssp.2011.05.007.
2005;20:336–43. https://doi.org/10.1109/Tec.2004.842393. [18] Kentli F. Effects of broken rotor bars in asynchronous motors. Istanbul Technical
[4] Yang T, Pen HB, Wang ZX, Chang CS. Feature knowledge based fault detection of University; 1992.
induction motors through the analysis of stator current data. IEEE Trans Instrum [19] Naha A, Samanta AK, Routray A, Deb AK. A method for detecting half-broken rotor
Meas 2016;65:549–58. https://doi.org/10.1109/Tim.2015.2498978. bar in lightly loaded induction motors using current. IEEE Trans Instrum Meas
[5] Riera-Guasp M, Pineda-Sanchez M, Perez-Cruz J, Puche-Panadero R, Roger-Folch J, 2016;65:1614–25. https://doi.org/10.1109/Tim.2016.2540941.
Antonino-Daviu JA. Diagnosis of induction motor faults via gabor analysis of the [20] Thomson WT, Gilmore RJ. Motor current signature analysis to detect faults in in-
current in transient regime. IEEE Trans Instrum Meas 2012;61:1583–96. https:// ductionmotor drives-fundamentals, data interpretation, and industrial case his-
doi.org/10.1109/Tim.2012.2186650. tories. 32nd turbomach symp. 2003. p. 145–56.
[6] Sadeghian A, Ye ZM, Wu B. Online detection of broken rotor bars in induction [21] Gu F, Wang T, Alwodai A, Tian X, Shao Y, Ball AD. A new method of accurate
motors by wavelet Packet decomposition and artificial neural networks. IEEE Trans broken rotor bar diagnosis based on modulation signal bispectrum analysis of motor
Instrum Meas 2009;58:2253–63. https://doi.org/10.1109/Tim.2009.2013743. current signals. Mech Syst Signal Process 2015;50(51):400–13. https://doi.org/10.
[7] Hou ZW, Huang J, Liu H, Wang T, Zhao LH. Quantitative broken rotor bar fault 1016/j.ymssp.2014.05.017.
detection for closed-loop controlled induction motors. IET Electr Power Appl [22] Chen Y, Lin YW, Lee CY. A block scaling FFT/IFFT processor for WiMAX applica-
2016;10:403–10. https://doi.org/10.1049/iet-epa.2015.0440. tions. 2006 IEEE Asian solid-state circuits conf. 2006. p. 203–6. https://doi.org/10.
[8] Pezzani C, Donolo P, Bossio G, Donolo M, Guzman A, Zocholl SE. Detecting broken 1109/ASSCC.2006.357886.
rotor bars with zero-setting protection. IEEE Trans Ind Appl 2014;50:1373–84. [23] Scarpa G. Digital signal processing course notes. 2017. n.d.
https://doi.org/10.1109/Tia.2013.2276116. [24] Chang YT, Chuang TF. A study of using grey system theory and artificial neural
[9] Hu NQ, Xia LR, Gu FS, Qin GJ. A novel transform demodulation algorithm for motor network on the climbing ability of Buergeria Robusta frog. Open J Ecol
incipient fault detection. IEEE Trans Instrum Meas 2011;60:480–7. https://doi.org/ 2013;03(No:11). https://doi.org/10.4236/oje.2013.32010.
10.1109/Tim.2010.2050980. [25] Hasani H, Tabatabaei SA, Amiri G. Grey relational analysis to determine the op-
[10] Camarena-Martinez D, Perez-Ramirez CA, Valtierra-Rodriguez M, Amezquita- timum process parameters for open-end spinning yarns. J Eng Fibers Fabr
Sanchez JP, Romero-Troncoso RD. Synchrosqueezing transform-based methodology 2012;7:81–6.
for broken rotor bars detection in induction motors. Measurement 2016;90:519–25. [26] Lin YC, Yeh CH. Grey relational analysis based artificial neural networks for product
https://doi.org/10.1016/j.measurement.2016.05.010. design: a comparative study. Proc 12th int conf informatics control autom robot,
[11] Garcia-Calva TA, Morinigo-Sotelo D, Romero-Troncoso RD. Non-uniform time re- vol. 1. 2015. p. 653–8. https://doi.org/10.5220/0005577506530658.
sampling for diagnosing broken rotor bars in inverter-fed induction motors. IEEE [27] Chen JH, Chang TT, Ho CR, Diaz JF. Grey relational analysis and neural network
Trans Ind Electron 2017;64:2306–15. https://doi.org/10.1109/Tie.2016.2619318. forecasting of REIT returns. Quant Finance 2014;14:2033–44. https://doi.org/10.
[12] Palacios RHC, da Silva IN, Goedtel A, Godoy WF. A comprehensive evaluation of 1080/14697688.2013.816765.
intelligent classifiers for fault identification in three-phase induction motors. Elec [28] Chen WH. Quantitative decision-making model for distribution system restoration.
Power Syst Res 2015;127:249–58. https://doi.org/10.1016/j.epsr.2015.06.008. IEEE Trans Power Syst 2010;25:313–21. https://doi.org/10.1109/TPWRS.2009.
[13] Camarena-Martinez D, Valtierra-Rodriguez M, Amezquita-Sanchez JP, Granados- 2036811.
Lieberman D, Romero-Troncoso RJ, Garcia-Perez A. Shannon entropy and K-Means [29] Shi J, Ding Z, Lee WJ, Yang Y, Liu Y, Zhang M. Hybrid forecasting model for very-
method for automatic diagnosis of broken rotor bars in induction motors using short term wind power forecasting based on grey relational analysis and wind speed
vibration signals. Shock Vib 2016;2016:4860309. distribution features. IEEE Trans Smart Grid 2014;5:521–6. https://doi.org/10.
[14] Delgado-Arredondo PA, Morinigo-Sotelo D, Osornio-Rios RA, Avina-Cervantes JG, 1109/TSG.2013.2283269.
Rostro-Gonzalez H, Romero-Troncoso RD. Methodology for fault detection in in-
duction motors via sound and vibration signals. Mech Syst Signal Process

221