Journal of Control, Automation and Electrical Systems (2018) 29:489–499

https://doi.org/10.1007/s40313-018-0388-5

A Neuro-Fuzzy Approach for Locating Broken Rotor Bars in Induction

Motors at Very Low Slip
Cleber Gustavo Dias1 · Cristiano Morais de Sousa1

Received: 22 October 2017 / Revised: 17 March 2018 / Accepted: 17 May 2018 / Published online: 30 May 2018
© Brazilian Society for Automatics–SBA 2018

Abstract
Squirrel-cage induction motors are widely used in a number of applications throughout the world. This paper proposes a
neuro-fuzzy approach to identify and to classify a typical fault related to the induction motor damage, such as broken rotor
bars. Two fuzzy classifiers are obtained by an adaptive-network-based fuzzy inference system model whose parameters can be
identified by using the hybrid learning algorithm. A Hall effect sensor was installed between two stator slots of the induction
machine, and a magnetic flux density variation is measured according to the failure. The data from the Hall sensor were
used to extract some harmonic components by applying fast Fourier transform. Thus, some frequencies and their amplitudes
were considered as inputs for the proposed fuzzy model to detect not only adjacent broken bars, but also noncontiguous
faulted scenarios. In the present work it is not necessary to estimate the rotor slip, as required by the traditional condition
monitoring technique, known as motor current signature analysis. This method was able to detect broken bars for induction
motor running at low-load or no-load condition. The intelligent approach was validated using some experimental data from
a 7.5-kW squirrel-cage induction machine.

Keywords Broken rotor bars · Neuro-fuzzy · Induction motor · Fault diagnosis · Low slip · Hall sensor

1 Introduction et al. (2011), Chandralekha and Jayanthi (2016), Sauer

et al. (2015), Palacios et al. (2016), Naha et al. (2016) and
Squirrel-cage induction motors (SCIMs) are currently Lizarraga-Morales et al. (2017) some techniques are pre-
responsible for several tasks, including pumps, fans and com- sented to detect faults in electrical machines, such as short
pressors, for example. Induction motors are also applied to circuit in the stator coils, bearing faults and broken rotor bars.
many fields, such as petrochemical, pulp and paper indus- For this reason, motor fault diagnosis is usually applied
tries as shown by Zhang et al. (2011), as well as applications to predictive maintenance programs in the industry. In recent
in electrical vehicles, as described by Guan et al. (2014). years, condition monitoring solutions have been investigated
According to Sauer et al. (2015), induction motors are to identify some electrical and mechanical faults in induction
responsible for 35% of the electrical energy consumption motors.
in Brazil. Other countries, such as Canada, the USA, India Broken rotor bars are a typical failure that may occur
and the European Union, have more than 70% of energy con- during motor operation in both transient and steady-state con-
sumption related to induction motor usage. ditions. Many researches have invested time and resources to
Although SCIMs are robust machines due to their reli- detect and to classify broken bars, in order to avoid downtime
able and simple construction, in many cases these motors and high maintenance costs. Some recent studies have shown
are subjected to relevant mechanical, thermal and electri- a concern about detecting incipient rotor faults, such as the
cal strengths. In the works by Jung et al. (2006), Zhang works by Chandralekha and Jayanthi (2016), Sauer et al.
(2015), Palacios et al. (2016), Naha et al. (2016) , Lizarraga-
Morales et al. (2017) and el Malek et al. (2017).
B Cleber Gustavo Dias
According to Lee et al. (2015), motor current signature
[email protected]
analysis (MCSA) is currently applied for broken bars detec-
1 Informatics and Knowledge Management Graduate Program, tion and this technique has been used over the last decades,
Nove de Julho University - UNINOVE, Rua Vergueiro, particularly due to its noninvasive characteristics and attrac-
n.235/249, Liberdade, São Paulo, SP, Brazil

123
490 Journal of Control, Automation and Electrical Systems (2018) 29:489–499

tive applications in the industrial environment. On the other

hand, MCSA has some drawbacks related to rotor failures
diagnosis, such as detection at very low slip and nonadja-
cent broken bars, as cited by Sizov et al. (2009), Faiz and
Ebrahimi (2009), Puche-Panadero et al. (2009), Riera-Guasp
et al. (2010), Laala et al. (2011), Lee et al. (2015) and Naha
et al. (2016).
The articles by Riera-Guasp et al. (2010) and Sizov et al.
(2009), for example, described the limitations of MCSA
when two nonadjacent broken bars are separated by a half- Fig. 1 Rotor structure and the overhang portion of the rotor bars
pole pitch (90◦ electrical). The work by Lee et al. (2015)
highlights the typical root causes of false-positive and false-
negative indications for rotor fault detection using MCSA.
Other studies have investigated the use of pattern recog-
nition and artificial intelligence techniques to detect not only
broken rotor bars, but also other types of failures, such as
Sadeghian et al. (2009), Singh et al. (2013), Zarei (2012),
Laala et al. (2011), Palacios et al. (2016) and Godoy et al.
(2016). In Sadeghian et al. (2009), for instance, an algorithm
for broken bars detections was proposed using wavelet packet
decomposition (WPD) and neural networks. This work dis-
cusses the use of the algorithm in large induction motors,
especially in reduced load conditions. In this sense, and as
cited by Al-Badri et al. (2015) and Lu et al. (2008), today, Fig. 2 Motor used for tests and Hall sensor installed near rotor bars
the motors below 500 hp operate at approximately 60% of
their rated load, thus at low-slip conditions and hence with
reduced efficiency, which results in wasted energy and addi- presents the neuro-fuzzy model to identify and to classify
tional cost. the rotor fault, and Sect. 4 shows the experimental results. In
It should be noted that large induction motors, such as Sect. 5, the conclusions are drawn.
medium-voltage SCIMs, with hundreds or thousands of kilo-
watts, usually work under very low slip, even during rated
load, as shown by Thomson and Fenger (2003), Zhang et al. 2 A Hall Effect Sensor Installed Inside the
(2011) and Lee et al. (2015). Motor
The works by Palacios et al. (2016) and Godoy et al.
(2016), for example, have disclosed different pattern recog- Here, a Hall sensor is installed between two stator slots of the
nition methods to detect and to classify broken rotor bars, but machine. As described by Dias and Chabu (2014) and Dias
none of them have addressed this failure under low-slip con- and de Sousa (2017), this kind of sensor is able to detect
ditions or breakage bars located at distinct magnetic poles. the air gap flux variation produced by rotor currents and also
Based on the aforementioned state of the art, this work disturbances caused by broken bars.
presents the use of a neuro-fuzzy model to detect broken rotor The magnetic flux density was measured by the Hall sen-
bars for both adjacent and nonadjacent cases. This method sor, and this unit is installed to monitoring the flux produced
takes into account the frequency sidebands provided by a by the overhang portion of the rotor bars. Figure 1 shows a
Hall effect sensor, installed between two stator slots of the rotor of an induction motor and its structure used in the exper-
machine, as better described in the next section. The fre- imental tests. Figure 2 shows a Hall sensor installed near the
quencies and their amplitudes have been obtained using fast rotor bars and the stator winding extension.
Fourier transform (FFT). In addition, this paper discusses
the neuro-fuzzy application for broken bars detection when 2.1 A Frequency Domain Data Evaluation
induction motors are operating at reduced load or low slip.
This paper is divided into five sections: Sect. 1 presents a The fast Fourier transform (FFT) has been applied to evalu-
brief introduction about the current use of induction motors ate the resultant magnetic flux density in a frequency domain,
round the world and the previous works related to broken bars from the Hall sensor and particularly the first harmonic com-
detection. Section 2 discloses the use of a Hall sensor inside ponents according to the air gap flux variation. As cited by
the machine to extract frequency domain features. Section 3 Dias and Chabu (2014) and Dias and de Sousa (2017), the

123
Journal of Control, Automation and Electrical Systems (2018) 29:489–499 491

Table 1 Comparison between first and third harmonic amplitudes for

motor running at slip = 1% . (Reproduced with permission from Dias
and de Sousa 2017)
Rotor condition First harm. (dB) Third harm. (dB)

Healthy − 96 − 84.2
One broken bar BB − 85 − 81.2
Two adjacent BB − 81.1 − 75
Three adjacent BB − 76.4 − 72.2
Four adjacent BB − 75 − 69.7

frequency domain analysis is able to detect broken bars by

observing the amplitudes of the first and the third harmonic
components.
On the other hand, although third harmonic amplitudes
change with broken bars, this work carried out the use of
only first harmonic components to detect or locate the rotor
failure, since the amplitudes of their frequencies have pre-
sented a greater variation when compared to amplitudes of
the third harmonics (Table 1). A first work published by one
of the authors Dias and de Sousa (2017) showed the medium-
amplitude disturbances of the first and third harmonics for a
damaged rotor with broken bars.
Equation 1 shows the frequencies in the air gap flux, after
applying FFT:

 
k Fig. 3 Spectral analysis of the magnetic flux density (Hall sensor) with
fk = f0 (1 − s) ± s , (1)
p two broken bars

where f 0 is the stator or line frequency, k the harmonic index

This hybrid topology is an extension of the Takagi–Sugeno
(1,2,3,…), p the number of pole pairs, and s the rotor slip.
(TSK) model and allows the fuzzy system to learn the
In this case, there are two frequencies for each harmonic
parameters by using the adaptive backpropagation learning
and load condition. Figure 3 shows a rotor fault scenario,
algorithm. If we assume that the inference system has two
wherein the FFT was applied to the magnetic flux density
inputs, for example, x and y, and one output z, the rule base
obtained from the Hall sensor, for a motor with two adjacent
is determined as:
broken bars. In this operational condition the motor was run-
Rule 1: if x is A1 and y is B1 then f 1 = p1 x + q1 y + r1
ning at no load. This figure shows the harmonic components
Rule 2: if x is A2 and y is B2 then f 2 = p2 x + q2 y + r2 ,
used as inputs for the fuzzy classifier.
where p, q and r are parameters of function f i .
In this example, A1, B1, A2 and B2 are linguistic terms;
thus, each input has two terms, as shown in Fig. 4. The output
3 A Neuro-Fuzzy Approach to Detect Broken of Layer 2 is the product of all inputs. In Fig. 4 only two
Bars rules are shown. Jang (1993) states that Layer 3 is used to
normalize the input firing strengths and every node in Layer
3.1 The ANFIS Structure 4 is a parameterized function, as shown in Rules 1 and 2.
Layer 5 is a simple summing function that computes the
In this paper, an adaptive-network-based fuzzy inference sys- overall output as the summation of all incoming signals, as
tem, also called ANFIS architecture, has been proposed to defined by Eq. 2:
detect and to classify broken bars in induction motors. The

ANFIS is basically a combination of fuzzy logic and neural  (wi f i )
networks, although the ANFIS is much more complicated ∅5,1 = (wi f i ) = i . (2)
than fuzzy inference systems as cited by Jang (1993). i i (w i )

123
492 Journal of Control, Automation and Electrical Systems (2018) 29:489–499

Fig. 6 Membership functions used for fuzzy inputs

Fig. 4 A typical two-input ANFIS network

Fig. 7 Broken rotor bars located at two poles

As mentioned before, this ANFIS approach is capable

of learning from data and a backpropagation algorithm was
applied to the learning step. A total of 100 patterns were
stored. In this case, 14 patterns were considered for healthy
motor, including different load torque conditions and the
other samples (86 patterns) were used as faulted rotor sce-
Fig. 5 Fuzzy classifier with six inputs
narios. A total of 90 patterns were used for the learning step
and 10 patterns for the validation process. As can be seen
Therefore, the ANFIS topology is a multilayer feed- in Fig. 6, each input was modeled with three membership
forward network able to search for fuzzy decision rules functions.
efficiently and to construct membership functions (MFs). Figure 7 shows a scenario with four broken bars, but
After the learning process, the ANFIS model implements located at two magnetic poles. In this case, the fuzzy classifier
a fuzzy inference system (FIS). In this paper, two ANFIS should indicate a value equal to two poles.
models were used for experimental tests, including four and The rule editor window was used for the learning process,
six inputs, as better described in the next section. using the ANFIS MATLAB toolbox. For six inputs fuzzy
model, the generated FIS structure contains 729 fuzzy rules.
3.2 Learning of the ANFIS and FIS Model The motor used for experimental tests has 38 rotor bars and 4
magnetic poles; thus, this machine has 9 21 bars per pole. The
A data set was configured as an input/output relationship, rotor bars were numbered from 1 to 38. The ANFIS model
and the toolbox ANFIS, available in the MATLAB software, should indicate a value equal to 4, for example, in a motor
has been used to build membership functions and the two operational condition with broken bars numbered 2, 12, 22
fuzzy classifiers. For each motor operational scenario the and 32. Figure 8 shows the complete methodology used in
higher amplitude was considered in dB, and its respective this work.
frequency, in Hz, from the first harmonic as fuzzy inputs. The final FIS model should indicate value equal to zero
Figure 3 shows FFT of the Hall sensor signal and the highest for a rotor without failure. Table 2 shows some fault rotor
three amplitudes. Thus, there are six values in this case, since patterns used for the ANFIS learning step.
we have three amplitudes and their frequencies. Figure 5
shows the fuzzy classifier with six inputs.
The output of the ANFIS system defines the number of 4 Detection of Broken Bars Located in
magnetic poles with broken bars, or a healthy condition. The Different Positions
induction motor used for the experimental tests has four mag-
netic poles; thus, the ANFIS output value varies between 0 4.1 Experimental Setup
and 4. Therefore, it should be noted that the present approach
does not identify the number of broken bars, but the number As mentioned before, a Hall effect sensor was installed
of magnetic poles where there are faulty rotor bars. between two stator slots and the resultant magnetic flux den-

123
Journal of Control, Automation and Electrical Systems (2018) 29:489–499 493

Fig. 8 Methodology used in this work

sity was collected using a PC and a USB Digital Oscilloscope, Fig. 10 shows the setup used for experimental tests. Table 3
Hantek, model HT6022BE, with bandwidth in 20 MHz and shows the induction motor characteristics.
maximum real-time sample rate of 48 MS/s. The data col- The experimental database was used for learning and val-
lected with the Oscilloscope were evaluated in a computer idation purposes. The data were collected from some tests
using the MATLAB and Origin Pro software to perform the performed in laboratory, considering the motor running at
fast Fourier analysis. Figure 9 shows an example of a signal rated frequency and under distinct load levels. It should be
with four values (two amplitudes and their frequencies), and noted that the induction machine has a Foucault braking sys-

123
494 Journal of Control, Automation and Electrical Systems (2018) 29:489–499

Table 2 Some fault rotor patterns used for the learning process
Rotor condition A(freq(11)) freq(11) A(freq(12)) freq(12) A(freq(13)) freq(13) Output

Healthy (10 V) − 95.66 29.99 − 96.54 30.25 − 105.22 30.5 0

Healthy (50 V) − 91.44 30.25 − 94.47 30.5 − 102.87 30 0
Two adjacent broken bars (60 V) − 81.15 30.5 − 84.48 29.25 − 88.51 30.25 1
Two adjacent broken bars (100 V) − 78.63 30.75 − 84.34 28.25 − 91.55 30.5 1
Four adjacent broken bars (no load) − 87.34 30.25 − 90.94 30 − 103.51 30.5 1
Four adjacent broken bars (130 V) − 72.75 31 − 81.07 27.25 − 84.51 27.5 1
Broken bars 1 and 12 (100 V) − 86.75 30.75 − 86.80 28 − 87.87 28.25 2
Broken bars 2, 12, 22 and 32 (100 V) − 88.5 30.75 − 100.01 31 − 101.29 28.25 4

Fig. 9 Two amplitudes and their frequencies used as inputs for the first fuzzy classifier

Table 3 Induction motor characteristics

Characteristic Value

Rated power 7.5 kW

Rated frequency 60 Hz
Rated voltage 220 Vca
Rated primary current 25.2 A
Rated sleep 1740 rpm
Rated slip 0.033
Number of pole pairs 2
Number of rotor bars 38
Number of stator slots 48

Fig. 10 Setup for experimental tests Each load was applied using a varivolt and a rectifier
bridge for the Foucault braking supplying system. It should
be mentioned that the induction motor used for experimental
tests was manufactured for research purposes and the rotor
tem capable of applying different load torques. This brake bars can be easily connected to the end-ring through bolts
system was fed with voltage in a range from 0 V (no load) to and nuts (Fig. 11). Thus, it is possible to simulate distinct
240 V (rated load). In this case, it was possible to consider a scenarios for adjacent and nonadjacent broken bars.
torque variation from 0 (no load) to 40 Nm (rated load) with Figure 12 shows a braking system diagram. The rated
balanced voltage condition. resistance of the braking apparatus is 240 , and its rated cur-

123
Journal of Control, Automation and Electrical Systems (2018) 29:489–499 495

Fig. 11 Bolts and nuts used for emulating broken bars

Fig. 12 Braking system diagram

rent is 1A. The rated braking voltage is 240 Vdc. The motor
and its braking system were manufactured by the Brazilian
company Equacional Eletrica e Mecanica Ltda.

4.2 Experimental Results

Experiments were carried out on a 7.5-kW, four-pole and

three-phase induction motor particularly manufactured for
research purposes. The SCIM was supplied with rated fre-
quency, and the rotor conditions were evaluated considering
a healthy structure, adjacent broken bars and noncontiguous
broken bars.
The brake system was supplied with voltage in the range
from 0 V (no load) to 60 V (slip lower than 1%). The time
window for FFT analysis was 4 s, and the data were sampled
at a frequency equal to 10 kHz. Thus, in this case, the FFT
was performed with 40,000 samples and a 0.25 Hz frequency
resolution. The motor tests have been performed under dis- Fig. 13 FFT analysis applied to a a healthy condition, b one broken
tinct environment temperatures and also under different load bar and c four adjacent broken bars
conditions, thus with several winding temperatures. A PT100
temperature sensor was installed near a stator winding (one
hot point), and it was possible to monitor the one-phase stator
temperature. As supplied by the Hall transducer manufac- netic sensitivity equal to − 0.06%/◦ C (max) and temperature
turer, the sensor and its transducer, used during the tests, have coefficient of resistance equal to + 0.3%/◦ C. In general, the
an accuracy of 2% in 25 ◦ C. Moreover, the manufacturer sup- temperature of the tests varied between 25 and 43 ◦ C; thus,
plied the following information: operating temperature range the operational cases for the 100 scenarios were built for
between − 40 and 100 ◦ C, temperature coefficient of mag- ANFIS approach in this temperature variation.

123
496 Journal of Control, Automation and Electrical Systems (2018) 29:489–499

Table 4 Validation data used for fuzzy classification

Scenario Description (Brake voltage) Slip (%)

VC1 Healthy motor (20 V) 0.22

VC2 One broken bar (60 V) 1
VC3 Two adjacent broken bars (60 V) 1
VC4 Broken bars 1 and 6 (10 V) 0.17
VC5 Broken bars 1 and 12 (60 V) 1
VC6 Broken bars 1, 2 and 15 (50 V) 0.77
VC7 Broken bars 2, 3, 23 and 24 (no load) 0.11
VC8 Broken bars 2, 12, 22 and 32 (60 V) 1
VC9 Broken bars 1, 2, 12 and 22 (10 V) 0.17
VC10 Broken bars 1, 2, 12, 22 and 23 (30 V) 0.33

4.2.1 Results and Discussion

Fig. 14 Second fuzzy classifier in Simulink environment
4.2.2 Frequency Domain Data Obtained from the Hall
Sensor

The FFT was applied to the Hall sensor data to identify fre-
quencies and amplitudes from the first harmonic components.
Figure 13 shows the FFT analysis for three conditions, such
as healthy rotor, one broken bar and four adjacent broken
bars.
Table 3 shows the spectral analysis of the first harmonic
data for no-load and low-load conditions, for example. Noted
that the amplitudes are lower for adjacent failure cases when
compared to noncontiguous breakage conditions. As men-
tioned before, 100 scenarios were considered for the fuzzy
classifier approach, but the cases wherein the SCIM was run- Fig. 15 Validation results for both fuzzy classifiers
ning at very low slip (s ≤ 1%) have a special interest in this
work.

4.2.3 Experimental Results for Fuzzy Classification

The experimental results were evaluated for two fuzzy clas-

sifiers. As previously described, the first one considered four
inputs and only one output, and the second one considered six
inputs and also only one output. The validation cases (VC)
are shown in Table 4, and these values were used for both
fuzzy models.
The fuzzy model, generated by ANFIS, was implemented
in MATLAB/Simulink to proceed with the simulation, as Fig. 16 Validation and learning data used for the first fuzzy classifier
shown in Fig 14.
Figure 15 shows the validation results for the fuzzy classi-
fiers. Note that the first model (4 inputs) considered a failure In addition, Figs. 16 and 17 show the classification results
case even when the rotor is healthy, but the second one (6 using both validation and learning data for the first and the
inputs) achieved a good result in this case. The second model second fuzzy models, respectively.
was able to detect and to classify the breakage scenarios bet- Some false-positive cases were detected by the first model,
ter than the first fuzzy classifier. as shown in Fig. 16, but the second fuzzy model was able to

123
Journal of Control, Automation and Electrical Systems (2018) 29:489–499 497

detected the broken bars. These scenarios took into account

the motor running at slip lower than 1%. For 39 scenarios,
i.e., by using operational scenarios with slip ≤ 1%, the accu-
racy for broken bars detection is shown in Fig. 18, which
also shows that both fuzzy models have good accuracy to
detect adjacent and nonadjacent broken bars. Moreover, the
second fuzzy model reached good accuracy to identify mag-
netic poles with broken bars (around 90%) for SCIM running
at very low slip.
The performance of the second fuzzy model was mea-
sured by the mean absolute error (MAE). For scenarios where
Fig. 17 Validation and learning data used for the second fuzzy classifier SCIM was running at very low slip, i.e., lower than 1%, the
MAE value was around 0.07, thereby with a good approxi-
mation for broken bars detection and evaluation of magnetic
poles with damaged bars.
Table 5 presents a summary of other works and their
respective results for comparison purposes. From these
comparisons, the main contributions of this paper can be
summarized as follows:

i-) It is not necessary to estimate the rotor slip, as required

by traditional MCSA technique and also addressed by
others;
ii-) This approach was able to detect adjacent and nonad-
jacent broken rotor bars, as well as to identify distinct
Fig. 18 Accuracy to detect and to identify adjacent and nonadjacent magnetic poles with broken bars;
broken bars iii-) It is a robust method to detect rotor faults for SCIM run-
ning at low-load or no-load condition, i.e., at very low
slip;
better classify a healthy condition when compared to failure iv-) It requires a short acquisition time (around 4 s), thus not
scenarios. requiring a large memory and high computing power to
The second model failed in scenario number 12 (bro- perform FFT; and
ken bars 1 and 6 at no load), as shown in Fig. 17, since v-) The proposed technique uses a low-cost Hall effect sensor
a false-negative condition was detected. In these cases, 31 to measure the air gap flux.
operational scenarios were used for comparison. However,
the second model has been capable of classifying the faulted
cases in around 90%. In 31 scenarios, for example, only
3 cases were classified incorrectly, although two of them

Table 5 Summary of the recently published papers in comparison with the present research
Reference Measurement time Low load or no load Slip estimation Nonadjacent cases

Faiz and Ebrahimi (2009) Not reported No Yes Yes

Puche-Panadero et al. (2009) 100 s Yes Yes No
Sizov et al. (2009) Not reported No Yes Yes
Riera-Guasp et al. (2010) 50 s No Yes Yes
Maouche et al. (2014) 10 s No Yes No
Choi et al. (2015) 200 s No Yes No
Naha et al. (2016) Not reported Yes Yes No
Samanta et al. (2017) 30 s Yes Yes No
Burriel-Valencia et al. (2017) 10 s No Yes No
This research 4s Yes No Yes

123
498 Journal of Control, Automation and Electrical Systems (2018) 29:489–499

5 Conclusions for harsh industrial application. IEEE Transactions on Industrial

Electronics, 62(3), 1760–1769.
Dias, C. G., & Chabu, I. E. (2014). Spectral analysis using a hall effect
This paper presents a neural-fuzzy approach to detect and to sensor for diagnosing broken bars in large induction motors. IEEE
classify broken rotor bars in squirrel-cage induction motors Transactions on Instrumentation and Measurement, 63(12), 2890–
running at very low slip. The proposed method was able to 2902.
detect the relative position of the broken bars around mag- Dias, C. G. & de Sousa, C. M. (2017). An experimental approach
for diagnosis of adjacent and nonadjacent broken bars in induc-
netic poles of the SCIM. The detection and classification tion motors at very low slip. In 2017 IEEE International Electric
process took into account the use of major amplitudes and Machines and Drives Conference (IEMDC) (pp. 1–6).
their frequencies components, using the fast Fourier trans- el Malek, M. A., Abdelsalam, A. K., & Hassan, O. E. (2017). Induction
form analysis, whose data were obtained from a Hall sensor motor broken rotor bar fault location detection through envelope
analysis of start-up current using hilbert transform. Mechanical
installed inside the machine. The amplitudes and frequencies
Systems and Signal Processing, 93(Supplement C), 332–350.
change according to the position of the faulty bars and also Faiz, J., & Ebrahimi, B. (2009). Locating rotor broken bars in induc-
with fault severity. This research employed a short acquisi- tion motors using finite element method. Energy Conversion and
tion time (around 4 s) and a frequency resolution equal to Management, 50(1), 125–131.
Godoy, W. F., da Silva, I. N., Goedtel, A., Palacios, R. H. C., & Lopes,
0.25 Hz, since the sampling frequency was 10 kHz. It should
T. D. (2016). Application of intelligent tools to detect and clas-
be noted that MCSA, for example, requires a frequency res- sify broken rotor bars in three-phase induction motors fed by an
olution around 0.01 Hz, to estimate slip and to detect broken inverter. IET Electric Power Applications, 10(5), 430–439.
bars. Guan, Y., Zhu, Z. Q., Afinowi, I. A. A., Mipo, J. C., & Farah, P. (2014).
Calculation of torque-speed characteristic of induction machine
The experimental results showed the possibility to detect for electrical vehicle application using analytical method. In 2014
and to classify adjacent and nonadjacent broken bars, even International Conference on Electrical Machines (ICEM) (pp.
for a motor running at low slip, i.e., with braking system 2715–2721).
voltage lower than or equal to 60 V. The accuracy in these Jang, J. S. R. (1993). Anfis: Adaptive-network-based fuzzy inference
system. IEEE Transactions on Systems, Man, and Cybernetics,
cases was greater than 90%. However, the present fuzzy 23(3), 665–685.
approach should be improved to better identify magnetic Jung, J. H., Lee, J. J., & Kwon, B. H. (2006). Online diagnosis of
poles with broken bars, although the second fuzzy model induction motors using mcsa. IEEE Transactions on Industrial
reached a good precision (≈ 87%). Although the noninva- Electronics, 53(6), 1842–1852.
Laala, W., Guedini, S., & Zouzou, S. (2011). Novel approach for diagno-
sive techniques are currently widespread, this method allows sis and detection of broken bar in induction motor at low slip using
detecting and classifying broken bars not only for new large fuzzy logic. In 2011 IEEE international symposium on diagnostics
induction machines, but also for those undergoing rewind for electric machines, power electronics drives (SDEMPED) (pp.
processes. In addition, a low-cost Hall sensor, around US$ 511–516).
Lee, S., Hyun, D., Kang, T., Yang, C., Shin, S., Kim, H., et al. (2015).
400.00 (four hundred dollars), was used. Finally, the present Identification of false rotor fault indications produced by on-line
fuzzy logic approach should be embedded in low-cost hard- mcsa for medium voltage induction machines. In 2015 61st IEEE
ware for developing an alternative tool, capable of diagnosing pulp and paper industry conference (PPIC) (Vol. 1, pp. 1–9).
broken rotor bars in large induction motors. Lizarraga-Morales, R. A., Rodriguez-Donate, C., Cabal-Yepez, E.,
Lopez-Ramirez, M., Ledesma-Carrillo, L. M., & Ferrucho-
Alvarez, E. R. (2017). Novel FPGA-based methodology for early
Acknowledgements The authors are thankful to the São Paulo Research
broken rotor bar detection and classification through homogeneity
Foundation (FAPESP), Grant #2016/02525-1, and Nove de Julho Uni-
estimation. IEEE Transactions on Instrumentation and Measure-
versity (UNINOVE) for their support.
ment, 66(7), 1760–1769.
Lu, B., Habetler, T. G., & Harley, R. G. (2008). A nonintrusive and in-
service motor-efficiency estimation method using air-gap torque
with considerations of condition monitoring. IEEE Transactions
References on Industry Applications, 44(6), 1666–1674.
Maouche, Y., Oumaamar, M. E. K., Boucherma, M., & Khezzar, A.
Al-Badri, M., Pillay, P., & Angers, P. (2015). A novel in situ efficiency (2014). Instantaneous power spectrum analysis for broken bar fault
estimation algorithm for three-phase IM using GA, IEEE method detection in inverter-fed six-phase squirrel cage induction motor.
F1 calculations, and pretested motor data. IEEE Transactions on International Journal of Electrical Power and Energy Systems,
Energy Conversion, 30(3), 1092–1102. 62(Supplement C), 110–117.
Burriel-Valencia, J., Puche-Panadero, R., Martinez-Roman, J., Sapena- Naha, A., Samanta, A. K., Routray, A., & Deb, A. K. (2016). A method
Bano, A., & Pineda-Sanchez, M. (2017). Short-frequency Fourier for detecting half-broken rotor bar in lightly loaded induction
transform for fault diagnosis of induction machines working in motors using current. IEEE Transactions on Instrumentation and
transient regime. IEEE Transactions on Instrumentation and Mea- Measurement, 65(7), 1614–1625.
surement, 66(3), 432–440. Palacios, R., Goedtel, A., & Godoy, W. F. (2016). Fault identification in
Chandralekha, R., & Jayanthi, D. (2016). Diagnosis of faults in three the stator winding of induction motors using PCA with artificial
phase induction motor using neuro fuzzy logic. International Jour- neural networks. Journal of Control, Automation and Electrical
nal of Applied Engineering Research, 11(8), 5735–5740. Systems, 27(406), 1–13.
Choi, S., Pazouki, E., Baek, J., & Bahrami, H. R. (2015). Iterative con- Puche-Panadero, R., Pineda-Sanchez, M., Riera-Guasp, M., Roger-
dition monitoring and fault diagnosis scheme of electric motor Folch, J., Hurtado-Perez, E., & Perez-Cruz, J. (2009). Improved

123
Journal of Control, Automation and Electrical Systems (2018) 29:489–499 499

resolution of the mcsa method via hilbert transform, enabling the Singh, H., Seera, M., & Abdullah, M. Z. (2013). Detection and diagnosis
diagnosis of rotor asymmetries at very low slip. IEEE Transactions of broken rotor bars and eccentricity faults in induction motors
on Energy Conversion, 24(1), 52–59. using the fuzzy min-max neural network. In The 2013 international
Riera-Guasp, M., Cabanas, M. F., Antonino-Daviu, J. A., Pineda- joint conference on neural networks (IJCNN) (pp. 1–5).
Snchez, M., & Garc-a, C. H. R. (2010). Influence of nonconsec- Sizov, G. Y., Sayed-Ahmed, A., Yeh, C. C., & Demerdash, N. A.
utive bar breakages in motor current signature analysis for the O. (2009). Analysis and diagnostics of adjacent and nonadjacent
diagnosis of rotor faults in induction motors. IEEE Transactions broken-rotor-bar faults in squirrel-cage induction machines. IEEE
on Energy Conversion, 25(1), 80–89. Transactions on Industrial Electronics, 56(11), 4627–4641.
Sadeghian, A., Ye, Z., & Wu, B. (2009). Online detection of broken rotor Thomson, W. T. & Fenger, M. (2003). Case histories of current sig-
bars in induction motors by wavelet packet decomposition and nature analysis to detect faults in induction motor drives. In IEEE
artificial neural networks. IEEE Transactions on Instrumentation international electric machines and drives conference. IEMDC’03
and Measurement, 58(7), 2253–2263. (Vol. 3, pp. 1459–1465).
Samanta, A. K., Naha, A., Routray, A., & Deb, A. K. (2017). Fast and Zarei, J. (2012). Induction motors bearing fault detection using pattern
accurate spectral estimation for online detection of partial broken recognition techniques. Expert Systems With Applications, 39(68–
bar in induction motors. Mechanical Systems and Signal Process- 73), 68–73.
ing, 98(Supplement C), 63–77. Zhang, P., Du, Y., Habetler, T. G., & Lu, B. (2011). A survey of condition
Sauer, I. L., Tatizawa, H., Salotti, F. A., & Mercedes, S. S. (2015). A monitoring and protection methods for medium-voltage induction
comparative assessment of Brazilian electric motors performance motors. IEEE Transactions on Industry Applications, 47(1), 34–
with minimum efficiency standards. Renewable and Sustainable 46.
Energy Reviews, 41, 308–318.

123