Application of Different Machine Learning

Strategies for Current- and Vibration-based Motor

Bearing Fault Detection in Induction Motors
Hugo D. L. Rações∗ , Fernando J. T. E. Ferreira† , João M. Pires‡ and Carlos V. Damásio§
Dept. of Computer Science, Universidade Nova de Lisboa (FCT/UNL), Lisbon, Portugal∗‡§
Dept. of Electrical and Computer Engineering, Universidade de Coimbra, Coimbra, Portugal†
Altran Portugal∗ , NOVA-LINCS∗‡§
Email: ∗ hd.racoes@campus.fct.unl.pt, † ferreira@deec.uc.pt, ‡ jmp@fct.unl.pt, § cd@fct.unl.pt

Abstract—In this paper, the application of different machine Regarding bearing faults in IM motors, vibration and stator
learning strategies for current- and vibration-based detection of currents analysis are two valid CM strategies and several
bearing faults in squirrel-cage induction motors is studied. This works have demonstrated these techniques to be effective for
study compares several feature extraction strategies such as a
statistical and spectral analysis of vibration, a statistical analysis use in Machine Learning (ML). In [5], 95% accuracy was
of the Hilbert’s Transform envelope of vibration, an analysis achieved by performing a statistical analysis of the vibration
of the currents deviation to a perfect sinusoid and a statistical signal, while in [6] the success rate was of 96% by merging
and spectral analysis of the Park’s Vector Modulus, with its information from the Fast Fourier Transform (FFT) of the
performances being evaluated with the Support Vector Machine, vibration and the statistical analysis. In [7], perfect results
Artificial Neural Network, Random Forests and Extreme Gradi-
ent Boosting algorithms. A comparison of results obtained using were obtained by analyzing the envelope spectrum of the
sampling frequencies of 0.8 kHz, 1 kHz, 2 kHz, 5 kHz and 10 vibration using the FFT of the Hilbert Transform (HT) and,
kHz and analysis periods between 20 ms and 100 ms is made in [8], a success rates of 100% was registered by statistically
and promising models are achieved even with the lowest sampling analyzing the Continuous Wavelet Transform (CWT) of the
frequencies. vibration. This transform was also used in [9] to process the
Index Terms—induction motor, bearing fault detection, vi-
bration analysis, current analysis, machine learning, predictive stator currents and 100% accuracy was achieved. In [10], the
maintenance, condition monitoring. classification was perfect by analyzing the Park Transform of
the stator currents and, in [11] and [12], the spectral analysis
I. I NTRODUCTION of the Park’s Vector modulus (PVM), commonly known as
Electric Motors (EMs) and systems that derive from them, the Extended Park’s Vector Approach (EPVA), was shown to
according to the International Energy Agency (IEA), account indicate the presence of bearing faults.
for more than 50% of global electricity consumption [1]. In In this paper, the detection of bearing faults in IMs with the
industry, three-phase induction motors (IMs) are used in more use of ML is studied, comparing the performances obtained
than 90% of the industrial power drive systems [2], consuming by analyzing both vibration and stator currents with differ-
approximately 2/3 of the electrical energy in that sector [3]. ent techniques, as well as the impact of different sampling
Most IM faults develop gradually, evolving from an initial frequencies on both since, ideally, the lower the sampling
defect to an actual fault (the propagation time of the fault frequency required, the greater the economic benefit (a lower
depends on several factors) [4] and, in industry, they are amount of data needs to be collected and a lower cost sensor
typically repaired 2 to 4 times over their lifetime of 12 to is used). The variation of the analysis period (duration of the
20 years. Despite the low breakdown probability, there is time interval from which the analyses and their extraction of
great interest in avoiding an unexpected breakdown since an features is performed) was also evaluated in order to study its
unscheduled stop of the motor has much higher costs than effect on the success rate of the classification with the different
a scheduled one (as result of the unscheduled stop of the sampling frequencies.
systems dependent on the motor) thus, from a technical and
economic perspective, investment in the improvement of the II. E XPERIMENTAL SETUP AND DATASET DESCRIPTION
efficiency and reliability of electric motors is, in general, very A dataset was generated by simulating different bearing
attractive for the vast majority of the industries and a predictive fault states on a IE3-class, 4-kW, 4-pole, 400-V, 50-Hz,
maintenance strategy based on condition monitoring (CM) can three-phase, squirrel-cage induction motor. This motor has
tremendously increase this reliability. FAG 6207-2RS bearings of dimensions and characteristics
described in Fig. 1.
This work is supported by NOVA LINCS (UID/CEC/04516/2013) with the
financial support of FCT - Fundação para a Ciência e a Tecnologia, through Data collection was performed at sampling frequencies of
national funds. 0.8 kHz, 1 kHz, 2 kHz, 5 kHz and 10 kHz with the experi-

978-1-7281-4878-6/19/$31.00 ©2019 IEEE 68

 TABLE I
N UMBER OF PERIODS OF 20 MS ( CYCLES ) FOR EACH BEARING STATE ,
MOTOR LOAD AND SAMPLING FREQUENCY.

Sampling Frequency (kHz)

Diameter Load Speed 0.8 1 2 5 10
(mm) (%) (RPM)
0 0 1499 562 549 499 599 999
[Healthy] 50 1482 562 549 499 499 999
100 1460 562 549 499 599 999
1 0 1499 1062 1049 874 1099 1499
[Faulty] 50 1480 1062 1049 874 999 1499
Fig. 1. Motor bearing dimensions. 100 1460 1062 1049 849 999 1499
2 0 1499 562 549 499 599 999
[Faulty] 50 1482 562 549 499 599 999
100 1460 562 549 499 599 999
mental setup shown in Fig. 2 (accelerometer was centered on 4 0 1499 874 949 899 999 1249
the side of the motor housing), obtaining records regarding [Faulty] 50 1482 874 999 824 999 1249
100 1460 937 999 874 999 1249
the electrical currents, electrical voltages and vibration (radial
acceleration) of the motor. The voltages were not considered
for the classification of the fault condition since a balanced a decrease in the rotational speed of the rotor depending on
voltage system with very low harmonic content (total harmonic the increase in load.
distortion lower than 0.1%) was used in the experimental tests
(Programmable Power Supply, Pacific Power 3300AFX). III. F EATURE E XTRACTION
Two different bearing states were considered: healthy bear- A. Statistical analysis of the vibration and the Hilbert’s Trans-
ing and bearing with an outer race fault. The last state was form envelope of vibration
simulated by performing small central holes in the outer race Regarding the statistical analysis of vibration, several fea-
with diameters of 1 mm, 2 mm and 4 mm, and the bearing tures were obtained: rms, mean, median, variance, standard
defect was placed in the motor at the 6 o’clock position, that deviation, the maximum value, the minimum value, the peak-
is, pointing to the motor base (the outer race of the bearing to-peak, the skewness, the kurtosis, the crest factor and the
always remains stationary). In terms of motor load, this was maximum value of the module of the vibration. The mean
imposed by a dynamometer brake and the collection was made and median were calculated over the module of the vibration
at 3 different loads: no load (0%), half load (50%) and full given the nature of the movement of radial acceleration of the
load (100%) for each one of the fault states. motor and consequent tendency to register values close to 0.
Data collected under the same conditions, but with different For the statistical analysis of the Hilbert’s Transform enve-
sampling frequencies, were collected sequentially and the only lope of the vibration, the same features were extracted except
variation is not just the sampling frequency but the data itself. the maximum of the module (which would equal the maximum
In addition to showing the motor bearing states, fault value) and the crest factor.
diameters, load level, and rotor rotational speed during the
experimental tests, in Table I the amount of data available for B. Vibration spectral analysis
each of the situations in terms of periods of 20 ms (a cycle of In order to perform this analysis, the vibration signal was
electric currents/voltages) is also described. It can be seen, for converted from the time domain to the frequency domain using
example, that for the situation where the bearing has a defect the FFT and isolating the signal by frequencies.
with a diameter of 1 mm, the motor has a load of 0%, and data For each evaluated frequency component, its frequency,
collection was performed with a sampling frequency of 1 kHz, magnitude and spectral density defined as
we have a record of 1049 periods of 20 ms (corresponding to a
Mf
data collection period of about 21 seconds). This Table shows Df = Pn (1)
i=0 Mi
where Mf represents the magnitude of frequency f and
Df indicates the percent signal magnitude associated with
frequency f in relation to the sum of the magnitudes of all
frequency components were considered.
C. Analysis of the currents deviation to a perfect sinusoid
The evaluation of the deviation of the currents to a perfect
sinusoid was carried out since it is theoretically expected that a
faulty bearing condition will be revealed in the electric motor
currents by increasing its deviation to a perfect sinusoid.
In order to perform the calculation of this deviation, the
Fig. 2. Experimental setup. approximation of a sine function with the same period of the

69
currents is made to each of the electric currents, being later B. Fault Detection through Vibration Analysis
calculated the deviation between the function obtained and the
As result of the multiple analyses performed on vibration
current according to
mentioned on section III, 7 feature sets (FSs) were defined:
FS 1. Statistical analysis of vibration
r Pn
x x 2
i=1 (ci − si )
Deviationx = (2) FS 2. Statistical analysis of Hilbert’s Transform envelope of
n
vibration
where x represents one of the electric currents, cx i represents
FS 3. Vibration spectral analysis (top 6 frequency components
the signal of that current, sx i represents the perfect sinusoid
with the higher magnitudes)
and n is the number of points sampled.
FS 4. Statistical analysis of vibration and Hilbert’s Transform
Besides the deviation of each of the currents to a perfect
envelope of vibration
sinusoid, the average of the deviation of the three currents (A,
FS 5. Statistical analysis of vibration + Vibration spectral
B and C) was also evaluated.
analysis
D. Analysis of the Park’s Vector Modulus FS 6. Statistical analysis of Hilbert’s Transform envelope of
The current PVM is given by (3) and, under ideal condi- vibration + Vibration spectral analysis
tions, it is clear from any spectral component (only a direct FS 7. Statistical analysis of vibration and Hilbert’s Transform
current value is present in the current PVM) [13]. envelope of vibration + Vibration spectral analysis
q Comparing the performance of all FSs with a 20 ms analysis
PVM = id (t)2 + iq (t)2 (3) period (Fig. 3) using RF as a reference algorithm, it is observed
that all FSs show a high predictive power with relatively high
where id (t) and iq (t) are given by kappa values. Those with features resulting from the vibration
 q q q spectral analysis stand out with average kappas (average of
id (t) = 2 iA (t) − 1 iB (t) − 1 iC (t) the kappa obtained with the sampling frequencies of 0.8 kHz,
3 q 6 q 6
(4)
 1 1
iq (t) = 2 iB (t) − 2 iC (t) 1 kHz, 2 kHz, 5 kHz and 10 kHz) above 0.98.
In order to study the extent to which an extended analysis
where iX (t) represents the intensity of current X at instant t. period could improve the fault classification with all the
Similarly to the statistical evaluation of the vibration, several different sampling frequencies used, analysis periods of more
statistical indicators were calculated over the PVM current than 20 ms were also considered and defined from 20 ms to
and it was possible to identify features whose variation is 100 ms with 10 ms intervals (20 ms, 30 ms, 40 ms, etc.).
not directly related to the variation of the load, such as the For FS 1 and FS 2 in Fig. 4 and Fig. 5 (the same color can
skewness, peak-to-peak, variance and kurtosis. have different meanings in each figure since a general mapping
Regarding the spectral analysis of the PVM, the same of colors difficults the visualization), it can be observed that
procedure seen in section III-B was followed by analyzing the the classification tends to be perfect with a analysis period of
frequencies, magnitudes and spectral densities of the spectral 20 ms and a sampling frequency of 5 kHz or more. With an
components. analysis period of 30 ms or 40 ms similar results are achieved
with a sampling frequency of 2 kHz and, from an analysis
IV. M OTOR B EARING FAULT D ETECTION
period of 50 ms, excellent results are obtained with any
A. Machine Learning Procedure sampling frequency. FS 1 presents slightly better results with
The data were normalized with the min-max technique given lower analysis periods (less than 40 ms), especially for lower
the large range of values recorded in the evaluated features sampling frequencies, where FS 2 shows a slight advantage
and, to ensure that the training and test set contained the same for analysis periods between 40 ms and 60 ms (from 70 ms
proportion of different fault condition situations (type of fault, both FSs have the same performance).
fault diameter, and motor load), stratified sampling of both sets
was performed and the division used was 70% for the training
set and 30% for the test set. The validation of the model was
executed using 10-fold cross-validation repeated 3 times.
The ML algorithms Support Vector Machine (SVM), Ar-
tificial Neural Network (ANN), Random Forests (RF) and
Extreme Gradient Boosting (XGBOOST) were used (the
chosen parameters are available on https://gitlab.com/hdlr/
iecon2019-bearingfaultdetection) and the main metric of eval-
uation considered for the evaluation of results was Kappa due
to the high class imbalance. This metric compares the observed
accuracy with the expected accuracy (accuracy that a random
classifier would obtain) and its result varies between -1 and 1,
with 1 being the best result possible. Fig. 3. Average Kappa obtained for all FSs and an analysis period of 20 ms.

70
 algorithms used, in Table II, it is verified that the higher
the sampling frequency or the more extended the analysis
period, the better the results achieved (excellent classifications
are achieved with several FSs and algorithms). Generally, for
lower sampling frequencies/analysis periods, the best result is
obtained by one FS and algorithm (with a sampling frequency
of 0.8 kHz and an analysis period of 20 ms it was the SVM
algorithm, 30 ms the XGBOOST algorithm and 40 ms the
ANN algorithm).
With a sampling frequency of 1 kHz and an analysis period
of 20 ms, the best classification was obtained with the SVM
algorithm. With an analysis period of 30 ms, the classification
Fig. 4. FS 1 kappa variation by analysis period and sampling frequency.
is perfect with the FS 3 (with the SVM algorithm) and FS
7 (with multiple algorithms), and perfect classifications are
verified with several FSs and algorithms for higher analysis
periods. When the sampling frequency is of 2 kHz and the
analysis period is 20 ms, the best result was obtained with FS
5 and the RF algorithm and, for greater analysis periods, the
classification is perfect with several FSs and algorithms.
From a sampling frequency of 5 kHz the classification
obtains excellent results with several FSs independently of
the analysis period. With an analysis period of 20 ms the
classification is perfect with several FSs and algorithms and,
for more extended analysis periods, perfect classifications are
registered with all FSs. With a sampling frequency of 10 kHz
Fig. 5. FS 2 kappa variation by analysis period and sampling frequency.
barely all FSs with all the different analysis periods obtained
perfect classifications with several algorithms.

With FS 3 (Fig. 6), the classification is practically perfect TABLE II

B EST RESULTS OBTAINED FOR ALL SAMPLING FREQUENCY AND
with an analysis period of 30 ms and a sampling frequency of ANALYSIS PERIODS BETWEEN 20 MS AND 50 MS
1 kHz or higher, with the same results being verified for any
sampling frequency with a analysis period of 40 ms or more. kHz ms FS Kappa TP FP TN FN
0.8 20 5 0.9746 2251 9 495 12
FS 3 shows better classification results than those achieved 30 5 0.9909 1506 4 332 1
by FSs 1 and 2 (particularly for sampling frequencies up to 40 3 0.9976 1131 1 251 0
5 kHz) nevertheless, for sampling frequencies higher than 5 50 2, 4, 5, 6, 7 1 902 0 201 0
1 20 7 0.9889 2310 3 489 6
kHz or periods of analysis of about 40 ms or higher, similar 30 3, 7 1 1541 0 327 0
results are obtained with the previously mentioned FSs. 40 3, 5, 6, 7 1 1157 0 246 0
Analyzing the best kappa and confusion matrix (positive 50 3, 4, 5, 6, 7 1 921 0 195 0
2 20 5 0.9986 2003 1 446 0
case corresponds to the faulty state and negative case to 30 3, 4, 6, 7 1 1331 0 297 0
the healthy state) for each sampling frequency (kHz) and an 40 All 1 999 0 222 0
analysis period (ms) between 20 ms and 50 ms with all the 50 All 1 795 0 177 0
5 20 3, 5, 6 1 2361 0 507 0
30 All 1 1571 0 337 0
40 All 1 1176 0 252 0
50 All 1 939 0 201 0
10 20 1, 3, 5, 6, 7 1 3366 0 897 0
30 All 1 2241 0 597 0
40 All 1 1680 0 447 0
50 All 1 1341 0 357 0

C. Fault detection through currents analysis

As result of the multiple analyses performed on the currents
mentioned on section III, 6 FSs were defined:
FS 1. Currents deviation to a perfect sinusoid
FS 2. PVM Statistical Analysis
FS 3. PVM Spectral Analysis (top 5 frequency components
Fig. 6. FS 3 kappa variation by analysis period and sampling frequency. with the higher magnitudes)

71
FS 4. PVM Spectral Analysis + PVM Statistical Analysis
FS 5. PVM Spectral Analysis + Currents deviation to a perfect
sinusoid
FS 6. PVM Spectral Analysis + PVM Statistical Analysis +
Currents deviation to a perfect sinusoid
Comparing the average kappa obtained with each FS with
an analysis period of 20 ms (one cycle of the currents) using
RF as a reference algorithm, it is possible to observe in Fig. 7
that FSs 1 and 2 present a relatively low performance when
compared with the rest, showing a relative low predictive
power. The rest of the FSs show promising results with very
slight differences between them and practically perfect (or
Fig. 8. FS 1 kappa variation by analysis period and sampling frequency.
perfect) classifications with high sampling frequencies.
Similarly to what was done for the vibration analysis,
analysis periods of more than 20 ms were also considered
and defined as 20, 40, 60, 80 and 100 ms.
Evaluating in detail the performance of FS 1 and FS 2 with
the variation of the sampling frequency and the analysis period
in Fig. 8 and Fig. 9, it is verified that FS 1 has no promising
results in any situation (when compared to the others), where
FS 2 demonstrates a greater potential especially when used
with an analysis period of 40 ms or more and from a sampling
frequency of about 2 kHz, and from an analysis period of 80
ms the results are quite good with any sampling frequency.
By observing the variations of the kappa obtained according
to the analysis period and frequency of sampling for FS 3
Fig. 9. FS 2 kappa variation by analysis period and sampling frequency.
in Fig. 10, it is verified that the performances are generally
excellent and that the classification tends to be perfect with
any frequency of sampling starting from an analysis period of
40 ms or with a sampling frequency of 5 kHz for an analysis results with a sampling frequency of 0.8 kHz, 1 kHz and 2 kHz
period of 20 ms. A kappa in the order of 0.97 is displayed were obtained by a single FS and algorithm. With a sampling
with the minimum sampling frequency and analysis period. frequency of 0.8 kHz it was with the XGBOOST algorithm
and, with 1 kHz and 2 kHz, it was with the RF algorithm.
Table III shows the best confusion matrix obtained for each
With these sampling frequencies and greater analysis periods,
sampling frequency (kHz) and an analysis period (ms) between
the best classifications were registered with several algorithms.
20 ms and 60 ms with all the algorithms used and it is possible
to verify that, as seen on the vibration analysis, the higher Similarly to what was observed on the vibration analysis,
the sampling frequency or the more extended the analysis from a sampling frequency of 5 kHz the classification obtains
period, the better the results achieved. FS 1 and FS 2 were not excellent results with several FSs independently of the analysis
evaluated with any other algorithm given their plainly worst period and, with a sampling frequency of 10 kHz, even FS 2
performances compared to FS 3. (that initially showed an inferior predictive power) registered
With an analysis period of 20 ms, the best classification a perfect classification with the RF algorithm.

Fig. 7. Average Kappa obtained for all FSs and an analysis period of 20 ms. Fig. 10. FS 3 kappa variation by analysis period and sampling frequency.

72
 TABLE III
B EST RESULTS OBTAINED FOR ALL SAMPLING FREQUENCY AND
ANALYSIS PERIODS BETWEEN 20 MS AND 60 MS

kHz ms FS Kappa TP FP TN FN
0.8 20 4 0.9916 2256 0 504 7
40 5, 6 1 1131 0 252 0
60 3, 4, 5, 6 1 753 0 168 0
1 20 5, 6 0.9988 2315 0 492 1
40 3, 4, 6 1 1157 0 246 0
60 3, 4, 5, 6 1 766 0 162 0
2 20 6 0.9986 2003 1 446 0
40 3, 4, 5, 6 1 999 0 222 0
60 3, 4, 5, 6 1 663 0 147 0
5 20 3, 4, 5, 6 1 2361 0 507 0
40 3, 4, 5, 6 1 1176 0 252 0
60 3, 4, 5, 6 1 781 0 167 0
10 20 3, 4, 5, 6 1 3366 0 897 0 Fig. 11. Difference in the obtained kappa for current and vibration analyses
40 2, 3, 4, 5, 6 1 1680 0 447 0 (in favor of the current) for each sampling frequency and analysis period.
60 3, 4, 5, 6 1 1116 0 297 0

and, although it is more common the use of motor frame

V. C ONCLUSION vibration to diagnose mechanical faults and line current to
diagnose electrical faults, the use of the latter is potentially
The work described in this paper has shown that it is cheaper and easier to apply in industry, since all the acquisition
possible to achieve excellent results in the detection of bearing system can be installed in the motor switchboard, contrarily
faults in IMs by analyzing its vibration or line currents to the accelerometers used in the vibration measurement that
sampled with all the tested frequencies and an adequate have to be installed in the motor frame.
analysis period using ML techniques. This demonstrates that,
in industrial use, the decision lies in the convenient and R EFERENCES
appropriate time to make the data collection and the cost of [1] U4E, “Accelerating the Global Adoption of Energy-efficient Electric
the sensors for each sampling frequency. Motors and Motor Systems.” United Nations Environment Programme,
2017.
Among the different analyses studied, the spectral analysis [2] F. J. Ferreira, G. Baoming, and A. T. de Almeida, “Reliability and oper-
of the vibration and of the line current PVM led to the best ation of high-efficiency induction motors,” in Industrial & Commercial
results for the tested motor and bearings (achieved by several Power Systems Technical Conference (I&CPS), 2015 IEEE/IAS 51st,
pp. 1–13, IEEE, 2015.
parameterizations of the used ML algorithms and showcasing [3] P. Waide and C. U. Brunner, “Energy-efficiency policy opportunities for
the quality of the proposed FSs). For sampling frequencies electric motor-driven systems,” pp. 11–17, 2011.
up to 2 kHz, FSs that result from the merging of features [4] E. L. Bonaldi, L. E. de Lacerda de Oliveira, J. G. B. da Silva,
G. Lambert-Torresm, and L. E. B. da Silva, “Predictive Maintenance
from these analyses with the others are the slightly better per- by Electrical Signature Analysis to Induction Motors,” 2012.
forming while, for higher sampling frequencies, FSs resulting [5] P. K. Kankar, S. C. Sharma, and S. P. Harsha, “Vibration-based fault
exclusively from these analyses match the best results. diagnosis of a rotor bearing system using artificial neural network and
support vector machine,” International Journal of Modelling, Identifica-
In comparison with the work mentioned in section I, this tion and Control, vol. 15, no. 3, pp. 185–198, 2012.
paper demonstrates the ability to detect faults using lower [6] B. Li, M.-Y. Chow, Y. Tipsuwan, and J. C. Hung, “Neural-network-based
motor rolling bearing fault diagnosis,” IEEE transactions on industrial
sampling frequencies and proposes a different feature extrac- electronics, vol. 47, no. 5, pp. 1060–1069, 2000.
tion approach regarding the spectral analysis of vibration and [7] L. Guo, J. Chen, and X. Li, “Rolling Bearing Fault Classification
of the current PVM. Analyzing the detection of the bearing Based on Envelope Spectrum and Support Vector Machine,” Journal
of Vibration and Control, vol. 15, no. 9, pp. 1349–1363, 2009.
fault through each of the physical quantities in question, the [8] P. Konar and P. Chattopadhyay, “Bearing fault detection of induction
analysis of the currents appears to offer the possibility of using motor using wavelet and Support Vector Machines (SVMs),” Appl. Soft
a lower sampling frequency to obtain the same level of results Comput., vol. 11, no. 6, pp. 4203–4211, 2011.
[9] A. Jawadekar, S. Paraskar, S. Jadhav, and G. Dhole, “Artificial neural
(not necessarily perfect) when compared to the vibration network-based induction motor fault classifier using continuous wavelet
analysis, as seen in Fig. 11. Regarding the minimum sampling transform,” vol. 2, pp. 684–690, 12 2014.
frequency required to register a perfect classification, both [10] R. Bandyopadhyay, P. Puhan, P. Konar, and P. Chattopadhyay, “A Hybrid
Park-ANN approach in fault detection of Induction Motor,” 08 2010.
current and vibration analyses achieve similar results except [11] J. L. H. Silva and A. J. M. Cardoso, “Bearing failures diagnosis in
with an analysis period of 40 ms where the electric currents three-phase induction motors by extended Park’s vector approach,” in
provide a perfect classification with 0.8 kHz and vibration only 31st Annual Conference of IEEE Industrial Electronics Society, 2005.
IECON 2005., pp. 6 pp.–, 11 2005.
with 1 kHz (for an analysis period over 40 ms, both achieve [12] B. Corne, B. Vervisch, S. Derammelaere, J. Knockaert, and J. Desmet,
perfect results from 0.8 kHz). With the appropriate sampling “The Reflection of Evolving Bearing Faults in the Stator Currents
frequencies, both show an excellent classification potential. Extended Park Vector Approach for Induction Machines,” Mechanical
Systems and Signal Processing, vol. 107, pp. 168 – 182, 2018.
In conclusion, the analysis of any of the physical quantities [13] L. Oliveira and A. Cardoso, “Extended Park’s vector approach-based dif-
in question for the detection of bearing faults in the used motor ferential protection of three-phase power transformer,” vol. 6, pp. 463–
shows remarkable results for the used experimental datasets 472, 09 2012.

73