Mechanical Systems and Signal Processing 98 (2018) 63–77

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Fast and accurate spectral estimation for online detection of

partial broken bar in induction motors
Anik Kumar Samanta a,⇑, Arunava Naha b, Aurobinda Routray b, Alok Kanti Deb b
a
Advanced Technology Development Centre, Indian Institute of Technology Kharagpur, 721302, India
b
Dept. of Electrical Engineering, Indian Institute of Technology Kharagpur, 721302, India

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, an online and real-time system is presented for detecting partial broken rotor
Received 8 December 2016 bar (BRB) of inverter-fed squirrel cage induction motors under light load condition. This
Received in revised form 14 April 2017 system with minor modifications can detect any fault that affects the stator current. A fast
Accepted 24 April 2017
and accurate spectral estimator based on the theory of Rayleigh quotient is proposed for
Available online 3 May 2017
detecting the spectral signature of BRB. The proposed spectral estimator can precisely
determine the relative amplitude of fault sidebands and has low complexity compared
Keywords:
to available high-resolution subspace-based spectral estimators. Detection of low-
Autocorrelation matrix
Broken rotor bar
amplitude fault components has been improved by removing the high-amplitude funda-
Closely spaced sinusoids mental frequency using an extended-Kalman based signal conditioner. Slip is estimated
Empirical cumulative distribution function from the stator current spectrum for accurate localization of the fault component.
Fault diagnosis Complexity and cost of sensors are minimal as only a single-phase stator current is
MCSA required. The hardware implementation has been carried out on an Intel i7 based embed-
MUSIC ded target ported through the Simulink Real-Time. Evaluation of threshold and detectabil-
Rayleigh-quotient ity of faults with different conditions of load and fault severity are carried out with
Simulink real-time
empirical cumulative distribution function.
Squirrel cage induction motor
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction

A condition monitoring system for early detection of squirrel cage induction motor (SCIM) faults like broken rotor bar
(BRB) can significantly enhance the operation efficiency of any industry. As reported, BRB accounts for about 5–10% of all
induction motor faults [1]. BRB is initiated when small cracks develop at the junction between the bars and the end ring.
The resultant signatures for cracked and broken rotor bars on the current spectrum are the effect due to rotor circuit asym-
metries [2–4] giving rise to multiple frequency components around the fundamental as

f brb ¼ ð1  2ksÞf o ð1Þ

where s is the slip, k is any integer, and f o is the fundamental frequency. High-resolution spectral estimators like multiple
signal classification (MUSIC) [5,6] and estimation of signal parameters via rotational invariance technique (ESPRIT) [7,8] have
gained prominence over the classical method of the power spectrum because of its robustness and resolution capacity for
detecting BRB under low load conditions. Still there are concerns over critical issues related to computational complexity
and accurate amplitude estimation of the detected fault components. Use of parametric spectral estimators for fault

⇑ Corresponding author.
E-mail address: ece.anik@gmail.com (A.K. Samanta).

http://dx.doi.org/10.1016/j.ymssp.2017.04.035
0888-3270/Ó 2017 Elsevier Ltd. All rights reserved.
64 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77

diagnosis using maximum lilkihood estimation with a particular signal model can be found in [9]. Spectral estimation based
on total least square methods was developed by [10] for fault detection.
Fault diagnosis can be accomplished by analyzing a variety of signals of the motor such as vibration [11–13], current
[14–16,7,17], magnetic fields [18,19], supply voltage modulation [20], active-reactive power [21,22], acoustics [11], ther-
mal field [23], thermal imaging [24], slot harmonics [25], torque [26], low-voltage offline testing [27], etc. Signals like flux
[19] and instantaneous power factor with phase [22] were used to alleviate the problems faced by motor current signature
analysis (MCSA) due to load-torque. Indicators of BRB, which are independent of load-torque oscillations were proposed in
[28–30]. In [31], a low sampling rate of 200 samples/s was used for improved resolution in detecting BRB under low load
conditions. Transient and non-stationary signal processing using the wavelet decomposition of the startup current
[32,12,16], adaptive slope transform [33] are also popular. Effect of simultaneous occurrence of static eccentricity, BRB,
and speed ripples were studied analytically and experimentally in [15]. A winding function model based method with
parameter estimation using current, rotor speed, and torque is shown in [34]. The authors in [35] have used empirical
mode decomposition for direct, and inverter-fed motors to detect BRB. A comprehensive review of recent advances for
detection of BRB can be found in [36].
The severity of BRB is indicated by the magnitude of the fault component [37,17]. However, subspace-based methods can-
not give exact information about the amplitude of the fault components. Hence, simulated annealing algorithm was used to
determine the correct amplitude and eventual fault severity in [8]. In [7,38] the amplitude was estimated using least square
estimation, which is equivalent to computing the discrete Fourier transform (DFT) for a single frequency. However, DFT is not
suitable for estimating the amplitude of closely spaced sinusoids [39,40]. Moreover, these methods require extra computa-
tional resources for their execution when used in conjunction with MUSIC and ESPRIT. Modulation of the stator current due
to partial BRB is weak, and its detection in a light load condition is challenging as the fault components are very close to the
fundamental [17] and have low amplitude [41]. Detecting weak faults require the fundamental frequency to be suppressed
effectively without affecting the closely spaced fault components. Commonly, this is achieved by using a sharp notch filter
[31]. But, variable frequency operation with load changes require the central frequency of the notch filter to track the fun-
damental frequency, and its cutoff bandwidth to be adaptive to the slip. In developing unsupervised fault detectors, imple-
menting notch filter with these characteristics is inconvenient. Therefore, an extended Kalman filter (EKF) based signal
conditioning method is adopted to remove the fundamental component. This method tracks and attenuates only the funda-
mental component to improve the detection of close sidebands. Detection of faults in low load condition was accomplished
by Hilbert modulus with FFT (0.2% slip) [42], Hilbert modulus with ESPRIT (0.33% slip) [43], Fourier analysis (1.38% slip) [44],
Teager - Kaiser energy operator (0.4% slip) [45]. In [46], Fourier analysis of stator current envelop is carried out to detect BRB
under a low slip of 0.11%. However, a high initial sampling of 50 kHz makes it disadvantageous for low-cost hardware imple-
mentation. The majority of the research focussed on detecting single and multiple BRB. Detection of partially BRB were
demonstrated in [47,4,3]. However, detecting partial BRB in low-slip applications especially for the inverter-fed motor is
yet to be addressed. In the proposed paper, detection of a partially broken bar for the inverter-fed SCIM has been achieved
with 0.2% slip.
The major contribution of this paper can be enumerated by (a) Development of a novel spectral estimator which can

1. estimate the location of fault frequency components with very high accuracy in noisy environment and has a lower com-
putational complexity than MUSIC,
2. estimate the magnitude of fault frequencies accurately unlike subspace-based methods like MUSIC, ESPRIT, etc.,
3. avoid spurious peaks as it doesn’t require the information about the number of sinusoids. This decreases the chance of
false-alarms and missed detections,

(b) An elegant fault detection algorithm is developed using the novel spectral estimator having the following attributes:

1. can detect single BRB fault with different levels of damage under low-slip. The lowest slip for the medium-sized motor to
detect a partially broken bar has been found to be 0.2% under 1.9% of the rated load,
2. a novel EKF-based signal conditioner is developed to estimate and remove the fundamental supply frequency component
from the input. The detectability of closely spaced fault frequencies due to partial BRB under low slip has been improved
by this conditioning. Moreover, being a time-domain based sequential technique, it can also be used for non-stationary
applications,
3. the overall system is implemented on an embedded hardware platform for online and real-time (RT) fault detection with
only a single phase stator current as input.

2. The proposed spectral estimator

The signal model used for developing the spectral estimator is represented by
X
P
x½n ¼ ai ejðxi nþ/i Þ þ m½n; n ¼ 0; 1; . . . ; ðN  1Þ ð2Þ
i¼1
 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77 65

where x½n is the nth sample of the input signal out of N samples and is a sum of P complex exponentials with additive white
gaussian noise m½n. The ith sinusoid have an amplitude ai , random phase /i , and normalized frequency xi ¼ 2pf i =F s . f i is the
ith frequency component, and F s is the sampling frequency.

2.1. Formation of autocorrelation matrix

The autocorrelation matrix (Rx ) of size L, is constructed by forming the data matrix (X) of same size [39] without any
interleaved samples as shown in (3).
2 3
xð0Þ  xðL  1Þ
6 7
X¼4    5 ð3Þ
xðL  1Þ    xð2L  2Þ
Assuming the process to be ergodic, Rx is estimated as

b x ¼ 1 fXH  Xg
R ð4Þ
L

2.2. Theoretical background of the spectral estimator

If A is a symmetric matrix with eigenvector v , then there exist a corresponding eigenvalue (k), which can be approxi-
mated by using the theory of Rayleigh quotients [48, Pg. 301–304] as

^k ¼ v Av
H
ð5Þ
v v
H

b x is a symmetric matrix, and its eigenvectors are related to the independent frequency components present in the signal.
R
wðxÞ as given in (6) will become an eigenvector of R b x if a sinusoid of frequency x is present [49, Pg. 452] in the signal.
^ xÞ (7) when x matches the frequency of a sinusoid present in the signal.
Sweeping x between 0 and 2p will give a peak of hð
^ xÞ is an eigenvalue corresponding to the eigenvector wðxÞ and is proportional to the amplitude of the sinusoid
The peak hð
present in the signal.
 H
wðxÞ ¼ ejx0 ejx    ejxðL1Þ ð6Þ

and wðxÞH wðxÞ ¼ L, and x 2 ½0; 2p.

Hb
^ xÞ ¼ wðxÞ R x wðxÞ
hð ð7Þ
wðxÞH wðxÞ

2.3. Amplitude estimation

Putting the estimate of (4) in (7) gives

h i
^ xÞ ¼ 1 ðXwðxÞÞH  ðXwðxÞÞ
hð ð8Þ
L2
The product XwðxÞ of (8) can be found as
2 32 3
xð0Þ  xðL  1Þ ejx0 " #T
6 76 .. 7 X
L1 X
L1
XwðxÞ ¼ 4    54 . 5¼ xðiÞejxi  xði þ L  1Þejxi ð9Þ
xðL  1Þ    xð2L  2Þ ejxðL1Þ
i¼0 i¼0

Now, we define that

X
L1
X q ðxÞ ¼ xðr þ qÞejxr ð10Þ
r¼0

and using this in (9), it is obtained that

XwðxÞ ¼ ½ X 0 ðxÞ    X L1 ðxÞ 

T
ð11Þ

putting the value of (11) in (8) gives

X
L1
^ xÞ ¼ 1
hð jX i ðxÞj
2
ð12Þ
2
L i¼0
66 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77

Now taking square of the absolute value on both sides in (10) and putting it in (12) gives
 2
L1 X 
^ 1X 
L1
jxv 
hðxÞ ¼ 2  xðv þ iÞe  ð13Þ

L i¼0 v ¼0 

From (13), it is observed that the proposed spectral estimator takes a form that is similar to the spectral information
obtained from Welch’s periodogram without windowing. This result was obtained because of the choice of the data matrix.
However, the inception of the method was derived from the fact that the eigenvalues of the autocorrelation matrix corre-
sponding to the signal eigenvectors obtained from the Rayleigh quotients can give vital information about the sinusoidal
amplitude present. In the next section, a simplified way to implement this method in matrix form will be discussed. Now
a single sinusoid at x with spectral peak magnitude hð^ x Þ without any noise can be modeled using (2) as
k k

x½n ¼ ak ejðxk nþ/k Þ ð14Þ

Using the value of x½n in (13) gives
 2
L1 X
X L1 
^ x Þ¼ 1  jxk ðv þiÞ j/k jxk v 
hð k  ak e e e 
L i¼0  v ¼0
2 
 2
L1  L1 
1X  jxk i j/k X 
¼ 2 ak e e 1
L i¼0  v ¼0
 ð15Þ

a2k L2 X
L1
¼ 1
L2 i¼0

¼ La2k
^k as
From (15) the amplitude ak is estimated by a
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
^k ¼
a ^ x Þ=L
hð ð16Þ
k

Fig. 1 shows the efficacy of the proposed spectral estimator in estimating the amplitude of closely spaced sinusoids. Two
sinusoids at 49.8 Hz and 50 Hz with equal amplitude have been chosen as input for this experiment. Different signal to noise
ratio (SNR) has been achieved by varying the sinusoidal amplitude, and keeping the noise variance constant. The peak has
been detected from the spectrum and its amplitude is evaluated by (16). It is observed that the amplitude estimation error is
very low even in low SNR. Hence, the spectral estimator can be used to quantify the severity of faults accurately under light
loads when the fault components are weak and close to the fundamental.
The performance of the proposed spectral estimator is evaluated in terms of its execution time, accuracy of frequency
detection, and resolvability of closely spaced sinusoids. A comparison between the DFT, MUSIC and the proposed method
for spectral estimation is shown in Fig. 2. It is to be noted that the proposed spectral estimator doesn’t require windowing
to avoid spectral leakage. Whereas, DFT without windowing give rise to false peaks as illustrated by Fig. 2a. This can be detri-
mental for the motor fault diagnosis problem. A Chebyshev window [31] was applied on the data for the DFT to reduce the
false peaks as in Fig. 2b. Also, it was found that the presence of another sinusoid can hamper the frequency estimates of clas-
sical periodogram as shown in Fig. 3a. For this simulation, two sinusoids (50 Hz, 53 Hz) with unity amplitude were used. The
frequency estimate for a sinusoid in the presence of other sinusoids using DFT can also be improved by using a window over
the input samples as shown in Fig. 3b. However, on application of a window, the location error estimates of DFT is found to
be inferior compared to the proposed method. The performance of the proposed method with respect to MUSIC is almost
identical.

8
Original Input Amplitude
6 Estimated Amplitude
Magnitude

Square Error
4

-15 -10 -5 0 5 10 15
SNR (dB)

Fig. 1. Estimation of amplitude, L ¼ 2000; F s ¼ 200 Hz.

 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77 67

1 Proposed method
1 Proposed method
MUSIC MUSIC
DFT Windowed DFT
Amplitude

Amplitude
0.5 0.5
False peaks

0 0
48 49 50 51 52 48 49 50 51 52
Frequency Frequency
(a) Spectrum without windowing for DFT. (b) Spectrum with windowed input data for DFT.
Fig. 2. Comparison of spectral estimators for possible spectral leakages resulting in false peaks. The frequency f 1 ¼ 50 Hz was used for this simulation,
N ¼ 399.

1 Proposed method 1 Proposed method

MUSIC MUSIC
DFT Windowed DFT
Amplitude

Amplitude
0.9 0.9

0.8 0.8

0.7 0.7
49.6 49.8 50 50.2 50.4 49.6 49.8 50 50.2 50.4
Frequency Frequency
(a) Spectrum without windowing for DFT. (b) Spectrum with windowed input data for DFT.
Fig. 3. Comparison of spectral estimators for peak frequency estimation in the presence of another sinusoid. The frequency f 1 ¼ 50 Hz and f 2 ¼ 53 Hz was
used for this simulation, N ¼ 399.

The execution time was evaluated by the mean time required for 1000 trials for each N for the proposed spectral estima-
tor, MUSIC, and DFT. It is evident from Fig. 4 that the proposed spectral estimator has a lower time complexity than MUSIC,
but is slightly slower than DFT.
The accuracy of frequency detection and the robustness of the proposed method have been compared with DFT and
MUSIC. In Fig. 5a, mean square error (MSE) between the input sinusoidal frequency and the location of the peak obtained
from the respective spectral estimator is evaluated by 100 trials for each N. It is inferred that MUSIC and the proposed spec-
tral estimator have slightly higher accuracy for similar data length when compared to DFT with windowed data. In Fig. 5b
MSE is evaluated with different SNR levels for 100 trials each. It is found that the performance of the proposed spectral esti-
mator in a noisy environment is quite robust and is equivalent to that of MUSIC and is marginally better than DFT with win-
dowed data. A single sinusoid of 50.1111 Hz and unity amplitude has been used for both the experiments. F s ¼ 200 Hz was
used. The proposed method is most suitable for fault detection as it is fast, and can estimate the value of fault frequency and
its amplitude accurately without any information about the number of sinusoids. The simulations were conducted with
MATLAB R2014a on a HP Z420 workstation, having 2.80 GHz Intel Xeon CPU E5-1603 processor, 16 GB RAM, and 64-bit Win-
dows 7 operating system.

2.4. Implementation of the proposed spectral estimator

The proposed spectral estimator can be formulated using simple matrix operations for easy implementation. This formu-
lation is suitable for finding the spectral magnitudes in a single specific band or in multiple bands. The formulation involves a
matrix multiplication followed by extraction of diagonal elements as given below
h i
^ ¼ 1 WH R
h b xW i ¼ 1; 2; . . . ; m ð17Þ
L i;i

b x WÞ, contains the spectral peaks in the region where the search is made.
The vector of diagonal elements of the matrix ðWH R
W ¼ ½ w1 w2 ... wm H is known as the search manifold matrix, and wi is defined as
 H
wi ¼ 1 ejxi    ejxi ðL1Þ ; i ¼ 1; 2; . . . ; m

where xi 2 ½xl ; xu  represents the normalized frequency band of m atoms where the search is made and L is the size of the
autocorrelation matrix R b x.
68 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77

0.4
Proposed Method
MUSIC
0.3 DFT

Time (s)
0.2

0.1

0
500 1000 1500 2000
Data length

Fig. 4. Mean execution time for DFT, MUSIC and the proposed method over 1000 trials for each data length (N).

3. Optimal choice of data length with probability of resolution

For accurately resolving two closely spaced sinusoids, with minimum execution time, an optimal length of data (N) needs
to be set for any spectral estimator [50]. The optimum value of N for the proposed spectral estimator is determined statis-
tically with probability of resolution. The decision statistic, c as in (18) decides whether two known frequency components
(f 1 and f 2 ) are resolved with a particular estimator [51].

1
cðf 1 ; f 2 Þ , hðf m Þ  fhðf 1 Þ þ hðf 2 Þg < 0 ð18Þ
2
where f m is the mean of f 1 and f 2 , and hðf i Þ is the magnitude of the f i th component obtained from the spectrum. The prob-
ability of resolution for various data lengths with three levels of SNR are shown in Fig. 6a. Two sinusoids of unity amplitude
and a frequency difference of 1 Hz are considered. For each N, 100 trials have been conducted to find whether the sinusoids
are resolved. It is evident from the figure that N ¼ 400 is sufficient to resolve two sinusoids of equal magnitude having a
frequency difference of 1 Hz and SNR > 0 dB with 100% probability. The required minimum value of N decreases with
decreasing noise.
From Fig. 6b, it is observed that the classical periodogram without windowing has a higher resolution than the proposed
method. This is mainly because the spectral estimation action of the proposed method is based on an average spectrum
obtained from each column of the data matrix. Whereas, the classical periodogram is using a single vector of data-length
that is approximately twice than that of each column of the data-matrix used by the proposed method. The use of window-
ing is necessary for fault detection to avoid spectral leakage and false alarms while using the classical periodogram. However,
using windowed data, the classical periodogram has a lower resolution shown in Fig. 6c when compared with the proposed
method. In this paper, the value of L is fixed at 2000 (N ¼ 3999) to detect weak faults under low-slip.

4. The proposed fault detection scheme

A schematic block diagram of the proposed fault detection algorithm is shown in Fig. 7. The signal conditioning unit esti-
mates and removes the fundamental frequency component from the input stator current. The spectrum of the conditioned
signal is used for finding the slip and the fault specific frequency components. Using the slip information, a fault search band
is formed. Fault specific frequencies and their peak magnitudes are evaluated in this band. The amplitude estimator deter-
mines the amplitude of the sinusoids from the peak magnitude with (16) and sends them to the decision block for threshold
comparisons. The major subsystems other than the spectral estimator involved in the fault detection scheme are discussed
below:

-30 CRLB -30 CRLB

Proposed method Proposed Method
Music MUSIC
MSE (dB)

MSE (dB)

-40 Windowed DFT -40 Windowed DFT

-50
-50
-60
-60
600 800 1000 1200 1400 1600 -5 0 5 10 15 20
Data length (N) Signal to Noise Ratio (dB)

(a) Frequency estimate error for different values of(b) Frequency estimate error for different SNR lev-
N. els, N = 599.
Fig. 5. Evaluation of frequency estimate error for different data length and SNR levels.
 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77 69

Probability of Resolution
1
No noise
SNR = 6.02 dB
SNR = 0 dB
0.5

0
300 350 400 450 500 550
Input data length

(a) Probability of resolution for the proposed method

Probability of Resolution

Probability of Resolution
1 1
No noise
No noise
SNR = 6.02 dB
SNR = 6.02 dB
SNR = 0 dB
0.5 0.5 SNR = 0 dB

0 0
80 100 120 140 160 180 350 400 450 500 550
Input data length Input data length

(b) Probability of resolution for classical periodogram (c) Probability of resolution for classical periodogram
without windowing with windowing

Fig. 6. Probability of resolution versus data length (N) for different levels of SNR (F s ¼ 200 Hz).

fo Fix search band for

EKF-based
Stator
fundamental f mixed
current
frequency estimation fo
input
and removal

Conditioned Spectral estimator Find f mixed and s Fix fault search

signal (for slip calculation) use (18) band

Spectral estimator Detect fault peaks

Decision
(for fault detection) and their magnitudes

Fig. 7. The complete SCIM fault detection scheme.

4.1. Estimation of fundamental frequency and signal conditioning with EKF

The pioneering work of Routray et al. [52] can efficiently estimate and track the fundamental frequency of single-phase
current using EKF. In this paper, the algorithm was modified to estimate the fundamental signal yo ðkÞ in addition to the fun-
damental frequency f o ðkÞ. The estimated fundamental signal is then subtracted from the measured stator current yðkÞ to gen-
erate the conditioned signal eðkÞ which is void of the fundamental component. Fig. 8 illustrates the method used in this
paper. The state vector is given by (19).

^xs ðkÞ ¼ ½2 cosðxÞ y ^ðk  2ÞT

^ðk  1Þ y ð19Þ

For a detailed implementation and analysis of EKF, the readers are referred to [52].
This method, though simple is very effective in eliminating the fundamental component. The spectrum of the stator cur-
rent without input signal conditioning is shown in Fig. 9a. It is observed in Fig. 9b, that, with EKF signal conditioning, the
amplitude of the fundamental component is effectively attenuated. The magnitude was reduced from 3.959E+05 units to
0.03252 units. As a consequence, the spectral leakage due to the fundamental component is reduced, and the detectability
of low amplitude sidebands close to the fundamental frequency is enhanced. The spectra were obtained using the proposed
spectral estimator.

4.2. Estimation of slip

Fault frequency components related to BRB are motor slip dependent. The value of slip is used to form the fault frequency
search bands for the spectral estimator. The slip of the motor can be measured using a speed sensor or can be estimated
using soft sensing techniques like observer-based [53] and slot-harmonic-based [54] methods. The observer based method
requires the acquisition of all the phase voltages and currents. It also requires accurate estimate of certain motor parameters
70 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77

State space formulaon required for EKF: : state vector

: stator current
: fundamental
component of stator current
: stator current without
fundamental

EKF
-
+

Fig. 8. The signal conditioning and fundamental frequency estimator. ^

xs1 ðkÞ is the first element of the estimated state vector.

× 105 0.2
4 Fault component
Fault component

Absolute Magnitude
Absolute Magnitude

Fundamental
3 component 0.15

2 0.1
Suppressed
fundamental
1 0.05

0 0
48 49 50 51 52 48 49 50 51 52
Frequency (Hz) Frequency (Hz)

(a) Unconditioned stator current with undetectable(b) Conditioned stator current with suppressed fun-
fault components damental and clearly visible fault components

Fig. 9. Effect of signal conditioning with EKF (the motor running with 33% load and 0.9% slip under single BRB fault).

that are difficult to determine and may change with time. The slot harmonic based method requires high sampling rate,
which has its own disadvantage. Most SCIMs carry an inherent low-frequency mixed eccentricity component due to machine
saturation [54]. The SCIMs used for this work also exhibited the above component defined by

f mixed ¼ j1  kð1  sÞ=pjf o ð20Þ

This component was present even when all precautions were taken during the fault incorporation and motor assembling
process. Using the lower component of (20), and putting k ¼ 1 (the principal frequency component) and p ¼ 2 (number of
pole-pairs for experimental motor under test), the slip is obtained as

2f mixed  f o
s¼ ð21Þ
fo

The proposed spectral estimator (Section 2) is used for finding f mixed in the frequency band of 0:5f o to 0:6f o , with F s ¼ 200
samples/s. This band has been obtained by evaluating f mixed from (21) with the slip ranging from 0% to 20% for the four-pole
machine. The highest peak obtained in this band is the mixed eccentricity component given by f mixed . Mean of the relative
square error for the speed estimator was found to be 1.908E07. The error was evaluated from full BRB fault dataset (refer
Table 1), which consisted of 72 data for different loads. The error was calculated with respect to the measurement obtained
from a proximity speed sensor.

4.3. Online implementation of the fault detection scheme

The complete fault detection algorithm has been realized on an Intel-based embedded hardware for online and RT exe-
cution. The hardware consists of an ASUS Z87 main board with Intel Core i7 (3.4 GHz) processor, 8 GB DDR3 RAM, and NI PCI
6024E analog interface to acquire analog input signal from the motor. A photograph of the system is shown in Fig. 10a. The

Table 1
Description of different BRB faults used for the experiments.

Fault Healthy BRB Partial BRB Half BRB Full BRB

Drill depth (mm) 0 4 16 34
 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77 71

analog interface has 12-bit resolution and supports up to 200 kHz sampling. DOS based RT kernel known as Simulink Real
Time (SLRT) developed by Mathworks is used as the operating system. A similar implementation for the detection of arc
faults with SLRT can be found in [55]. SLRT is a host-target based system, where the fault detection program is developed
on the host machine with SIMULINK and MATLAB. The host compiles the code and sends the executable to the target com-
puter for RT execution. The NI PCI 6024E of the target acquires a single phase stator current with 200 samples/s for 20 s. The
acquired signal is conditioned using EKF and a data matrix (X) of size 2000 is created. The data matrix is used for spectral
estimation and eventual fault evaluation as discussed in the preceding sections. A screenshot to give a feel of the online sys-
tem console to the reader is shown in Fig. 10b.

5. Experimental results

Experiments were carried out on a 22 kW, four-pole, and three-phase induction motor manufactured by ABB (see A).
Power to the motor was supplied with variable frequency drive (VFD) from ABB (model: ACS550-U1-045A-4). A single rotor
was damaged at different levels as given in Table 1 and shown in Fig. 12 for the experiment. Variable loading was achieved
through rheostatic loads, with a 24 kW separately excited DC generator coupled to the motor as shown in Fig. 11. Clamp-
type current transducer from Fluke (model: i1010s) was used to sense the current signal. Tests were performed in off-
line as well as in online mode. For off-line data analysis, Yokogawa make DL850v oscilloscope was used for recording,
and the analysis was carried out with MATLAB. Single phase stator current was recorded for 30 s with 20 kSamples/s and
was downsampled to 200 samples/s for the offline analysis. For online mode, experiments were carried out with the system
discussed in Section 4.3.

5.1. Results and discussion

Experiments for BRB and its analysis can be classified into two categories. With the first set of analysis, detectability with
partial damage to a single bar with different slip is illustrated. For these experiments, frequency spectra with the proposed
method and MUSIC are compared. To bring uniformity in the curves, each spectrum has been normalized by their fundamen-
tal peak value obtained after EKF-conditioning. In Fig. 13, results for the experiment with 1.9% load and 0.2% slip for healthy
BRB is shown. In Fig. 14, results with 1.9% load and 0.2% slip for partial BRB fault is presented. In this case, the upper and
lower sidebands at ð1  2 sÞf o appears in the spectrum and the difference between healthy and faulty condition is quite com-
prehensive. Fig. 15 illustrates the partial BRB fault with 0.33% slip. Spectral signature with increased fault severity in form of
the half BRB is shown in Fig. 16. The motor has been operating with 0.33% slip on the application of 9% load. It is observed,
that, MUSIC doesn’t give rise to fault peaks, which may be due to improper estimation of the number of sinusoids. Effective-
ness of the method for detection of BRB fault with high load is shown in Fig. 17. The peaks obtained by MUSIC don’t provide
any information about the fault severity. It is also to be noted that the location of fault frequencies obtained by both the spec-
tral estimators are equivalent. This experiment demonstrates the effectiveness and reliability of the proposed spectral esti-
mator when compared with MUSIC for detecting weak BRB faults.
The second set of analysis deals with the effect of different levels of BRB fault on the magnitude spectra, with similar load-
ing in all the cases. The spectrum for this set has also been normalized with the magnitude of the attenuated fundamental
component. It is observed in Fig. 18 that, with increase in the severity of fault in a single bar, the magnitude of the fault com-
ponent also increase. The fault frequency peaks for different cases don’t exactly match due to slight variations in the supply
frequency in each case. For partial and half BRB cases, fault components with k ¼ 4, and k ¼ 5 in (1) are quite prominent for
the 50 Hz supply. Fig. 19 shows the spectrum of the stator current obtained by the proposed spectral estimator for 40 Hz

(a) Photograph of the developed fault detection sys- (b) Online fault detector console
tem

Fig. 10. The online fault detection system and a screenshot taken during a test run.
72 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77

Proximity speed Cooling

sensor arrangement for
DC generator

22 kW Inducon 24 kW DC 3-phase AC
motor Generator tacho generator

Fig. 11. The motor-generator experimental setup.

(a) Half BRB fault (b) Full BRB fault

Fig. 12. Photograph of different levels of BRB fault.

200
Normalized magnitude

150 MUSIC
Proposed Method

100

50

0
48 49 50 51 52
Frequency (Hz)

Fig. 13. Normalized stator current spectrum of healthy motor with 1.9% load and 0.2% slip using MUSIC and the proposed spectral estimator.

200
Normalized magnitude

MUSIC
Proposed Method
150

100 (1-2s)f (1+2s)f

o o

50

0
47 48 49 50 51 52 53
Frequency (Hz)

Fig. 14. Normalized stator current spectrum of the motor with partial BRB running with 1.9% load and 0.2% slip using MUSIC and the proposed spectral
estimator.
 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77 73

200

Normalized magnitude
MUSIC
Proposed Method
150
(1-2s)fo (1+2s)f
o

100

50

0
47 48 49 50 51 52 53
Frequency (Hz)

Fig. 15. Normalized spectrum of stator current with partial BRB running with 13% load and 0.46% slip.

200
Normalized magnitude

MUSIC
Proposed Method
150

100
(1-2s)f
o (1+2s)f
o

50

0
48 49 50 51 52
Frequency (Hz)

Fig. 16. Motor with half BRB with 9% load and 0.33% slip.

200
(1+2s)fo
Normalized magnitude

MUSIC
(1-2s)fo
Proposed Method
150

100

50

0
47 48 49 50 51 52 53
Frequency (Hz)

Fig. 17. Motor with full BRB operating under a high load of 71% and 2.06% slip.

80
(1-2s)fo (1+2s)fo
Normalized Magnitude

(1+10s)fo Healthy
(1-10s)fo Partial
60 (1-8s)fo Half
(1+8s)f
o Full
40

20

0
48 49 50 51 52
Frequency (Hz)

Fig. 18. Spectrum of stator current with different levels of BRB fault for 50 Hz supply (0.26% slip) using the proposed spectral estimator.

supply frequency and 0.33% slip for all the cases. Fig. 20 shows the spectrum with 30 Hz supply frequency and 0.45% slip for
all the fault cases. With 40 Hz and 30 Hz supply, the fault components with k ¼ 1 are most prominent and the peak magni-
tude increase with increasing fault severity.
74 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77

15
(1-2s)fo Healthy

Normalized Magnitude
(1+2s)f
o Half
Full
10

(1-2s)f (1+2s)fo
o

0
38 39 40 41 42
Frequency (Hz)

Fig. 19. Spectrum of stator current with different levels of BRB fault for 40 Hz supply (0.33% slip) using the proposed spectral estimator.

20 (1-2s)f
o (1+2s)fo
Normalized Magnitude

Healthy
15 Half
Full

(1-2s)fo fo (1+2s)f
10 o
(1-10s)f
o
(1+10s)fo
(1-6s)f (1+6s)f
5 o o

0
28 29 30 31 32
Frequency (Hz)

Fig. 20. Spectrum of stator current with different levels of BRB fault for 30 Hz supply (0.45% slip) using the proposed spectral estimator.

1
False alarm
Healthy
0.75
CDF [FY(α )]

Threshold Faulty

0.5

0.25
Missed detection
0
0 1
10 10 102 103
Normalised peak magnitude

Fig. 21. Emperical CDF of normalized peak magnitude (in log scale) for healthy and faulty cases (50 Hz data), with all the faulty cases augmented in a single
vector.

Table 2
Statistics for BRB fault.

Threshold value Missed detection (%) False alarm (%)

20.30 27.38 0.00
17.65 23.50 1.09
14.16 19.35 8.70
10.03 13.00 23.00

From the above results, it is clear, that, the sideband peaks normalized by the fundamental frequency magnitude can dis-
tinguish between the healthy and the faulty cases. The normalized magnitude (Y) is the ratio of the magnitude of the fault
^ ) to the magnitude of the fundamental component (h
specific magnitude (h ^ ) obtained after suppression given by
fault fo

^
h fault
Y¼ ð22Þ
^
h fo
 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77 75

The dispersed nature of the peak magnitudes makes it prudent to employ a probabilistic measure to fix a threshold.
Hence, cumulative distribution function (CDF) of the normalized sidebands has been computed as plotted in Fig. 21. The dis-
tribution function is given by [49, Pg. 61]
F Y ðaÞ ¼ probðY 6 aÞ ð23Þ
where Y is a random variable and a is any number. In the present scenario, Y represents the normalized peak magnitude of a
particular dataset (healthy/faulty) and a is the chosen threshold. An overlap between the CDF magnitudes corresponding to
the faulty and the healthy cases have been observed. Therefore, an optimal threshold is found to represent the acceptable
tradeoff between the missed detection and false alarm rates. Any normalized peak value above a is considered as faulty.
The fraction of data on the right side of the threshold in Fig. 21 for the healthy CDF representing the false-alarm (FA) is given
by
FA ¼ 1  probðY healthy 6 aÞ ð24Þ
while, the fraction of data on the left side of the threshold for faulty CDF representing the missed detection (MD) is given by
MD ¼ probðY faulty 6 aÞ ð25Þ
Y healthy and Y faulty are the normalized peak magnitudes of healthy and faulty data respectively. A range of suitable threshold
from Fig. 21, are given in Table 2. It is observed that a normalized side-band peak magnitude greater than 20.3 can be con-
sidered as faulty without any false alarm.

6. Conclusion

This paper presented the development of an online SCIM fault detection system based on the spectral analysis of a single
phase stator current. The theory of Rayleigh quotient was utilized for designing the proposed spectral estimator. The perfor-
mance of the spectral estimator for detection of frequency components is better than DFT and is similar to that of MUSIC.
Additionally, it can estimate the amplitude of the fault components and is also faster unlike MUSIC. The use of the EKF-
based signal conditioner has enhanced the detection of partial BRB fault under very low load conditions. The system is adap-
tive to changes in supply frequency and loads for a particular frame of data but is unsuitable for the detection of faults under
transient and non-stationary conditions. However, the EKF-based signal conditioner can be used independently for transient
applications for improved detection. The proposed BRB fault detection system was validated by extensive experiments with
different levels of fault under various loading conditions. Empirical CDF was adopted to determine a threshold required for
assessment of the faults. Detection of low-amplitude closely spaced sinusoids makes this method a powerful tool for detec-
tion of weak failure modes such as partial BRB in low-load applications. The online implementation of the system has been
accomplished on a low-cost embedded hardware with Simulink Real Time. With this basic framework, portable solution
with ARM-based systems and system-on-chip which can be incorporated in the VFD is also feasible.

Acknowledgements

The authors highly appreciate the time and effort the anonymous reviewers have exercised in improving the quality of
this manuscript. This research was funded by Research Design and Standards Organization, Indian Railways, Government
of India.

Appendix A. Motor specification

Three-phase squirrel cage induction motor (ABB), delta connection. Rated characteristics: Power = 22 kW, f o = 50 Hz,
voltage = 415 V, current = 40.5 A, rated speed = 1460 rpm, No. of pole pairs = 2.

References

[1] S. Nandi, H.A. Toliyat, X. Li, Condition monitoring and fault diagnosis of electrical motors – a review, IEEE Trans. Energy Convers. 20 (4) (2005) 719–729.
[2] H. Henao, C. Bruzzese, E. Strangas, R. Pusca, J. Estima, M. Riera-guasp, S.H. Kia, Trends in fault diagnosis for electrical machines: a review of diagnostic
techniques, IEEE Ind. Electron. Mag. 8 (June) (2014) 31–42.
[3] S.H. Kia, H. Henao, G.-A. Capolino, Zoom-MUSIC frequency estimation method for three-phase induction machine fault detection, in: 31st Annu. Conf.
of IEEE Ind. Electron. Soc., 2005 (IECON 2005), IEEE, 2005, p. 6.
[4] G. Didier, E. Ternisien, O. Caspary, H. Razik, Fault detection of broken rotor bars in induction motor using a global fault index, IEEE Ind. Appl. 42 (1)
(2006) 79–88.
[5] A. Naha, A.K. Samanta, A. Routray, A.K. Deb, A method for detecting half-broken rotor bar in lightly loaded induction motors using current, IEEE Trans.
Instrum. Meas. PP (99) (2016) 1–12, http://dx.doi.org/10.1109/TIM.2016.2540941.
[6] A. Garcia-perez, R.D.J. Romero-troncoso, E. Cabal-yepez, R.A. Osornio-rios, The application of high-resolution spectral analysis for identifying multiple
combined faults in induction motors, IEEE Trans. Ind. Electron. 58 (5) (2011) 2002–2010, http://dx.doi.org/10.1109/TIE.2010.2051398.
[7] Y.-H. Kim, Y.-W. Youn, D.-H. Hwang, J.-H. Sun, D.-S. Kang, High-resolution parameter estimation method to identify broken rotor bar faults in induction
motors, IEEE Trans. Ind. Electron. 60 (9) (2013) 4103–4117, http://dx.doi.org/10.1109/TIE.2012.2227912.
[8] B. Xu, L. Sun, L. Xu, G. Xu, An ESPRIT-SAA-based detection method for broken rotor bar fault in induction motors, IEEE Trans. Energy Convers. 27 (3)
(2012) 654–660.
76 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77

[9] V. Choqueuse, M. Benbouzid, et al, Induction machine faults detection using stator current parametric spectral estimation, Mech. Syst. Signal Process.
52 (2015) 447–464.
[10] G. Bouleux, Oblique projection pre-processing and tls application for diagnosing rotor bar defects by improving power spectrum estimation, Mech.
Syst. Signal Process. 41 (1) (2013) 301–312.
[11] D.-J. Kim, H.-J. Kim, J.-P. Hong, C.-J. Park, Estimation of acoustic noise and vibration in an induction machine considering rotor eccentricity, IEEE Trans.
Magn. 50 (2) (2014) 857–860, http://dx.doi.org/10.1109/TMAG.2013.2285391.
[12] J. Seshadrinath, B. Singh, B.K. Panigrahi, Investigation of vibration signatures for multiple fault diagnosis in variable frequency drives using complex
wavelets, IEEE Trans. Power Electron. 29 (2) (2014) 936–945, http://dx.doi.org/10.1109/TPEL.2013.2257869.
[13] P.A. Delgado-arredondo, D. Morinigo-sotelo, R.A. Osornio-rios, J.G. Avina-cervantes, H. Rostro-gonzalez, R.D.J. Romero-troncoso, Methodology for fault
detection in induction motors via sound and vibration signals, Mech. Syst. Signal Process. 83 (2017) 568–589.
[14] A. Sapena-bano, M. Pineda-sanchez, R. Puche-panadero, J. Perez-cruz, J. Roger-folch, M. Riera-guasp, J. Martinez-roman, Harmonic order tracking
analysis: a novel method for fault diagnosis in induction machines, IEEE Trans. Energy Convers. 30 (3) (2015) 833–841.
[15] M.Y. Kaikaa, M. Hadjami, Effects of the simultaneous presence of static eccentricity and broken rotor bars on the stator current of induction machine,
IEEE Trans. Ind. Electron. 61 (5) (2014) 2452–2463.
[16] J. Pons-Llinares, J.A. Antonino-daviu, M. Riera-guasp, S.B. Lee, T.-J. Kang, C. Yang, Advanced induction motor rotor fault diagnosis via continuous and
discrete time-frequency tools, IEEE Trans. Ind. Electron. 62 (3) (2015) 1791–1802.
[17] B. Xu, L. Sun, H. Ren, A new criterion for the quantification of broken rotor bars in induction motors, IEEE Trans. Energy Convers. 25 (1) (2010) 100–106.
[18] A. Ceban, R. Pusca, R. Romary, Study of rotor faults in induction motors using external magnetic field analysis, IEEE Trans. Ind. Electron. 59 (5) (2012)
2082–2093, http://dx.doi.org/10.1109/TIE.2011.2163285.
[19] M.F. Cabanas, F. Pedrayes, M.G. Melero, C.H.R. García, J.M. Cano, G.A. Orcajo, J.G. Norniella, Unambiguous detection of broken bars in asynchronous
motors by means of a flux measurement-based procedure, IEEE Trans. Instrum. Meas. 60 (3) (2011) 891–899.
[20] M. Nemec, K. Drobnic, D. Nedeljkovic, F. Rastko, V. Ambrozic, R. Fiser, Detection of broken bars in induction motor through the analysis of supply
voltage modulation, IEEE Trans. Ind. Electron. 57 (8) (2010) 2879–2888.
[21] M.A. Cruz, An active reactive power method for the diagnosis of rotor faults in three-phase induction motors operating under time-varying load
conditions, IEEE Trans. Energy Convers. 27 (1) (2012) 71–84.
[22] M. Drif, A.J.M. Cardoso, Discriminating the simultaneous occurrence of three-phase induction motor rotor faults and mechanical load oscillations by
the instantaneous active and reactive power media signature analyses, IEEE Trans. Ind. Electron. 59 (3) (2012) 1630–1639.
[23] X. Ying, Performance evaluation and thermal fields analysis of induction motor with broken rotor bars located at different relative positions, IEEE
Trans. Magn. 46 (5) (2010) 1243–1250.
[24] M.J. Picazo-Rodenas, R. Royo, J. Antonino-Daviu, J. Roger-Folch, Use of infrared thermography for computation of heating curves and preliminary
failure detection in induction motors, in: 2012 XXth Int. Conf. on Elect. Mach., Ieee, 2012, pp. 525–531, http://dx.doi.org/10.1109/
ICElMach.2012.6349920.
[25] A. Khezzar, M.Y. Kaikaa, M.E.K. Oumaamar, M. Boucherma, H. Razik, On the use of slot harmonics as a potential indicator of rotor bar breakage in the
induction machine, IEEE Trans. Ind. Electron. 56 (11) (2009) 4592–4605.
[26] K.N. Gyftakis, D.V. Spyropoulos, J.C. Kappatou, E.D. Mitronikas, A novel approach for broken bar fault diagnosis in induction motors through torque
monitoring, IEEE Trans. Energy Convers. 28 (2) (2013) 267–277, http://dx.doi.org/10.1109/TEC.2013.2240683.
[27] T.-j. Kang, J. Kim, S.B. Lee, C. Yung, Experimental evaluation of low-voltage offline testing for induction motor rotor fault diagnostics, IEEE Trans. Ind.
Appl. 51 (2) (2015) 1375–1384.
[28] C. Bruzzese, Analysis and application of particular current signatures (symptoms) for cage monitoring in nonsinusoidally fed motors with high
rejection to drive load, inertia, and frequency variations, IEEE Trans. Ind. Electron. 55 (12) (2008) 4137–4155.
[29] J. Kim, S. Shin, S.B. Lee, K.N. Gyftakis, M. Drif, A.J.M. Cardoso, Power spectrum-based detection of induction motor rotor faults for immunity to false
alarms, IEEE Trans. Energy Convers. (2015) 1–10.
[30] C. Yang, T.-j. Kang, S.B. Lee, J.-y. Yoo, A. Bellini, L. Zarri, F. Filippetti, Screening of false induction motor fault alarms produced by axial air ducts based on
the space harmonic-induced current components, IEEE Trans. Ind. Electron. 0046 (c) (2014) 1, http://dx.doi.org/10.1109/TIE.2014.2331027.
[31] B. Ayhan, H.J. Trussell, M.-y. Chow, M.-H. Song, On the use of a lower sampling rate for broken rotor bar detection with DTFT and AR-based spectrum
methods, IEEE Trans. Ind. Electron. 55 (3) (2008) 1421–1434.
[32] J. Chen, Z. Li, J. Pan, G. Chen, Y. Zi, J. Yuan, B. Chen, Z. He, Wavelet transform based on inner product in fault diagnosis of rotating machinery: a review,
Mech. Syst. Signal Process. 70 (2016) 1–35.
[33] J. Pons-llinares, M. Riera-guasp, J.A. Antonino-daviu, T.G. Habetler, Pursuing optimal electric machines transient diagnosis: the adaptive slope
transform, Mech. Syst. Signal Process. 80 (2016) 553–569.
[34] P. Shi, Z. Chen, Y. Vagapov, Z. Zouaoui, A new diagnosis of broken rotor bar fault extent in three phase squirrel cage induction motor, Mech. Syst. Signal
Process. 42 (1–2) (2014) 388–403.
[35] J. Faiz, V. Ghorbanian, B.M. Ebrahimi, EMD-based analysis of industrial induction motors with broken rotor bars for identification of operating point at
different supply modes, IEEE Trans. Ind. Informat. 10 (2) (2014) 957–966.
[36] V. Ghorbanian, J. Faiz, A survey on tand frequency characteristics of induction motors with broken rotor bars in line-start and inverter-fed modes,
Mech. Syst. Signal Process. 54 (2015) 427–456.
[37] A. Bellini, F. Filippetti, G. Franceschini, C. Tassoni, G.B. Kliman, Quantitative evaluation of induction motor broken bars by means of electrical signature
analysis, IEEE Trans. Ind. Appl. 37 (5) (2001) 1248–1255.
[38] Y. Trachi, E. Elbouchikhi, V. Choqueuse, M.E.H. Benbouzid, Induction machines fault detection based on subspace spectral estimation, IEEE Trans. Ind.
Electron. 63 (9) (2016) 5641–5651.
[39] B. Halder, T. Kailath, Efficient estimation of closely spaced sinusoidal frequencies using subspace-based methods, IEEE Signal Process. Lett. 4 (2) (1997)
49–51.
[40] P. Stoica, H. Li, J. Li, Amplitude estimation of sinusoidal signals: survey, new results, and an application, IEEE Trans. Signal Process. 48 (2) (2000) 338–
352.
[41] C. Concari, G. Franceschini, C. Tassoni, Differential diagnosis based on multivariable monitoring to assess induction machine rotor conditions, IEEE
Trans. Ind. Electron. 55 (12) (2008) 4156–4166.
[42] R. Puche-Panadero, M. Pineda-Sanchez, M. Riera-Guasp, J. Roger-Folch, E. Hurtado-Perez, J. Perez-Cruz, Improved resolution of the MCSA method via
hilbert transform, enabling the diagnosis of rotor asymmetries at very low slip, IEEE Trans. Energy Convers. 24 (1) (2009) 52–59.
[43] B. Xu, L. Sun, L. Xu, G. Xu, Improvement of the hilbert method via ESPRIT for detecting rotor fault in induction motors at low slip, IEEE Trans. Energy
Convers. 28 (1) (2013) 225–233.
[44] C.G. Dias, I.E. Chabu, Spectral analysis using a hall effect sensor for diagnosing broken bars in large induction motors, IEEE Trans. Instrum. Meas. 63 (12)
(2014) 2890–2902.
[45] M. Pineda-Sanchez, R. Puche-Panadero, M. Riera-Guasp, J. Perez-Cruz, J. Roger-Folch, J. Pons-Llinares, V. Climente-Alarcon, J.A. Antonino-Daviu,
Application of the Teager–Kaiser energy operator to the fault diagnosis of induction motors, IEEE Trans. Energy Convers. 28 (4) (2013) 1036–1044.
[46] A. Sapena-Bao, M. Pineda-Sanchez, R. Puche-Panadero, J. Martinez-Roman, Kanovi, Low-cost diagnosis of rotor asymmetries in induction machines
working at a very low slip using the reduced envelope of the stator current, IEEE Trans. Energy Convers. 30 (4) (2015) 1409–1419, http://dx.doi.org/
10.1109/TEC.2015.2445216.
[47] V. Climente-Alarcon, J.A. Antonino-Daviu, E.G. Strangas, M. Riera-Guasp, Rotor-bar breakage mechanism and prognosis in an induction motor, IEEE
Trans. Ind. Electron. 62 (3) (2015) 1814–1825.
 A.K. Samanta et al. / Mechanical Systems and Signal Processing 98 (2018) 63–77 77

[48] B.N. Datta, Numerical Linear Algebra and Applications, Siam, 2010.
[49] M.H. Hayes, Statistical Digital Signal Processing and Modeling, John Wiley & Sons, 2010.
[50] A. Naha, A.K. Samanta, A. Routray, A.K. Deb, Determining autocorrelation matrix size and sampling frequency for MUSIC algorithm, IEEE Signal Process.
Lett. 22 (8) (2015) 1016–1020, http://dx.doi.org/10.1109/LSP.2014.2366638.
[51] Q.T. Zhang, Probability of resolution of the MUSIC algorithm, IEEE Trans. Signal Process. 43 (4) (1995) 978–987, http://dx.doi.org/10.1109/78.376849.
[52] A. Routray, A.K. Pradhan, K.P. Rao, A novel Kalman filter for frequency estimation of distorted signals in power systems, IEEE Trans. Instrum. Meas. 51
(3) (2002) 469–479, http://dx.doi.org/10.1109/TIM.2002.1017717.
[53] M.S. Zaky, M.M. Khater, S.S. Shokralla, H. Yasin, Wide-speed-range estimation with online parameter identification schemes of sensorless induction
motor drives, IEEE Trans. Ind. Electron. 56 (5) (2009) 1699–1707, http://dx.doi.org/10.1109/TIE.2008.2009519.
[54] S. Nandi, S. Ahmed, H.A. Toliyat, Detection of rotor slot and other eccentricity related harmonics in a three phase induction motor with different rotor
cages, IEEE Trans. Energy Convers. 16 (3) (2001) 253–260.
[55] A. Mukherjee, A. Routray, A.K. Samanta, Method for on-line detection of arcing in low voltage distribution systems, IEEE Trans. Power Del. PP (99)
(2015), http://dx.doi.org/10.1109/TPWRD.2015.2392385, pp. 1-1.