Mechanical Systems and Signal Processing: El Houssin El Bouchikhi, Vincent Choqueuse, Mohamed Benbouzid

Mechanical Systems and Signal Processing ] (]]]]) ]]]–]]]
Contents lists available at ScienceDirect
Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp
Induction machine faults detection using stator current

parametric spectral estimation
El Houssin El Bouchikhi a,b, Vincent Choqueuse a, Mohamed Benbouzid a,n
a
University of Brest, EA 4325 LBMS, Rue de Kergoat, CS 93837, 29238 Brest, France
b
ISEN, EA 4325 LBMS, 20, Rue Cuirassé Bretagne, 29200 Brest, France
a r t i c l e i n f o abstract
Article history: Current spectrum analysis is a proven technique for fault diagnosis in electrical machines.
Received 25 June 2013 Current spectral estimation is usually performed using classical techniques such as
Received in revised form periodogram (FFT) or its extensions. However, these techniques have several drawbacks
7 June 2014
since their frequency resolution is limited and additional post-processing algorithms are
Accepted 26 June 2014
required to extract a relevant fault detection criterion. Therefore, this paper proposes a
new parametric spectral estimator that fully exploits the faults sensitive frequencies.
Keywords: The proposed technique is based on the maximum likelihood estimator (MLE) and offers
Induction machine high-resolution capabilities. Based on this approach, a fault criterion is derived for
Faults detection
detecting several fault types.
Bearing faults
The proposed technique is assessed using simulation signals, issued from a coupled
Stator current
Spectral estimation electromagnetic circuits approach-based simulation tool for mechanical unbalance and
Maximum likelihood estimator electrical asymmetry faults detection. It is afterward validated using experiments on a
0.75-kW induction machine test bed for the particular case of bearing faults.
& 2014 Elsevier Ltd. All rights reserved.
1. Introduction
Condition monitoring is of high concern in industrial applications since it minimizes the downtime and improves the
reliability, availability, safety and productivity of these systems. For electrical motors and generators, faults detection is
usually performed by vibration monitoring, temperature measurements, oil monitoring, flux monitoring and current
analysis [1,2]. Among these various techniques, current analysis has several advantages since it is a noninvasive technique
that avoids the use of extra sensors [3–7]. Moreover, the electrical signals (for instance, the stator current) are usually
available and inexpensive to measure.
Stator currents processing-based faults diagnosis of induction machine has received intense research interest for several
decades [8–10]. Moreover, the International Standard “ISO FDIS 20958” dealing with “Condition monitoring and diagnostics
of machine systems - Electrical signature analysis of three-phase induction motors” sets out guidelines for the online
techniques recommended for the purposes of condition monitoring and diagnostics of machines, based on electrical
signature analysis. Hence, many studies have shown that fault monitoring could be performed by supervising the current
spectrum. In particular, it has been demonstrated that faults introduce additional spectral components in the stator current
around the supply frequency [11,12]. For a faulty machine, the frequency location of these components is given by f k ðΩÞ,
n
Corresponding author. Tel.: þ33 2 98 01 80 07; fax: þ 33 2 98 01 66 43.
E-mail address: Mohamed.Benbouzid@univ-brest.fr (M. Benbouzid).
http://dx.doi.org/10.1016/j.ymssp.2014.06.015
0888-3270/& 2014 Elsevier Ltd. All rights reserved.
Please cite this article as: E.H. El Bouchikhi, et al., Induction machine faults detection using stator current parametric
spectral estimation, Mech. Syst. Signal Process. (2014), http://dx.doi.org/10.1016/j.ymssp.2014.06.015i
2 E.H. El Bouchikhi et al. / Mechanical Systems and Signal Processing ] (]]]]) ]]]–]]]
Nomenclature DSP digital signal processing

MCMFT maximum covariance frequency tracking
DFT discrete Fourier transform MDL minimum description length
FFT fast Fourier transform PSD power spectral density
ESPRIT estimation of signal parameters via rotational SNR signal to noise ratio
invariance techniques fs stator supply frequency
MLE maximum likelihood estimator s per-unit slip
MUSIC multiple signal classification p pole pair number
DFT discrete Fourier transform
where fk corresponds to the kth component (k A Z), and Ω is a set of parameters that must be estimated in order to
determine the induction machine health condition. These frequencies are associated with several mechanical and/or
electrical faults such as air-gap eccentricity, bearing failures or broken rotor bars. For instance, In the case of broken rotor
bars, the fault signature is given by

1s
f k ðΩÞ ¼ f s k 7s ð1Þ
p
where Ω ¼ fs; f s g.
Traditional current-based techniques for fault detection monitor the stator current spectrum and, more precisely, the
fault characteristic frequencies [13]. In steady-state conditions, techniques based on conventional PSD estimators have been
employed [14]. These techniques can be classified into two categories: non-parametric and parametric methods [15].
Non-parametric methods include the periodogram, which is usually implemented using the FFT, and its extensions.
The classical periodogram have been applied for fault detection in [11,12]. The main drawback of this technique relies on its
performance. Indeed, even though the FFT is computationally efficient, it suffers from a poor frequency resolution (inversely
proportional to the measurement time interval) and leakage effects (the energy of the main lobe leaks into sidelobes) due to
windowing. Moreover, frequencies in the Discrete Fourier Transform (DFT) are spaced at intervals of F s =N where Fs is the
sampling frequency and N is the length of the input time series. Attempting to estimate the amplitude of a sinusoid with a
frequency that does not correspond to a DFT bin can result in an inaccurate estimate. Therefore, several DFT interpolation
techniques have been proposed in order to enhance frequency accuracy such as analytical leakage compensation [16], zero-
padding [15], phase interpolation estimator [17] and many others [18–20]. Furthermore, the periodogram method provides
reasonably high resolution for sufficiently long data lengths, but it is a poor spectral estimator because its variance is high
and does not decrease with increasing data length. In addition of that, it is often advantageous to use a window function
other than a rectangular (natural) one and which has a Fourier transform with faster decaying side-lobes than sinc-type
function. The multiplication of the data with a particular window function can reduce the sidelobe amplitudes but increases
the width of the mainlobe. Common window functions are the Bartlett window, Hamming window, Hanning window, and
others [21]. The rectangular window leads to a narrow mainlobe but the highest sidelobes whereas the Hanning window
reduces strongly the sidelobe amplitudes but leads to the largest mainlobe [22,23]. Furthermore, the high variance of
the periodogram method motivates the development of modified methods that have lower variance, at the cost of
reduced resolution. Several modified methods have been introduced such as the Bartlett [24] and the Welch techniques
[24,25]. Moreover, the so-called Zoom-FFT (ZFFT) technique [26] has been introduced to improve the frequency accuracy in
a specified frequency range without increasing the computational complexity. Nevertheless, the periodogram and its
extensions suffer from a low frequency resolution, which is defined as the ability to distinguish two closely spaced
frequency components. In [27], demodulation technique based on Hilbert transform was used to improve the frequency
resolution of MCSA method for rotor asymmetries detection without concern about the signal nature (multi-component
signal).
If an a priori signal model is assumed, parametric methods can be employed to improve the frequency resolution. These
techniques are generally called high-resolution methods and include two sub-classes: the linear prediction methods and the
subspace techniques. The linear prediction class contains several algorithms like the Prony and Pisarenko methods. The use
of these methods for fault detection in electrical drives has been investigated in [28,29]. The subspace class includes
the MUSIC and ESPRIT approaches. Applications for induction machine faults diagnosis are available in [2,30–32]. In [33], the
MUSIC algorithm and a zooming method were combined to reduce the computational cost of the spectral estimator.
However these techniques are generally computationally intensive and lead to suboptimal estimators of the PSD. Moreover,
their performance decrease significantly if the noise level increases.
In addition to the aforementioned techniques, many faults detection procedures based on statistical analysis of the
current signal have been proposed such as MCMFT [34] and adaptive statistical Time-Frequency Methods [35] without
presenting any fault detection criteria for an automatic fault diagnosis.
The above review emphasizes the compromise between frequency accuracy, frequency resolution, statistical perfor-
mance and computational cost of spectrum analysis techniques for fault detection in induction machine. Furthermore, these
previous techniques are general and do not exploit the particular structure of the stator current frequency components
E.H. El Bouchikhi et al. / Mechanical Systems and Signal Processing ] (]]]]) ]]]–]]] 3
Fig. 1. Classical spectrum analysis techniques on experimental data. (a) Stator current waveform (Faulty case). (b) Periodogram-based spectrum (Hanning
window). (c) MUSIC-based pseudospectrum. (d) ESPRIT-based spectrum.
Fig. 2. Block diagram of the proposed technique. Symbol C corresponds to the fault detection criterion and v is a vector containing the estimated
parameters of the monitored stator current.
given in the existing works dealing with induction machine failure diagnosis. In order to illustrate the limitations of the
classical spectrum analysis techniques, Fig. 1 represents the current waveform and the spectrum using the periodogram,
MUSIC and ESPRIT techniques in the case of a healthy and faulty motor (bearing fault). The simulation parameters for the
periodogram are sampling frequency F s ¼ 1000 Hz, N ¼ 2000, a Hanning window, and a zero padding with N pad ¼ 4096.
The number of sinusoids for MUSIC and ESPRIT is setted equal to L¼6 and M þ 1 by M þ 1 correlation matrix is computed
with M ¼ N=8. It is obvious that these techniques provide information about the presence of fault (sidebands-related fault)
but no direct fault detection criterion is provided. Consequently, additional post-processing algorithms are required
to determine the fault-related frequencies (mainly done manually) and to extract a fault detection criterion. Once
the frequency spectrum is obtained and stored, empirical formula are used in order to reveal frequency signatures in the
spectrum within various frequency ranges depending on the fault to be diagnosed. If predicted frequency patterns are
present in the spectrum, a fault classification is achieved.
In this paper, in order to overcome the above mentioned limitations, we propose a parametric PSD estimator based on
the MLE combined with minimum description length principle (MDL). In [36], a technique based on the MLE has been
proposed for mechanical fault detection in induction machine. Nevertheless, this technique is limited to particular faults and
its extension to other faults is far from trivial. Moreover, its implementation requires the use of sophisticated optimization
tool. The fault detection algorithm proposed in this paper is composed of the two steps described in Fig. 2. First, the fault
related frequencies number and the PSD are estimated by using the MLE combined with information theoretic criteria
(MDL). This estimation algorithm implies a 3-D optimization problem if we consider the supply frequency unknown. Then, a
simple fault criterion is computed from the amplitude estimate of the fault characteristic frequencies. The main advantage
of the proposed method is its ability to automatically extract the fault frequency signature and to automatically compute a
fault indicator without the need of an expert to interpret the current spectrum [37].
The major contributions of this paper are threefold:
We propose a new model order and spectral estimation technique aiming at detecting the induction machine fault
frequency signatures. This technique is asymptotically optimal and exhibits high-resolution capabilities.
We demonstrate the appropriateness of the approach on major faults in induction machine. In fact, it makes a potential
failure identifiable and quantifiable.
We prove the effectiveness of the technique on simulated and experimental data.
The remaining parts of this paper are organized as follows.
Section 2 introduces the signal model for healthy and faulty machines.
Section 3 describes the proposed PSD estimator.
Section 4 presents the fault detection criterion.
Section 5 deals with the application of the proposed method for air-gap eccentricity, broken rotor bars, and bearing faults
detection, which are the most frequent faults in electric machines according to [38,39].
Fig. 3. Distribution of known motor failures for the petroleum and chemical industries [39]. (a) Breakdown of failed components by distribution.
(b) Extrapolated distribution of failure by motor components.
Table 1
Induction machine faults signatures [11,12].
Induction machine state Frequency signature Parameters Ω
Bearing damage jf s 7 kf o j k ¼ 1; 2; 3; … Ω ¼ ff s ; f 0 g

Broken rotor bars 1 s k Ω ¼ ff s ; sg
fs k 7s ¼ 1; 3; 5; 7; 11; 13…
p p

Air gap eccentricity 1s k ¼ 1; 2; 3; … Ω ¼ ff s ; sg
f s 17 k
p

Load oscillation 1s k ¼ 1; 2; 3 ,... Ω ¼ ff s ; sg
f s 17 k
p
2. Stator current signal model
2.1. Faults effect on stator current spectrum
The induction machine is subjected to various failures that affect mainly three components: the stator, the rotor and/or
the bearings (Fig. 3a). Recent papers dealing with induction machine faults distribution [39] have shown that bearings
(69%), stator windings (21%), rotor (7%), and shaft/coupling (3%) are the most failing components (Fig. 3b). Most of the recent
researches on induction machine faults detection has been directed toward electrical monitoring with emphasis on stator
current supervision. In particular, the current spectrum is analyzed to extract the frequency components introduced by the
fault. A summary of induction machine faults effects on stator current is presented in Table 1.
In the Table, fs corresponds to the supply frequency, s is the per unit slip, and p is the pole-pair number. Symbol fo
corresponds to one of the characteristic vibration frequencies introduced by bearing fault which depends on the bearing
dimensions and mechanical rotor frequency [40]. In [41], it has been demonstrated that depending on the physical
phenomena (eccentricity or torque variations) introduced by bearing faults, the corresponding frequency signatures are
different. Moreover, these signatures depend on the defected components. Concerning the load oscillation, the model
presented holds only if the load oscillates at the rotating frequency [41,42]. In fact, in the case of load oscillations, the stator
current is modulated by the shaft rotational speed, belt pass frequency, compression frequency, blades rotational frequency,
meshing frequency, resonance frequency of the control loop, etc. Finally, concerning the eccentricity fault, we only focus on
the monitoring of the current at the fundamental sidebands of the supply frequency given in [11]. In fact, since space
harmonics are usually of smaller amplitude, higher order terms of these harmonics can be supposed to be less affected by
the fault than the fundamental frequency. In general way, the fault detection methods monitor the behavior of the current at
the sidebands of the slot frequencies. These sidebands associated with eccentricity are given in [9].
The frequencies given by Table 1 are used in the faulty induction machine stator current model described in the following
section. When a fault occurs, the amplitude at these frequencies increases and reveal abnormal operating conditions.
2.2. Induction machine stator current signal model
The stator current signal model is based on the following assumptions:
H1 : The received signal is modeled as a sum of 2L þ1 sine waves embedded in noise. 2L corresponds to the number of the
sidebands around the supply frequency introduced by the fault (whose amplitudes rise when a fault occurs).
Fig. 4. Theoretical PSD for L ¼2 [15].
H2 : The noise is white, Gaussian with zero-mean and variance σ2.

H3 : The signal frequency content obeys to the particular structures as given by Table 1.
H4 : The matrix AðΩÞ has full column rank 2 L þ 1, i.e. rankðAðΩÞÞ ¼ 2 L þ 1. This is verified for N 4 2 L þ 1.
In practice, H1 may not be verified since the induction machine stator current may contain space- and time-harmonics.
However, some of these harmonics can be eliminated by filtering the stator current signal in order to focus on the fault-
affected frequency bandwidth. Moreover, we must note that H1 requires the knowledge of L. In the present work, we
propose a technique to estimate it from the stator current signal. Concerning the assumption H2 , it is not particularly
restrictive. In fact, if the noise process is not white and has unknown spectral shape, then accurate frequency estimates can
still be found if we estimate the sinusoids using the non-linear least squares (NLS) [15, Chapter 4, Introduction].
Under the assumption H1 H3 , the stator-current samples x½n can be expressed as
L

n
x½n ¼ ∑ ak cos 2π f k ðΩÞ þ ϕk þ b½n ð2Þ
k ¼ L Fs
where b½n corresponds to the noise sample. Symbols f k ðΩÞ, ak and ϕk correspond to the frequency, the amplitude and the
phase of the kth frequency component, respectively. Symbol Fs corresponds to the sampling rate. Note that the particular
case where k¼0 corresponds to the fundamental frequency component.
The PSD is defined as the Discrete Time Fourier Transform of the covariance function of x½n [15]. Under the assumption
H2 and that the initial phases ϕk, are independent random variables uniformly distributed on ½ ππ , the theoretical PSD of
x½n is given by Fig. 4 [15]. In practice, the PSD is unknown and must be estimated from N samples. Using a matrix notation,
x½n (n ¼ 0; …; N 1) can be expressed as
x ¼ AðΩÞv þ b ð3Þ
where
x ¼ ½x½0; …; x½N 1T is a N 1 column vector containing the stator current samples,
b ¼ ½b½0; …; b½N 1T is a N 1 column vector containing the noise samples,
v is a 2ð2L þ 1Þ 1 column vector containing the amplitudes and phases of the characteristic fault frequencies. This vector
is given by
v ¼ ½a L cos ðϕ L Þ…aL cos ðϕL Þ; a L sin ðϕ L Þ… aL sin ðϕL ÞT ð4Þ
AðΩÞ is a N 2ð2L þ 1Þ matrix whose elements are given by
AðΩÞ ¼ ½z L …zL ; y L …yL ð5Þ
where

1 N1 T 1 N1 T
zk ¼ 1 cos 2π f k ðΩÞ … cos 2π f k ðΩÞ yk ¼ 0 sin 2π f k ðΩÞ … sin 2π f k ðΩÞ ð6Þ
Fs Fs Fs Fs
Non-parametric estimators estimate the PSD from x without using any a priori knowledge about the signal. Departing
from this approach, we propose a parametric estimator that exploits the signal model in (3) and the fault related frequency
presented in Table 1. In this context, the computation of the current spectrum from stator current samples x is treated as a
statistical estimation problem. The main object is to estimate the parameters Ω in Table 1 and a relevant fault detection
criterion.
3. Parametric PSD estimation
In this section, we present the fault detection scheme based on the MLE. The MLE is an asymptotically optimal estimator
since it has the asymptotic properties of being unbiased, achieving the Cramer–Rao lower bound, and having a Gaussian
Probability Density Function [43]. Application for PSD estimation has been investigated for harmonic and non-harmonic
signal models in [44] and [45], respectively. Unfortunately, the MLE has a higher computational cost since it requires the
optimization of a cost-function in high dimensional space. In the case of induction machine fault detection, we demonstrate
that the MLE is computationally efficient since it leads to a 2-D optimization problem that is easy to implement. Besides, an
efficient implementation of the MLE requires the knowledge of frequency components numbers which is called model order
estimation problem. In order to estimate the model order (fundamental frequency and 2 L sidebands produced by fault,
where L has to be estimated), we propose to combine the MLE with an order-dependent penalty term based on the MDL
principle [46]. It must be emphasized that the sidebands number (2 L) estimation is of great interest since it contribute to
inform us about the fault existence. Moreover, if the order is not estimated (chosen) correctly, the fault characteristic
frequency may erroneously be estimated at, for example, half or double of the true value. In addition to the exact MLE,
we propose an FFT-based approximate approach that leads to significant computational complexity reduction.
3.1. Proposed estimator
The MLE is used in order to estimate v and Ω. Then a penalty term is applied to the MLE cost function in order to
estimate the model order i.e. L.
3.1.1. Estimate of v and Ω

The ML estimator of v, and Ω is given by
b; Ω
fv b g ¼ arg max logðpðx; v; ΩÞÞ ð7Þ
v;Ω
where pðx; v; ΩÞ is the probability density function (pdf) of x which is given by

1 1
pðx; v; ΩÞ ¼ exp 2 ðx AðΩÞvÞT ðx AðΩÞvÞ ð8Þ
ð2πσ 2 Þ 2
N
2σ
where ð:ÞT denotes the matrix transpose, and a noise is considered white Gaussian noise with zero mean and variance equal
to σ2 i.e. b½n N ð0; σ 2 Þ. The ML estimates of Ω and v are obtained by maximizing the pdf with respect to the unknown
parameters. The maximization in (7) is equivalent to the minimization of the following cost function:
Lðx; v; ΩÞ ¼ ðx AðΩÞvÞT ðx AðΩÞvÞ ð9Þ

Differentiating Lðx; v; ΩÞ with respect to v and setting the derivative equal to 0 leads to the ML estimate of v denoted v
b
(see Appendix A for the proof).
b ¼ A† ðΩÞx
v ð10Þ
†
where A ðΩÞ is the pseudo-inverse of AðΩÞ, i.e.
A† ðΩÞ ¼ ðAT ðΩÞAðΩÞÞ 1 AT ðΩÞ ð11Þ

1
and where ð:Þ corresponds to the matrix inverse.
The ML estimate of Ω is obtained by minimizing Lðx; v
b ; ΩÞ with respect to Ω. By replacing v by vb in (9), we obtain (see
Appendix B)
fΩ
b g ¼ arg maxJ ðΩÞ ð12Þ
Ω
where
J ðΩÞ ¼ xT AðΩÞA† ðΩÞx ð13Þ
3.1.2. Estimate of model order L

The estimation of L enhances the performance of the ML estimates of Ω. The information theoretic criteria are used to
estimate L [47]. In fact, a penalty term based on MDL principle is applied to the MLE cost function in order to estimate L.
Fig. 5. Exact and approximate cost function (f s ¼ 50 Hz, f c ¼ 1 Hz, Lopt ¼ 1, F s ¼ 1 kHz and SNR ¼ 50 dB). (a) J ðf c Þ (N ¼ 500). (b) J ðf c Þ (zoom). (c) J a ðf c Þ
(N¼ 2000). (d) J a ðf c Þ (zoom).
Hence, the estimation of L can be performed by maximizing the penalized ML estimate of Ω [46] as follows:
fΩ
b ;b b; σ
b ; Ω; LÞ þ cðg; NÞÞ
2
Lg ¼ arg minð 2 log pðx; v ð14Þ
Ω;L
where
The noise variance estimate is given by

1
σb 2 ¼ J x AðΩ
b Þv
bJ ð15Þ
N
The penalty function cðg; NÞ depends on the number of free parameters g and the number of data samples N. Under the
assumption that the number of the components is equal to 2L þ 1, g ¼ 4L þ 4. The criterion information rule used within
this paper is the minimum description length (also called the Bayesian Information Criterion Rule) [46] and is given by
cðg; NÞ ¼ g logðNÞ ð16Þ
A straightforward computation leads to the following optimization problem (see Appendix C for the proof):

cðg; NÞ
fΩ Lg ¼ arg max xT x J ðΩÞ exp
b ;b ð17Þ
Ω;L N
Finally, the PSD estimate of the stator current is composed of two steps: (a) the estimates of Ω, and L are obtained from (17)
and (b) the vector v containing the amplitude and the phase of the fault characteristic frequencies is estimated by replacing Ω, and
L with its estimates in (10). Because of its statistical properties, one should note that the Maximum Likelihood estimate remains
the most accurate method for PSD estimation even in those cases where the noise is colored [15]. In particular, this estimator
overcomes the frequency resolution limitation of the periodogram. Furthermore, as opposed to other techniques, the proposed
approach exploits the faults characteristic frequencies to improve the accuracy of the PSD estimate.
About the implementation, the main difficulty relies on the optimization problem in (17). As the maximum cannot be
found analytically, numerical method should be used to estimate Ω and L. In our context, the cost function is only composed
of three parameters, which implies a maximization in a 3-D space. The search space is relatively limited since the variation
range of Ω and L are approximately known. For these reasons, we propose to perform the maximization of (17) with a grid-
search algorithm. This algorithm evaluates the cost function at the vertices of a rectangular grid, and chooses the vertex with
the highest value. It should be noted that the maximization step could be computationally demanding since it requires the
construction and the inversion of a large matrix for each vertex of the grid.
For grid connected induction machine, the fundamental frequency can be assumed to be known. Consequently, the
optimization problem in (17) reduces to 2-D problem. Fig. 5a illustrates the maximization step for a synthetic signal with
f k ðΩÞ ¼ f s 7kf c , f s ¼ 50 Hz, f c ¼ 1 Hz, L¼1, and SNR ¼50 dB. The acquisition time and the sampling frequency are equal to
1 s and F s ¼ 600 Hz, respectively. The grid search algorithm evaluates the cost function for fc ranging from 0 Hz to 2 Hz with
a step size of 0.001 Hz and L varying from 0 to 4. By looking at Fig. 5a, it can be observed that the cost function is maximized
for the true values of fc i.e. 1 Hz and L¼1. Concerning the computational complexity, the evaluation of the cost function
s
requires 26.7 s on a HP ProBook PC at 2.2 GHz, using Matlabsimulink .
3.2. Link with FFT
The computational complexity of the PSD estimator can be reduced when the number of samples, N, goes to infinity.
Indeed, by using the following limit (see, e.g., [43, Example 7.16]).
2 T
lim A ðΩÞAðΩÞ ¼ IN ð18Þ
N-1N
where IN corresponds to the N N identity matrix. The cost function can be approximated as (see Appendix D)
2
J a ðΩÞ ¼ lim J ðΩÞ ¼ xT AðΩÞAT ðΩÞx
N-1 N

2
L
1 N1
¼ 2 ∑
pffiffiffiffi ∑ x½ne 2jπ ðf k ðΩÞÞn=F s
ð19Þ
k ¼ L
N n ¼ 0
where the last equality comes from the fact that x½n A R. The last equation can be expressed according to the Discrete
Fourier Transform (DFT) of x½n. Indeed,
L
J a ðΩÞ ¼ 2 ∑ jDFT x ½f k ðΩÞ=F s j2 ð20Þ
k ¼ L
where DFT x ½f is the DFT computed at frequency f, i.e.

1 N1
DFT x f ¼ pffiffiffiffi ∑ x½ne 2jπ fn ð21Þ
Nn¼0
Finally the approximate ML estimate of Ω is simply obtained by replacing J ðΩÞ with J a ðΩÞ in (12). It must be stressed that
an approximate MLE can be obtained if f k ðΩÞ=F s is not close to 0 and 1=2.
Similarly, the approximate ML estimator of the vector v is then computed using (22):
2
b ¼ AT ðΩÞx
v ð22Þ
N
This approximate approach can be extended in order to estimate the model order L as follows:

cðg; NÞ
fΩ Lg ¼ arg max xT x J a ðΩÞ exp
b ;b ð23Þ
Ω;L N
Eqs. (23) and (20) show that the approximate cost function is reduced to a sum of DFT bins. This makes the approximate
approach attractive for the following reasons: (a) Most DSP-boards include functions for DFT computation (b) the DFT can be
efficiently computed using the FFT. However, it should be stressed that the accuracy of the approximation highly depends on
the signal length N. In particular, the approximation in (18) is no longer valid for short signals. In this case, the DFT of the
stator current signal exhibits sidelobes which affect the frequency resolution. The sidelobes can mask components close in
frequency and then lead to false conclusions. Moreover, the sidelobes could be interpreted as fault characteristic frequency
and then lead to false alarm. This may be overcame by choosing different windows (Hanning, Blackman) in order to
attenuate sidelobes [23]. Moreover, interpolating the DFT by zero padding the data can often increase the frequency
accuracy and thus improve the amplitude estimate [15]. The approximated method is then limited by the FFT algorithm
resolution: the parameters are estimated correctly as long as the observed signal length N=F s is large enough compared to
the inverse of the smallest frequency difference between two neighboring poles of the signal, i.e. (24).
1
Nb ð24Þ
mink1 a k2 jf k2 f k1 j
Fig. 5c illustrates the cost function for synthesized signal. One can notice that the exact cost function in Fig. 5a and its
approximation in Fig. 5c have roughly the same shape and differ only in low frequencies and while jf s 7 kf c j-0. Indeed,
these two shapes differs when the signal model contains closely spaced frequencies. For closely spaced frequencies, the
resolution limitation of the DFT leads to wrong results. In particular in Fig. 5c, one can notice the spurious peak located at
f c ¼ 0 Hz. When using the approximate approach, spurious peaks must be removed to obtain accurate estimate of fc.
This can be done by excluding small values of fc from the grid search. Despite this drawback, the approximate approach is
quite attractive since it leads to a drastic computational reduction. For example, the evaluation of the approximate cost
function in Fig. 5c requires only 4.2 s, while the evaluation of the exact one in Fig. 5a requires 26.7 s.
4. Fault detection criterion
In this section, we propose a fault detection criterion based on the amplitude of the fault characteristic frequencies.
As the PSD estimate directly exploits the fault signature, we show that the fault detection criterion can be easily derived
from the vector vb.
4.1. Proposed criterion
In order to successfully perform fault detection, a fault criterion is needed to measure the machine state. As the
information about the fault severity is carried out by ak (k a 0), we propose to compute the sum of the squares of the
Fig. 6. Fault detection algorithm.
(normalized) amplitude of the fault characteristic frequencies. This criterion is expressed mathematically as
!
L a2k
C¼ ∑ 2
ð25Þ
k ¼ L;k a 0 a0
This criterion exhibits the two following desirable properties.
Property 1. For healthy machines, the criterion is equal to 0 i.e.

C¼0 ð26Þ
Property 2. The proposed criterion is invariant to scale. Indeed, let us denote Cx the fault criterion relative to the stator current
x½n. It can be demonstrated that the fault criterion Cαx , relative to the signal αx½n (α 40), is equal to
Cαx ¼ Cx ð27Þ
The proposed criterion in (25) depends on the amplitudes ak (k ¼ L; …; L). Once the PSD is estimated, ak can be
extracted from the vector v. However, using the structure of v in (4), it can be shown that C can be obtained directly from v
without computing ak. Indeed, the fault criterion in (25) can be expressed under the following matrix form (see the
Appendix E for proof):
vT v
C¼ 1 ð28Þ
vT Mv
where M is a ð4L þ 2Þ ð4L þ2Þ matrix which is given by
" #
EL þ 1;L þ 1 0
ð29Þ
0 EL þ 1;L þ 1
and where Eu;v is the ð2L þ 1Þ ð2L þ 1Þ elementary matrix which is 1 in the uth row and vth column and is zero elsewhere.
In practice, one should note that v is unknown and must be replaced by its estimate v b in (28) to compute C.
4.2. Algorithm summary
The fault detection algorithm is summed up in Fig. 6. This algorithm can be implemented for real-time monitoring of an
induction machine. Compared to other PSD-based monitoring technique (Periodogram, MUSIC, ESPRIT), the proposed
approach is quite attractive since the vector v directly conveys information about the characteristic frequencies.
5. Application to induction machine faults detection
s
The proposed approach has been implemented in MatlabSimulink on a HP ProBook PC at 2.2 GHz with 2 Go of RAM.
Its performance is evaluated with simulated stator currents for eccentricity and broken rotor bars faults and experimental
data for bearing failures.
5.1. Simulation results
This section deals with the performance of the proposed approaches for induction machine faults detection. The
induction machine modeling is briefly presented, then the proposed approaches are used to detect eccentricity and broken
rotor bars faults.
5.1.1. Induction machine modeling

An induction machine model for healthy and faulty machines has been developed based on the coupled magnetic circuits
theory [13,48,49]. This method is based on the following assumptions:
Negligible saturation.
Negligible eddy current, friction and windage losses.
Insulated rotor bars.
The induction machine stator consists of three phase concentric winding. Each winding is considered as a separated coil.
The cage rotor has n bars which can be described as n identical and equally spaced rotor loops; each loop is formed by two
adjacent rotor bars connected by the end-ring portions.
The rotor cage equivalent circuit showing the rotor loops and different currents is displayed in Fig. 7.
The stator currents are obtained by solving the system of differential equations in (30):
8
> d d
>
> dt I ¼ L 1
R þ Ω L I þ L 1V
>
>
> dθ m
>
<d
1 d f 1
Ω ¼ IT L I Ω ΓC ð30Þ
>
>
>
dt 2J d θ m J J
>
>
>
> d
: θm ¼ Ω
dt
where
" #
Vs Is
V¼ ; I¼
0 Ir
" # " #
Rs 0 Lss Lsr
R¼ ; L¼
0 Rr Lrs Lrr
Symbols Vs , Is and Ir correspond to the stator voltage vector, the stator current vector and the rotor current vector,
respectively. Inductances Lss , Lrr , Lrs and Lsr correspond to the self and mutual inductances between stator windings and
Fig. 7. Rotor bars equivalent circuit [14].
rotor windings. Resistances Rs and Rr refer to the stator and cage resistances and Γ C , Ω and θm correspond to the load
torque, the rotor mechanical speed and the rotor angular position, respectively. Finally, J is the rotating masses inertia and f
is the viscous friction coefficient.
In (30), the machine inductances L are carried out using the air-gap magnetic energy, which is determined by means of
s
the actual geometry and winding layout of the machine. A Matlabsimulink based tool for faulty induction machines has
been developed to generate faults database and therefore to allow testing the proposed faults detection approach.
In these simulations dynamic eccentricities are introduced to emulate a bearing fault. It has been demonstrated that
single-point bearings faults have an effect over the machine eccentricity and/or load variations [40–42,50]. In fact, bearing
fault will induce mechanical eccentricities, but also load-torque variations. Hence, in the carried-out simulations, bearing
faults are emulated by generating only one sort of physical phenomena: rotating eccentricities at bearing characteristic fault
frequency fc. These eccentricities leads to periodical changes in the induction machine inductances [40]. Moreover, broken
rotor bars are emulated by suppressing the corresponding broken rotor bar.
Therefore, a 4 kW induction machine operating under nominal load condition have been simulated. Three machines have
been considered: a healthy machine, a faulty one affected by a 10% (static, dynamic and mixed) eccentricity, and a faulty one
affected by broken rotor bars. Simulations were performed with a supply frequency equal to f s ¼ 50 Hz and nominal speed
equal to Ωm ¼ 1425 rpm. The stator current signals have been recorded during 1 s with a 1 kHz sampling rate.
5.1.2. Eccentricity fault detection results

The PSD of the stator current using the periodogram, MUSIC algorithm, and the proposed approach are depicted in Fig. 8
for the same simulation parameters as the ones used in Fig. 1.
The algorithm shown in Fig. 6 is used to extract a fault detection criterion. The eccentricity fault characteristic frequency
is used, i.e. f k ðΩÞ ¼ f s 7 kf c where f c ¼ ðð1 sÞ=pÞf s . The grid search algorithm has been implemented using a fine search: fc
ranges from 0 Hz to 100 Hz with a step size of 0.01 Hz. Table 2 summarizes the simulation results. This table presents the
estimated fault frequency fc, the sidebands number L, and the fault detection criterion C. The fault detection criterion C is
computed using (28).
The analysis of Table 2 allows concluding that the two proposed approaches permit to detect the eccentricity fault.
Indeed in the presence of eccentricity faults, the fault criterion increases significantly. Therefore, a simple threshold-based
decision can distinguish between the healthy and the faulty cases. However, it is obvious that the exact method gives more
reliable results since the criterion is null for healthy case which confirms the theoretical study. The FFT-based approach
tends to over-estimate L which leads to values of fc equal to half the true value given by the exact method. The last
ascertainment is due to sidelobes (due to sampling) present in the signal FFT which are interpreted as sidebands
(characteristic fault frequency).
5.1.3. Broken rotor bars fault detection results

Broken rotor bar is one of the electrical faults that is difficult to detect since the squirrel cage current can not be acquired.
The PSD of the stator current using the periodogram, MUSIC algorithm, and the proposed approach are depicted in Fig. 9 for
the same simulation parameters as the ones used in Fig. 1.
The broken rotor bars characteristic frequency is used in the signal model i.e. f k ðΩÞ ¼ f k ðf s ; sÞ ¼ f s ½kðð1 sÞ=pÞ 7 s.
Computer simulations have been performed for assessment of operating features of the proposed fault detection scheme.
Table 3 gives simulation results for 1–3 broken rotor bars for both the exact and FFT-based techniques. The broken bars
are adjacent. The criterion has been evaluated for different fault degrees. It can be noticed that the fault criterion varies in
proportion to the number of broken rotor bars. The same conclusions can be drawn from the broken rotor bars results. It is
worthy to notice that in the exact approach case, the estimation of k=p leads to informations about faults presence. This is
not the case for FFT-based approach since the algorithm interprets sidelobes as sidebands which may lead to false alarm.
Fig. 8. Stator current PSD for a faulty motor (mixed eccentricity). (a) Periodogram (Hanning Window). (b) MUSIC pseudospectrum (L ¼16). (c) Proposed
approach.
Table 2
Eccentricity fault simulation results. (a) Exact PSD estimator. (b) Approximated PSD estimator.
State b
L b
f c (Hz) C ( 10 3)
(a)
Healthy 0 0 0
Static eccentricity 1 24.46 1.1
Dynamic eccentricity 1 48.87 5.3
Mixed eccentricity 2 24.43 8
(b)
Healthy 4 10 0.94
Static eccentricity 2 12.3 1.5
Dynamic eccentricity 2 12.22 5.5
Mixed eccentricity 4 12.21 8.4
Fig. 9. PSD of the stator current with 3 broken rotor bars. (a) Periodogram (Hanning window). (b) MUSIC pseudospectrum (L ¼16). (c) Proposed approach.
Table 3
Broken rotor bars simulation results. (a) Exact PSD estimator. (b) Approximated PSD estimator.
State b
L d
k=p b
s (%) C( 10 3)
(a)
Healthy 0 0 0 0
1 broken bar 3 1, 3, 5 5.8 4.6
2 broken bars 4 1, 3, 5, 7 6.2 24.4
3 broken bars 5 1, 3, 5, 7, 11 6.6 49.7
(b)
Healthy 1 1 3.13 1.8
1 broken bar 3 1, 3, 5 5.85 4.3
2 broken bars 4 1, 3, 5, 7 6.16 22.6
3 broken bars 5 1, 3, 5, 7, 11 6.57 45.2
5.2. Experimental results
This section reports on the performance of the proposed approaches. First, the test rig is presented, then the proposed
techniques was applied off-line using Matlab for bearing faults detection in a conventional induction machine.
5.2.1. Test rig

A conventional 0.75 kW induction machine drive test rig is used in order to test the proposed parametric spectral
estimation fault detection approach (Fig. 11). The rated data of the used induction machine are given in Appendix F.
The test rig mechanical part is composed by a synchronous and an induction machine. The induction machine is fed by
the synchronous generator in order to eliminate time-harmonics. Indeed, this will automatically eliminate supply harmonics
and therefore allow focusing only on bearing faults effect on the stator current.
The induction machine has two 6204.2 ZR type bearings (single row and deep groove ball bearings) with the following
parameters: outside diameter is 47 mm, inside diameter is 20 mm, and pitch diameter D is 31.85 mm. A bearing has n ¼8
balls with an approximate diameter d of 12 mm and a contact angle of α ¼ 0○ (Fig. 10). Based on these dimensions, the
theoretical values of the different bearing fault frequencies are given in Table 4.
Bearing failure is one of the foremost causes of breakdowns in rotating machinery, resulting in costly downtime (Fig. 3b)
[39]. One of the key issues in bearing prognostics is to detect the defect at its incipient stage and alert the operator before it
develops into catastrophic failure. Hence, bearing faults are obtained by simply drilling holes in different parts [51].
Fig. 10. Bearing structure with main dimensions ((α) outer race deterioration, (β) inner race deterioration, (γ) cage deterioration, δ ball deterioration). (a)
Geometry of a rolling-element bearing. (b) Artificially deteriorated bearing.
Table 4
Theoretical values of bearing faults frequencies.
Bearing fault Fault related frequency Theoretical values (Hz)

Cage defect f d [14.4, 27.077]
f c ¼ r 1 cos ðαÞ
2 D
!
Ball defect D d
2 [105.5, 119.47]
f bd ¼ f 1 2 cos 2 ðαÞ
d r D

Inner raceway nf d [154.04, 255.1]
f id ¼ r 1þ cos ðαÞ
2 D

Outer raceway nf d [115.5, 216.6]
f od ¼ r 1 cos ðαÞ
2 D
Fig. 11. Test rig scheme.
In order to study the load influence, the induction machine was operated with various load levels ranging from 0 W to
400 W.
The measured quantities for off-line bearing fault detection were the line-currents. For all the experiments, the stator
fundamental frequency was equal to f s ¼ 50 Hz. All the signals were acquired at a 10 kHz sampling frequency by a data
s
acquisition card and processed using Matlabsimulink . As the information relative to the bearing faults is mostly
contained in the low frequency content, these signals were down-sampled at a 600 Hz sampling rate.
5.2.2. Fault detection results

In presence of bearing faults, it has been shown in [40] that the fault characteristic frequencies are given by:
f k ðΩÞ ¼ jf s 7kf c j (k A Z). In [41], it has been demonstrated that depending on the bearing fault effect on the induction
Fig. 12. Performance of the proposed approaches on experimental data. (a) Exact PSD estimator: values of the fault detection criterion C for 1 – healthy
case, 2 – inner raceway fault, 3 – cage fault, and 4 – ball fault. (b) Approximate PSD estimator: values of the fault detection criterion C for 1 – healthy case, 2
– inner raceway fault, 3 – cage fault, and 4 – ball fault.
machine behavior (eccentricity or torque variations) the bearing fault-related frequencies in the stator current spectrum is
different. In this section, we consider only the case where the localized single-point bearing defects leads to the torque
variations. Hence, the model in [40] has been considered.
The algorithm in Fig. 6 is again used to extract a fault detection criterion. Similarly to the simulation configuration, the
number L has been estimated in order to decide whether the induction machine is operating with healthy bearings or
damaged once. If L ¼0 then the bearings are safe, otherwise the bearing is damaged and criterion in (28) is computed in
order to measure the fault severity and then take decision. Preliminary tests have shown that most of the characteristic fault
frequencies are greater than f s ¼ 50 Hz which means that most of the frequency components related to fault are located at
the right-side of the supply frequency fs due to the absolute value of the frequencies introduced by faults (Table 1).
Therefore, the proposed approach has been used to extract these right-side components. This simplification is not
equivalent to minimization of time-harmonics in power supply since these frequencies are different from multiples of the
fundamental frequency f s ¼ 50 Hz.
Experimental results are reported in Fig. 12a and b for the exact and approximate approaches, respectively. These figures
display the criterion C for healthy and faulty machines with various load levels.
For the exact algorithm, in the case of healthy machine, it is obvious that L¼0 which implies that the criterion is equal to 0.
From this bar chart, it could be observed that the proposed criterion significantly increases for each faulty machine, regardless of
the fault type and load level. Therefore, a simple estimation of L gives an indication of the existence of fault. However, in the case
of approximated approach based on the FFT, the estimation of L cannot be a reliable way to distinguish faulty from healthy case
since L is different from 0 even if the machine is operating correctly. That is why, the criterion computation is mandatory and
threshold-based fault detector must be defined in order to distinguish between healthy and faulty machines.
By comparing the two techniques, we can note that the exact approach is more reliable than the approximated method
which is less time consuming. However, it must be emphasized that the approximate approach has practical advantages
since it is based on the DFT (easiness of implementation, fast computation with FFT). The two figures also show that the
criterion C decreases with the load level; there is only one exception to this rule for the healthy machine in the case of
approximated algorithm. This could be explained by the fact that the load tends to hide the faults effect on the stator current
[52]. This is clearly illustrated in Fig. 12b when the load is equal to 400 W. This results obviously confirms the effectiveness of
the proposed technique over the FFT-based one.
The proposed technique allows detecting several induction machine faults. In order to characterize the fault (defected
component, fault severity, etc.) another step is required which has not been addressed within this paper. However, the proposed
technique may be used as an input for a fault classifier since it allows the extraction of the fault frequency signatures. Once the
fault sensitive frequencies are extracted, the signatures given in the literature may be used to discriminate several faults.
6. Conclusions
This paper has proposed a statistical-based approach for fault detection in induction machines. The proposed approach is
composed of two steps: (a) the estimation of the PSD with a new parametric technique and (b) the computation of a fault
detection criterion.
The proposed PSD estimator has been computed using the maximum likelihood estimation approach. As opposed to
non-parametric PSD estimators, the proposed technique exploits the fault frequency signatures in order to improve the
performance of the fault detection criterion. As a result, the proposed estimator has better frequency–resolution and
frequency–accuracy than other techniques such as the periodogram. When the number of samples goes to infinity, it has
also been demonstrated that the proposed PSD estimator can be efficiently implemented using the Discrete Fourier
Transform. However, this approximated method does not perform as well as the exact method for short signals. Concerning
the fault detection criterion, we have proposed a criterion based on the amplitude of the fault-related frequencies. This
criterion is theoretically equal to zero for healthy machine and increases for faulty once.
The proposed approach was successfully tested on simulations with eccentricity and broken rotor bars faults and
experimental test rig with various bearing faults and load conditions. Simulation and experimental results have
corroborated the efficiency of the proposed method, regardless of the fault type in contrary to the FFT-based approach.
Furthermore, these results have suggested that the estimation of L is very interesting since it allows to make a direct and fast
first idea about the machine state.
Further investigations are required in order to study the effect of the induction machine faults over the time- and space-
harmonics of the stator current. Furthermore, the impact of these harmonics over the reliability of the proposed technique
should be highlighted.
Appendix A. Estimation of v
By expanding (9), we obtain
Lðx; v; ΩÞ ¼ ðx AðΩÞvÞT ðx AðΩÞvÞ ¼ ðxT vT AT ðΩÞÞðx AðΩÞvÞ ¼ xT x 2vT AT ðΩÞx þ vT AT ðΩÞAðΩÞv ð31Þ
The derivative of Lðx; v; ΩÞ with respect to v is equal to (see Ref. [53])
∂Lðx; v; ΩÞ
¼ 2AT ðΩÞx þ2AT ðΩÞAðΩÞv ¼ 2AT ðΩÞðx AðΩÞvÞ ð32Þ
∂v
b:
Setting the above derivative to zero, we obtain the ML estimator of v, which is denoted v

∂Lðx; v; ΩÞ

v ¼ b ¼ 0 ) x ¼ AðΩÞv
b ð33Þ
∂v v
Finally, we obtain the following ML estimator of v:

b ¼ A† ðΩÞx
v ð34Þ
†
where A ðΩÞ is the pseudo-inverse of AðΩÞ i.e.
A† ðΩÞ ¼ ðAT ðΩÞAðΩÞÞ 1 AT ðΩÞ ð35Þ
Appendix B. Estimation of Ω
The ML estimate of f k ðΩÞ is obtained by minimizing Lðx; v

b ; ΩÞ with respect to Ω. By replacing v by vb in (9), we obtain
b ; ΩÞ ¼ ðx AðΩÞv
Lðx; v b Þ ¼ ðx AðΩÞA ðΩÞxÞ ðx AðΩÞA ðΩÞxÞ ¼ xT ðIN AðΩÞA† ðΩÞÞx
b Þ ðx AðΩÞv T † T †
ð36Þ
where the last equality has been obtained using (35). Neglecting the terms that do not depend on Ω, it can be shown that
the ML estimate of Ω is given by
fΩ
b g ¼ arg maxJ ðΩÞ ð37Þ
Ω
where
J ðΩÞ ¼ xT AðΩÞA† ðΩÞx ð38Þ
Appendix C. Model order estimation
Departing from Eq. (14) we have
b; σ
CF ¼ 2 log pðx; v b ; Ω; LÞ þcðg; NÞ
2
ð39Þ
1h i
CF ¼ 2 log ð2π σ
b 2 Þ 2 2 ðx AðΩÞv
b ÞT x AðΩÞv
N
b þcðg; N Þ ð40Þ
σb
1h i
2π σ ðx AðΩÞv
b ÞT x AðΩÞv
2
CF ¼ N log b b þ cðg; N Þ ð41Þ
σb 2
" #
1 ðx AðΩÞv
b ÞT ðx AðΩÞv
bÞ cðg; NÞ
2π σ
2
CF ¼ log b þ ð42Þ
σb 2 N N
Using (15) the cost function is equivalent to

cðg; NÞ
CF1 ¼ log 2π σ b2 þ ð43Þ
N
Since the exp function is strictly increasing, applying the exp on (43) gives

cðg; NÞ
CF2 ¼ 2π σb 2 exp ð44Þ
N
Replacing (15) in (44) gives the equation given in (17).
Appendix D. Link with FFT proof
2
J a ðΩÞ ¼ xT AðΩÞAT ðΩÞx ð45Þ
N
2
J a ðΩÞ ¼ ‖AT x‖F ð46Þ
N
2
J a ðΩÞ ¼ ‖½z L …zL ; y L …yL T ½x½0 x½1 … x½N 1T ‖F ð47Þ
N
2 3
∑N 1 x½n cos 2π f ðΩÞ n
n¼0 L Fs
6 7
6 ⋮ 7
6 7
6 7
6 ∑N 1 x½n cos 2π f L ðΩÞ n 7

2 6 n ¼ 0 F s 7
J a ðΩÞ ¼
6
6
7
7 ð48Þ
N 6 ∑ N 1
x½n sin 2π f L ðΩÞ F s 7
n
6 n ¼ 0 7
6 ⋮ 7
6 7
4 5
∑N 1 x½n sin 2π f ðΩÞ n
n¼0 L Fs
F
N 1 2 N 1 2
2 L n n
J a ðΩÞ ¼ ∑ ∑ x½n cos 2π f k ðΩÞ þ ∑ x½n sin 2π f k ðΩÞ ð49Þ
N k ¼ L n¼0 Fs n¼0 Fs

2
2 L
N 1
j2π f k ðΩÞFns
J a ðΩÞ ¼ ∑
∑ x½ne
ð50Þ
N k ¼ L
n ¼ 0
Appendix E
By using (4), we obtain

L L L
vT v ¼ ∑ a2k cos 2 ðϕk Þ þ ∑ a2k sin 2 ðϕk Þ ¼ ∑ a2k ð51Þ
k ¼ L k ¼ L k ¼ L
By using the structure of M, we also get

vT Mv ¼ ½0⋯0a0 cos ðϕ0 Þ0⋯0 a0 sin ðϕ0 Þ0⋯0v
¼ a20 cos 2 ðϕ0 Þ þ a20 sin 2 ðϕ0 Þ
¼ a20 ð52Þ
These two equations lead to the following result:
!
vT v L a2k a2
1 ¼ ∑ 2 02 ð53Þ
vT Mv a
k ¼ L 0 a0
!
vT v L a2k
1 ¼ ∑ ¼C ð54Þ
vT Mv k ¼ L;k a 0 a20
Appendix F
RATED DATA OF THE TESTED INDUCTION MACHINE
0.75 kW, 50 Hz, 220/380 V, 3.4/1.95 A, 2780 rpm, p ¼1
References
[1] S. Nandi, H.A. Toliyat, X. Li, Condition monitoring and fault diagnosis of electrical motors – a review, IEEE Trans. Energy Convers. 20 (December (4))
(2005) 719–729.
[2] A. Garcia-Perez, R. de Jesus Romero-Troncoso, E. Cabal-Yepez, R. Osornio-Rios, The application of high-resolution spectral analysis for identifying
multiple combined faults in induction motors, IEEE Trans. Ind. Electron. 58 (May (5)) (2011) 2002–2010.
[3] M. Seera, C.P. Lim, D. Ishak, H. Singh, Fault detection and diagnosis of induction motors using motor current signature analysis and a hybrid FMM-cart
model, IEEE Trans. Neural Netw. Learn. Syst. 23 (January (1)) (2012) 97–108.
[4] V. Choqueuse, M.E.H. Benbouzid, Y. Amirat, S. Turri, Diagnosis of three-phase electrical machines using multidimensional demodulation techniques,
IEEE Trans. Ind. Electron. 59 (April (4)) (2011) 2014–2023.
[5] X. Gong, W. Qiao, Imbalance fault detection of direct-drive wind turbines using generator current signals, IEEE Trans. Energy Convers. 27 (June (2))
(2012) 468–476.
[6] B. Trajin, Analyse et traitement de grandeurs électriques pour la détection et le diagnostic de défauts mécaniques dans les entraînements asynchrones.
Application à la surveillance des roulements à billes, Ph.D. Dissertation, Institut National Polytechnique de Toulouse-INPT, 2009.
[7] M.E.H. Benbouzid, G.B. Kliman, What stator current processing based technique to use for induction motor rotor faults diagnosis? IEEE Trans. Energy
Convers. 18 (June (2)) (2003) 238–244.
[8] P.J. Tavner, J. Penman, Condition Monitoring of Electrical Machines, vol. 1, Research Studies Press, Letchworth, 1987.
[9] J. Cameron, W. Thomson, A. Dow, Vibration and current monitoring for detecting airgap eccentricity in large induction motors, Electric Power Appl. IEE
Proc. B 133 (May (3)) (1986) 155–163.
[10] I. Mitchel, Study of Time and Frequency Domain Analysis Techniques for Rotor Cage Fault Detection (Ph.D. Dissertation), Thesis Gordon's Institute of
Technology Aberdeen, 1986.
[11] M.E.H. Benbouzid, A review of induction motors signature analysis as a medium for faults detection, IEEE Trans. Ind. Electron. 47 (October (5)) (2000)
984–993.
[12] A. Bellini, F. Filippetti, C. Tassoni, G.A. Capolino, Advances in diagnostic techniques for induction machines, IEEE Trans. Ind. Electron. 55 (December
(12)) (2008) 4109–4126.
[13] E.H. El Bouchikhi, V. Choqueuse, M.E.H. Benbouzid, J. Charpentier, G. Barakat, A comparative study of time-frequency representations for fault
detection in wind turbine, in: Proceedings of the 2011 IEEE IECON, Melbourne, Australia, November 2011, pp. 3584–3589.
[14] E.H. El Bouchikhi, V. Choqueuse, M.E.H. Benbouzid, Current frequency spectral subtraction and its contribution to induction machines' bearings
condition monitoring, IEEE Trans. Energy Convers. 28 (March (1)) (2013) 135–144.
[15] P. Stoica, R.L. Moses, Introduction to Spectral Analysis, Prentice-Hall, New Jersey, 1997.
[16] H. Renders, J. Schoukens, G. Vilain, High-accuracy spectrum analysis of sampled discrete frequency signals by analytical leakage compensation, IEEE
Trans. Instrum. Meas 33 (4) (1984) 287–292.
[17] D.R.A. McMahon, R.F. Barrett, An efficient method for the estimation of the frequency of a single tone in noise from the phases of discrete Fourier
transforms, IEEE Trans. Signal Process. 11 (1986) 169–177.
[18] B.G. Quinn, Estimation of frequency, amplitude, and phase from the DFT of a time series, IEEE Trans. Signal Process. 45 (3) (1997) 814–817.
[19] B.G. Quinn, Estimating frequency by interpolation using Fourier coefficients, IEEE Trans. Signal Process. 42 (5) (1994) 1264–1268.
[20] J. Schoukens, R. Pintelon, H. Van Hamme, The interpolated fast Fourier transform: a comparative study, IEEE Trans. Instrum. Meas. 41 (2) (1992)
226–232.
[21] F.J. Harris, On the use of windows for harmonic analysis with the discrete Fourier transform, Proc. IEEE 66 (1) (1978) 51–83.
[22] R.B. Randall, Vibration-Based Condition Monitoring: Industrial, Aerospace and Automotive Applications, John Wiley & Sons, 2011.
[23] A. Oppenheim, R. Schafer, W. Padgett, Discrete-Time Signal Processing, 3rd ed. Prentice Hall, 2009.
[24] G. Didier, E. Ternisien, O. Caspary, H. Razik, Fault detection of broken rotor bars in induction motor using a global fault index, IEEE Trans. Ind. Appl. 42
(January/February (1)) (2006) 79–88.
[25] M.E.H. Benbouzid, M. Vieira, C. Theys, Induction motors' faults detection and localization using stator current advanced signal processing techniques,
IEEE Trans. Power Electron. 14 (January (1)) (1999) 14–22.
[26] A. Bellini, A. Yazidi, F. Filippetti, C. Rossi, G. Capolino, High frequency resolution techniques for rotor fault detection of induction machines, IEEE Trans.
Ind. Electron. 55 (December (12)) (2008) 4200–4209.
[27] R. Puche-Panadero, M. Pineda-Sanchez, M. Riera-Guasp, J. Roger-Folch, E. Hurtado-Pearez, 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 (March (1)) (2009) 52–59.
[28] Z. Leonowicz, T. Lobos, J. Rezmer, Advanced spectrum estimation methods for signal analysis in power electronics, IEEE Trans. Ind. Electron. 50 (June
(3)) (2003) 514–519.
[29] J. Stack, T. Habetler, R. Harley, Bearing fault detection via autoregressive stator current modeling, IEEE Trans. Ind. Appl. 40 (May/June (3)) (2004)
740–747.
[30] E.H. El Bouchikhi, V. Choqueuse, M.E.H. Benbouzid, Induction machine fault detection enhancement using a stator current high resolution spectrum,
in: Proceedings of the 2012 IEEE IECON, Montréal, Canada, October 2012, pp. 3893–3898.
[31] A. Bracale, G. Carpinelli, L. Piegari, P. Tricoli, A high resolution method for on line diagnosis of induction motors faults, in: Proceedings of IEEE Power
Tech., Lausanne, Suisse, July 2007, pp. 994–998.
[32] F. Cupertino, E. de Vanna, L. Salvatore, S. Stasi, Analysis techniques for detection of IM broken rotor bars after supply disconnection, IEEE Trans. Ind.
Appl. 40 (March/April (2)) (2004) 526–533.
[33] S.H. Kia, H. Henao, G.A. Capolino, A high-resolution frequency estimation method for three-phase induction machine fault detection, IEEE Trans. Ind.
Electron. 54 (August (4)) (2007) 2305–2314.
[34] A. Bellini, G. Franceschini, C. Tassoni, Monitoring of induction machines by maximum covariance method for frequency tracking, IEEE Trans. Ind. Appl.
42 (January/February (1)) (2006) 69–78.
[35] B. Yazici, G.B. Kliman, An adaptive statistical time-frequency method for detection of broken bars and bearing faults in motors using stator current,
IEEE Trans. Ind. Appl. 35 (March/April (2)) (1999) 442–452.
[36] M. Blodt, M. Chabert, J. Regnier, J. Faucher, Mechanical load fault detection in induction motors by stator current time-frequency analysis, IEEE Trans.
Ind. Appl. 42 (6) (2006) 1454–1463.
[37] E.H. El Bouchikhi, V. Choqueuse, M.E.H. Benbouzid, A parametric spectral estimator for faults detection in induction machines, in: Proceedings of the
2013 IEEE IECON, Vienna (Austria), November 2013, pp. 7356–7361.
[38] P. Zhang, Y. Du, T. Habetler, B. Lu, A survey of condition monitoring and protection methods for medium-voltage induction motors, IEEE Trans. Ind.
Appl. 47 (January/February (1)) (2011) 34–46.
[39] A.H. Bonnett, C. Yung, Increased efficiency versus increased reliability, IEEE Ind. Appl. Mag. 14 (January/February (1)) (2008).
[40] R. Schoen, T. Habetler, F. Kamran, R. Bartheld, Motor bearing damage detection using stator current monitoring, IEEE Trans. Ind. Appl. 31 (November/
December (5)) (1995) 1274–1279.
[41] M. Blodt, P. Granjon, B. Raison, G. Rostaing, Models for bearing damage detection in induction motors using stator current monitoring, IEEE Trans. Ind.
Electron. 55 (April (4)) (2008) 1813–1822.
[42] M. Blödt, P. Granjon, B. Raison, J. Regnier, et al., Mechanical fault detection in induction motor drives through stator current monitoring-theory and
application examples, in: Wei Zhang (Ed.), Fault Detection, 2010, pp. 451–488.
[43] S. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory, Prentice-Hall Signal Processing Series, 1993, 17th Printing.
[44] H. Li, P. Stoica, J. Li, Computationally efficient parameter estimation for harmonic sinusoidal signals, Signal Process. 80 (September) (2000) 1937–1944.
[45] Y. Bresler, Exact maximum likelihood parameter estimation superimposed exponential signals in noise, IEEE Trans. Acoust. Speech Signal Process. 1
(October (5)) (1986) 1081–1089.
[46] P. Stoica, Y. Selén, A review of information criterion rules, IEEE Signal Process. Mag. 21 (July (4)) (2004) 36–47.
[47] M. Wax, T. Kailath, Detection of signals by information theoretic criteria, IEEE Trans. Acoust. Speech Signal Process. – ASSP 33 (April (2)) (1985)
387–392.
[48] A. Ceban, R. Pusca, R. Romary, Study of rotor faults in induction motors using external magnetic field analysis, IEEE Trans. Ind. Electron. 59 (May (5))
(2012) 2082–2096.
[49] G. Houdouin, G. Barakat, B. Dakyo, E. Destobbeleer, C. Nichita, A coupled magnetic circuit based global method for the simulation of squirrel cage
induction machines under rotor and stator faults, in: Proceedings of ELECTRIMACS'02, Montreal, Canada, August 2002, pp. 18–21.
[50] A. Knight, S. Bertani, Mechanical fault detection in a medium-sized induction motor using stator current monitoring, IEEE Trans. Energy Convers. 29
(December (4)) (2005) 753–760.
[51] Z. Obeid, S. Poignant, J. Regnier, P. Maussion, Stator current based indicators for bearing fault detection in synchronous machine by statistical
frequency selection, in: Proceedings of the 2011 IEEE IECON 2011, Melbourne, Australia, November 2011, pp. 2036–2041.
[52] R.R. Schoen, T.G. Habetler, Effects of time-varying loads on rotor fault detection in induction machines, IEEE Trans. Ind. Appl. 31 (4) (1995) 900–906.
[53] K.B. Petersen, M.S. Pedersen, The Matrix Cookbook, November 2008.
El Houssin El Bouchikhi was born in Khemisset, Morocco, in 1987. He received the Dipl.-Ing. and the M.Sc. degrees in
automatic and electrical engineering, from the National Polytechnic Institute of Toulouse (INP-ENSEEIHT), Toulouse, France, in
2010. He is currently working toward the Ph.D. degree on offshore wind and marine current turbines condition monitoring
with the University of Brest, Brest, France. His current research interests are electrical machines fault detection and diagnosis
through electrical quantities, especially in non-stationary operating conditions.
Vincent Choqueuse (M'08) was born in Brest, France, in 1981. He received the Dipl.-Ing. and the M.Sc. degrees in 2004 and
2005, respectively, from Troyes University of Technology, Troyes, France, and the Ph.D. degree in 2008 from the University of
Brest, Brest, France. Since September 2009, he has been an Associate Professor with the Institut Universitaire de Technologie of
Brest, University of Brest, Brest, France, and a member of the LBMS Lab (EA 4325). His research interests focus on signal
processing and statistics for diagnosis and MIMO systems.
Mohamed El Hachemi Benbouzid (S'92-M'95-SM'98) was born in Batna, Algeria, in 1968. He received the B.Sc. degree in
electrical engineering from the University of Batna, Batna, Algeria, in 1990, the M.Sc. and Ph.D. degrees in electrical and
computer engineering from the National Polytechnic Institute of Grenoble, Grenoble, France, in 1991 and 1994, respectively,
and the Habilitation à Diriger des Recherches degree from the University of Picardie “Jules Verne,” Amiens, France, in 2000.
After receiving the Ph.D. degree, he joined the Professional Institute of Amiens, University of Picardie “Jules Verne,” where he
was an Associate Professor of electrical and computer engineering. Since September 2004, he has been with the Institut
Universitaire de Technologie of Brest, University of Brest, Brest, France, where he is a Professor of electrical engineering. His
main research interests and experience include analysis, design, and control of electric machines, variable-speed drives for
traction, propulsion, and renewable energy applications, and fault diagnosis of electric machines.
Prof. Benbouzid is a Senior Member of the IEEE Power Engineering, Industrial Electronics, Industry Applications, Power
Electronics, and Vehicular Technology Societies. He is an Associate Editor of the IEEE Transactions on Energy Conversion, the
IEEE Transactions on Industrial Electronics, the IEEE Transactions on Sustainable Energy, and the IEEE Transactions on Vehicular
Technology. He was an Associate Editor of the IEEE/ASME Transactions on Mechatronics from 2006 to 2009.

Mechanical Systems and Signal Processing: El Houssin El Bouchikhi, Vincent Choqueuse, Mohamed Benbouzid

Uploaded by

Copyright:

Available Formats

Mechanical Systems and Signal Processing: El Houssin El Bouchikhi, Vincent Choqueuse, Mohamed Benbouzid

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mechanical Systems and Signal Processing: El Houssin El Bouchikhi, Vincent Choqueuse, Mohamed Benbouzid

Uploaded by

Copyright:

Available Formats

Mechanical Systems and Signal Processing ] (]]]]) ]]]–]]]

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

Induction machine faults detection using stator current

Nomenclature DSP digital signal processing

The remaining parts of this paper are organized as follows.

Induction machine state Frequency signature Parameters Ω

2. Stator current signal model

2.1. Faults effect on stator current spectrum

2.2. Induction machine stator current signal model

The stator current signal model is based on the following assumptions:

Fig. 4. Theoretical PSD for L ¼2 [15].

H2 : The noise is white, Gaussian with zero-mean and variance σ2.

3. Parametric PSD estimation

3.1. Proposed estimator

3.1.1. Estimate of v and Ω

where pðx; v; ΩÞ is the probability density function (pdf) of x which is given by

Lðx; v; ΩÞ ¼ ðx AðΩÞvÞT ðx AðΩÞvÞ ð9Þ

A† ðΩÞ ¼ ðAT ðΩÞAðΩÞÞ 1 AT ðΩÞ ð11Þ

J ðΩÞ ¼ xT AðΩÞA† ðΩÞx ð13Þ

3.1.2. Estimate of model order L

The noise variance estimate is given by

3.2. Link with FFT

where DFT x ½f  is the DFT computed at frequency f, i.e.

4. Fault detection criterion

4.1. Proposed criterion

Fig. 6. Fault detection algorithm.

This criterion exhibits the two following desirable properties.

Property 1. For healthy machines, the criterion is equal to 0 i.e.

4.2. Algorithm summary

5. Application to induction machine faults detection

5.1. Simulation results

5.1.1. Induction machine modeling

Fig. 7. Rotor bars equivalent circuit [14].

5.1.2. Eccentricity fault detection results

5.1.3. Broken rotor bars fault detection results

5.2. Experimental results

5.2.1. Test rig

Bearing fault Fault related frequency Theoretical values (Hz)

Fig. 11. Test rig scheme.

5.2.2. Fault detection results

By expanding (9), we obtain

Finally, we obtain the following ML estimator of v:

A† ðΩÞ ¼ ðAT ðΩÞAðΩÞÞ 1 AT ðΩÞ ð35Þ

The ML estimate of f k ðΩÞ is obtained by minimizing Lðx; v

J ðΩÞ ¼ xT AðΩÞA† ðΩÞx ð38Þ

Appendix C. Model order estimation

Departing from Eq. (14) we have

Using (15) the cost function is equivalent to

Appendix D. Link with FFT proof

By using (4), we obtain

By using the structure of M, we also get

RATED DATA OF THE TESTED INDUCTION MACHINE

0.75 kW, 50 Hz, 220/380 V, 3.4/1.95 A, 2780 rpm, p ¼1

You might also like

where DFT x ½f is the DFT computed at frequency f, i.e.