2014

International Symposium on Power Electronics,

Electrical Drives, Automation and Motion

The Use of a Modified Prony’s Method to Track the

Broken Rotor Bars Characteristic Frequencies and
Amplitudes, in Three-Phase Induction Motors
M. Sahraoui A. J. M. Cardoso A. Ghoggal
LGEB, University of Biskra, Algeria CISE - Electromechatronic Systems Electrical Engineering Department,
CISE - Electromechatronic Systems Research Centre, University of Beira University of Biskra, Algeria
Research Centre, University of Beira Interior, Covilhã, Portugal
Interior, Covilhã, Portugal ajmcardoso@ieee.org
m.s.sahraoui@ieee.org

Abstract—The knowledge of the broken rotor bars , in the stator current spectrum, where f s is the supply
characteristic frequencies and amplitudes has a great importance
for all related diagnostic methods. The monitoring of motor frequency and s is motor slip. Therefore, the identification of
faults requires a high resolution spectrum to separate different the sideband frequencies and the evaluation of their
frequency components. The Discrete Fourier Transform (DFT) amplitudes can be used as an efficient and reliable approach to
has been widely used to achieve these requirements. However, diagnose rotor bar faults. However, these sidebands are
DFT can give meaningful information only for stationary usually quite close to the fundamental frequency and they may
harmonics which cannot be guaranteed in real cases. In addition, become sensitive to load variations. In addition, the
a long data sequence is necessary for DFT to get high frequency magnitudes of the sideband components are in the range of -20
resolution. Nevertheless, the signals are time-varying, and the to -60 dB below the magnitude of the fundamental, which
steady-state conditions can be lost for a long time acquisition. As
a solution for these problems, this paper proposes an efficient
combined with the noise usually present in the signal, makes
time-domain technique based on a modified Prony’s method for difficult the task of separating the different harmonic
the estimation of the frequencies/amplitudes of broken rotor bar frequencies, and consequently, the harmonic estimation is still
faults. Using this technique, the stator current is divided into an open research topic.
short overlapped time windows, and each one is analyzed by the Many harmonic estimation techniques exist in the area of
Least Squares Prony’s method. The proposed technique provides signal processing, which are categorized as parametric and
a linear-time frequency/amplitude representation with high non-parametric methods [3]. Non parametric methods estimate
frequency resolution and adjustable time resolution. It is shown the spectrum directly from the signal, while the parametric
that this technique allows tracking the frequencies and ones use a suitable model to represent the signal and then
amplitudes of the sidebands around the fundamental frequency
component with a very high accuracy. The efficiency of the
estimate the parameters of this model from the available signal
proposed method is verified by simulation and experimental data.
tests. Among the numerous harmonic spectral analysis
techniques, the Discrete Fourier Transform (DFT) is the most
Keywords—Prony’s method; Filtering; MCSA; Frequency commonly adopted one. The DFT is a non-parametric method
resolution which transforms a periodic, discrete signal from time-domain
to frequency-domain with finite range of data samples.
I. INTRODUCTION
However, some disadvantages limit its applicability in
Nowadays, induction motors have a crucial importance in practice; in fact, the DFT is only efficient for stationary
industry. Internal faults such as broken rotor bars can cause harmonics (i.e. for signals under steady-state conditions). On
serious damage to the motor itself as well as to motor-related the other hand, the DFT requires a long data sequence to get a
equipment, which can lead to unexpected shutdown of the high frequency resolution. However, storing a big number of
industrial processes causing then considerable economic data samples is not always possible due to the system memory
losses. To avoid such problems, reliable diagnostic systems limitation. In addition, the broken rotor bar sideband
should be installed. Generally, a diagnostic system must frequencies are time-varying, and the steady-state conditions
include a robust fault detection algorithm, which allows can be lost for a long time acquisition. Since the frequency
detecting all defects, or any undesirable changes on the resolution is mandatory and it depends on the inverse of the
machine performances before a total failure occurrence. Motor time acquisition, reduction of this later decreases considerably
Current Signature Analysis (MCSA) has been widely used to the fault detection efficiency. In order to overcome the
detect broken rotor bars in squirrel cage induction motors [1], limitations of DFT for signals having non stationary
[2]. It is well known that broken rotor bars produce geometric harmonics, the Short Time Fourier Transform (STFT) has
and magnetic unbalances which induce sidebands, (1± 2ks ) f s been also widely used. The STFT is based on applying the
DFT algorithm on a sliding window. This technique allows a

978-1-4799-4749-2/14/$31.00 ©2014 IEEE

210
linear time-frequency representation; however, the time and variants. To overcome this major drawback, a new approach
frequency resolutions depend strongly on the type and the has been proposed in [10].This approach uses a subspace
length of the window. In fact, a very short window gives a method known as estimation of signal parameters via rotation
good time-resolution with a poor frequency resolution, invariance technique (ESPRIT). The ESPRIT has advantages
whereas a long window provides high frequency resolution in terms of estimation accuracy and computational efficiency
with reduced time resolution. Thus, a tradeoff among these compared to MUSIC method. The approach proposed in [10]
two mandatory requirements is difficult to fulfill. uses the zooming technique to reduce the complexity of the
Due to these limitations of DFT and STFT, other methods ESPRIT leading to more accurate frequency estimation then
for spectral analysis have been proposed in the literature. As ZMUSIC. Once the harmonic frequencies are known, their
far as the fault detection in induction motors is concerned, it is corresponding amplitudes are obtained using the least squares
known that the spectral fault signatures are usually localized in (LS) algorithm. Despite the good results obtained by this new
some frequency ranges, consequently not all the Nyquist approach, the amplitude/frequency tracking is still a problem
spectrum requires high resolution. Based on this idea, the for this approach.
Zoom-FFT (ZFFT) method was proposed in [4] to improve the An interesting approach to estimate harmonic frequencies
frequency resolution and reduce the computational time. of a signal is the Prony analysis (PA) [3], [11]. The PA is a
However, the frequency resolution obtained by ZFFT is still high-resolution spectral analysis method developed on the
dependent on the acquisition time. In [5], the Maximum basis of the original work of the French mathematician,
Covariance method for Frequency Tracking (MCFT) was used Gaspard de Prony. This method approximates a sampled
successfully to estimate and track the fundamental and the waveform as a linear combination of complex conjugate
rotor-slotting frequencies even under time-varying conditions. exponentials [12]. The PA has the ability to estimate
This method provides a high frequency resolution independent accurately all attributes of harmonics: frequency, amplitude,
of the sampling frequency and the time acquisition. However, phase and damping factor, from a short data record signal. The
the MCFT requires a long computation time when a large performances provided by PA overcome the drawback of the
frequency bandwidth and very high frequency resolution are DFT and many other methods. However, the original Prony’s
required. In addition, the MCFT cannot estimate the method is sensitive to noise and its direct application requires
amplitudes of the target harmonics which is extremely a complex and long computational time especially for high
necessary for the diagnostic procedures. In [6], a combined model order. To tackle these problems, many modified
use of MCFT and ZFFT techniques for the processing of versions of PA have been proposed in the literature. For
induction motor stator currents has been proposed to instance, in [13] a remedial technique for the root-finding
overcome the constraint of limited time or frequency process was proposed to overcome the computational burden
resolutions of classical DFT-based method. of the classical PA. This remedial technique uses a filter bank
Among various time-frequency methods, the Wavelet with a set of band-pass filters to eliminate the undesired
Transform (WT) and its variants have been extensively used harmonics and separate the signal into several frequency
for many fault diagnostic domains[7]. The WT methods can canals. Consequently, a few sampled points are needed for the
be considered as a band-pass filter with a varying window calculation, and the time-varying harmonic estimation can be
size, being wide for slow frequencies and narrow for the fast performed in a short time duration. However, the efficiency of
ones, which represents the main advantage. Thus, the WT this technique was tested only for simple generated signals
methods have multi-resolution capability and they also retain (not real signals) and it was found that the obtained phase
time information of the signal. The WT methods are suitable angles are deteriorated due to the non-linear response of the
for signals with gradual frequency changes or transient used band-pass filters. In [14] an adaptive Prony method is
conditions, however, their high computational burden and proposed for the evaluation of spectral components of power
interpretation complexities limit their stand-alone application system waveforms. This method is based on the application of
for harmonics estimation. the PA to some contiguous variable windows leading to a
The classical Multiple Signal Classification (MUSIC) reduction of the computational complexity. However, the
method is a noise subspace based parametric method that uses noise effect is still a problem for this method and its efficiency
sinusoidal or exponential model to estimate harmonic cannot be confirmed since it was only tested by numerical
frequencies present in the signal. It has been largely used in generated signals. In [15] an iterative reweighted least square
many fault detection applications [8]. Although the MUSIC Prony’s method is used to estimate the amplitudes and
method is a powerful tool for detecting frequencies, with very frequencies of the broken rotor bar sidebands. This method
high frequency resolution, from a short data record signal has shown a great ability of dealing with noisy signal. A
buried in a noise, it has the drawback of long computation comparison between the performances of the PA and the DFT
time when the dimension of the autocorrelation matrix was also performed in [15]; it was found that the minimum
increases. To tackle this problem, a combined use of the ZFFT window required by DFT can be 6-20 times longer than PA in
technique with the MUSIC method was proposed to improve light load conditions. However, the frequency/ amplitude
the accuracy of broken rotor bars detection. This method is tracking and the time-varying conditions were not studied. A
designated by ZMUSIC [9]. However, the amplitude combination of digital filtering and an algorithm using Prony's
estimation is still a problem for the MUSIC method and its method was presented in [16] for frequency estimation of

211
basic components of real measured signals from industrial ⎡ x[ P ] " x[1] ⎤ ⎡ a1 ⎤ ⎡ x[P + 1]⎤
circuits containing PWM converters. Another solution based ⎢ ⎥
# ⎥ , A = ⎢ # ⎥ , Y = ⎢⎢ # ⎥⎥
⎢ ⎥
where: X = ⎢ # %
on a combined use of wavelet and Prony methods has been
applied successfully for determining the beginning and the end ⎢⎣ x[2 P − 1] " x[ P ]⎥⎦ ⎢⎣a P ⎥⎦ ⎢⎣ x[2 P ] ⎥⎦
of transient phenomena and also for voltage dip detection [17]. The linear prediction parameters ak that fit the observed
The Kalman filter has been also used to improve the accuracy data are determined by solving (3), then, a characteristic
of the Prony’s method for amplitude estimation of electrical
polynomial with roots zk can be formed using the linear
signals of power systems [18]. The results showed that this
technique is an efficient tool to estimate time-varying prediction parameters as follows:
harmonics. P
In the present paper, a useful and efficient technique, based F ( z) = ∑ ak z P − k (4)
on a modified version of the Prony's method, is proposed to k =0
estimate and track the frequencies and corresponding
As a result, the damping factor and the frequency can be
amplitudes characteristic of the broken rotor bars occurrence
from a short data record current signal buried in a noise. Using deduced directly from the roots zk of (4):
this technique, the line stator current is divided into short ln z k 1 ⎡ Im( zk ) ⎤
overlapped time windows, and each one is analyzed by the αk = tan −1 ⎢
and f k = ⎥
Ts 2πTs ⎣ Re( zk ) ⎦
Least Squares Prony's method, which allows considering the
non-stationary aspect of the problem. The proposed technique Finally, the roots zk are used to write the P equations of (1)
provides a linear-time frequency/amplitude representation with in a matrix form as:
high frequency resolution and adjustable time resolution. This
⎡ 1 1 " 1 ⎤
technique shows a great ability of tracking frequencies and ⎢ z ⎥ ⎡ h1 ⎤ ⎡ x (1) ⎤
z " z P ⎥⎢ ⎥
# = ⎢ # ⎥⎥
amplitudes of the sidebands around the fundamental with very ⎢ 1 2 (5)
high. ⎢ # # # # ⎥⎢ ⎥ ⎢
⎢ P −1 ⎢
⎥ hP ⎥ ⎢ ⎥
x( P)⎦
II. THE PRONY'S METHOD ⎣ z1 z 2P −1 " z PP −1 ⎦ ⎣ ⎦ ⎣
The original Prony's method is a signal processing The complex parameters hk can be determined by solving
technique for extraction the sinusoid or exponential signals by (5), and consequently, the exponential amplitudes Ak and
solving a set of linear equations. Assuming the signal x(t ) and phases angle ϕ k can be obtained using the following
its N complex samples , the Prony's method approximates the
relationships:
sampling data with linear combination of P complex
exponential functions: ⎡ Im(hk ) ⎤
Ak = hk and ϕ k = tan −1 ⎢ ⎥
P ⎣ Re(hk ) ⎦
xˆ ( n ) = ∑ hk z kn −1 (1) It should be noticed that the performance of the classical
Prony’s method becomes unstable if there is noise in the
k =1
signal data. This is due to the fact that there is no error
with hk = and z k = e (α k + j 2πf k )Ts
Ak e jϕ k considered in the model. However, the real acquired signals
where Ts is the sampling time. The model parameters Ak , f k , are always embedded in noise. To deal with such situation, the
ϕ k and α k represent, respectively, the unknown amplitude, equation (1) should be modified to the following form:
P
frequency, phase angle and damping factor of the kth
component. The above equation is a difficult non linear
x[ n] = xˆ[ n] + ε [ n] = ∑ hk zkn −1 + ε [n] (6)
k =1
problem which can be solved using the Prony’s Method. In
fact, Prony's method turns the parameters estimation non where ε represents the approximation error between the
linear problem into the solution of a linear system and roots original data samples x[n] and the linear approximation xˆ[ n] .
calculation of the polynomial. On this direction, the PM This error is assumed to be white and has Gaussian
constructs an homogenous linear difference equation with distribution. The parameter estimation problem is based on the
constant coefficients (with a0 = 1 ): minimization of the squared error over the N samples:
P 2
N N P
∑ ak x[n − k ] = 0 (2) E= ∑ ε [ n] 2
= ∑ x[n] − ∑ hk z kn −1 (7)
k =0 n =1 n =1 k =1
In the classical Prony’s Method, the number of the In practice, the number of available data samples exceeds
available data samples is assumed to be equal to the unknown the number of unknown parameters ( N > 2P ). In the over-
model parameters, thus, the linear difference equation (2) can determined data case, the linear difference equation (2),
be expressed in matrix form as follows: should be modified to:
X . A = −Y (3)

212
 P III. SIGNAL PREPROCESSING
∑ ak x[n − k ] = ε [ n] (8) The success of amplitude and frequency estimation by the
k =0 Prony's method can be affected by several factors. Indeed, the
The available N data samples are used to rewrite (8) in a noise level, the number of the harmonics to be estimated, the
matrix form: sampling frequency, the model order, the number of samples
⎡ x[ P ] " x[1] ⎤ ⎡ a1 ⎤ ⎡ x ( P + 1) ⎤ considered in the estimation process and the signal nature
⎢ ⎥⎢ # ⎥ = −⎢ ⎥ (stationary or not) all affect the quality and the accuracy of the
# % # # (9)
⎢ ⎥⎢ ⎥ ⎢ ⎥ results. To reduce the impact of these influenced factors, a
⎢⎣ x[ N − 1] " x[ N − P ]⎥⎦ ⎢⎣ a P ⎥⎦ ⎢⎣ x ( N ) ⎥⎦ preprocessing of the acquired signal is extremely necessary.
The vector of the unknown parameters ak is selected to This represents a crucial step prior to Digital Signal
Processing procedures in order to obtain improved results.
minimize the linear prediction total squared error. The Data acquisition, filtering, DC components removal and dawn-
minimization can be solved by using the least square method. sampling are the main involved tasks.
In addition, the estimation of the complex parameters hk is
A. Data acquisition
switched also to a linear least square procedure. In this case,
the set of linear equations (5) should be modified to the During acquisition, the stator current which is an analog
following compact form: signal is transformed into a discrete signal characterized by a
Z .H = C (10) number of data points and a determined sampling frequency.
⎡ 1 1 " 1 ⎤ B. Filtering
⎢ z ⎡ h1 ⎤ ⎡ x (1) ⎤
z2 " z P ⎥⎥ It is an important task having a direct effect on the quality
With Z = ⎢ 1
, H = ⎢ # ⎥ , C = ⎢⎢ # ⎥⎥
⎢ ⎥
of the results. In the present case, filtering the stator current
⎢ # # # # ⎥
⎢ N −1 N −1 N −1 ⎥
⎢⎣hP ⎥⎦ ⎢⎣ x ( N ) ⎥⎦ has two main objectives: attenuation of noise and separation of
z
⎣ 1 z 2 " z P ⎦ the broken rotor bar sidebands from undesired spectral
In this paper, the treated signals (stator currents) are time- components. Using the “fdesign.bandpass” function from the
varying and they undergo sudden variations due to the Filter Design Toolbox of Matlab, a band-pass filter with a
occurrence of broken rotor bars. The Prony’s Method narrow pass band frequencies is designed and employed to
described above cannot deal with such situation, and it is process the current signal after acquisition (Fig.2). The
unable to track the harmonic properties. To overcome these parameters of this filter are specified to allow passing the
drawbacks, a Short Time Least Square Prony’s method fundamental and the two sidebands frequencies only. The
(STLSP) is proposed. filter specifications are presented in Table I.

TABLE I. THE BANDPASS FILTER SPECIFICATIONS

Band-pass Filter specifications Values

frequency at the edge of the start of the first
Fstop1 45Hz
stop band
frequency at the edge of the start of the pass
Fpass1 46Hz
band
frequency at the edge of the end of the pass
Fpass2 54Hz
band
frequency at the edge of the start of the
Fstop2 55Hz
second stop band
Astop1 attenuation in the first stop band in dB 40dB
Fig. 1. Description of the proposed method Astop2 attenuation in the second stop band in dB 40dB
Apass amount of ripple allowed in the pass band 0.5dB
Using the STLSP, the stator current is divided into short
overlapped time windows, and each one is analyzed by the
Least Squares Prony’s method (Fig.1). The Prony’s method is
well suitable for the proposed technique since it was originally
developed to work with small number of data samples.
Compared with the STFT, where there is a design trade-off
that must be made between the time and frequency resolution,
the STLSP provides very high frequency resolution even for
very short time windows. This can reduce considerably the
computational cost together with time acquisition. Moreover,
the time resolution can be ameliorated by increasing the
overlapping degree without affecting the frequency resolution. Fig. 2. Band pass filter design
Hence, tracking harmonic attributes of distorted signals can be
achieved with high accuracy.

213
C. The model order
h
The Prony's method is unfortunately a high computational
cost algorithm, because it requires the inverrsion of large size
matrices and deals with high order poolynomials. These
complexities depend strongly on the model order P and the
number of treated sampled data N . The propper selection of P
and N can be considered as a critical taskk to obtain correct
estimations. In fact, a model order smaller thhan the number of
the searched harmonics leads to great errorss on the harmonic
estimation. However, increasing the model order
o can improve
the estimation precision but it causes long computational time Fig. 3. Frequency estimation for the fundamental
f and the two sidebands
and produces many fake signal componentts. Some selecting
approaches have been proposed in [133] but they are
complicated and my not work in all situationns.
The efforts done for proper selection off P and N can be
considerably reduced if there is prior knowledge
k of the
number of searched harmonics. Fortunatelly, in the broken
rotor bar diagnosis applications, the numbeer of the searched
harmonics is known as there are onlly three spectral
components in the band pass filtered current signal. For our
application, it is found that P = 7 and N = 15 are largely
sufficient to obtain very good results.
IV. SIMULATION RESULTTS
Fig. 4. Frequency estimation error
The proposed method STLSP was impleemented in Matlab
and tested using a simulated stator current signal of a three- In order to demonstrate thhe tracking capabilities of the
phase induction motor with broken rotor r bars. The STLSP in the case of a tiime-varying signal, the same
mathematical model used to generate the stator currents is mathematical model was usedd to simulate the occurrence of
based on the winding function approacch which allows broken rotor bars. This fault inntroduces high distortions in the
considering the rotor slots harmonics. Thhis model is well stator current and the estimatioon frequencies can be regarded
described in [19]. The motor is 3 kW with four poles and 28 as a challenge.The results deppicted in Fig. 5 prove the high
rotor bars. The current signal is sampled at thhe frequency of 10 accuracy of the proposed methhod for estimating and tracking
kHz for 10 s. In order to track the frequenciies of the searched the amplitudes and the frequenncies of the fundamental and the
components, the acquired signal is divided into a lot of short two sidebands even for a high distorted
d signal.
overlapped time windows with small size (15 ( samples). The
overlapping degree is fixed as 14 sampless which represent
93.3% of the sliding window. Each window is analyzed by the
Least Squares Prony’s method. Fig. 3 show ws the frequencies
estimation using the STLSP. All figuress have the same
frequency axis range for better observations.
fr
It is important to note that the sideband frequencies, during
the simulation record, are not constant; thhey present small
fluctuations due to the presence of rotor sloots harmonics. The
estimation accuracy is evaluated usingg the frequency
estimation error defined as follows:
f true − f estim
E freq (%) = 100 (11) Fig. 5. Simulation test: Frequency annd amplitude tracking for a motor with
f true rotor bar fault occurred suddenly at 2s
where f true is the true frequency calculated frrom the motor slip
V. EXPERIM
MENTAL RESULTS
and f estim is the estimated frequency valuee obtained by the
In order to test the STLSP P for practical cases, measured
proposed method. Fig.4 shows the time evolution of the
signals from laboratory experim ments are used. A three-phase 3
frequency estimation error for the threee harmonics (the kW squirrel cage induction mootor, with four poles and 28 bars
fundamental and the two sidebands). Obvviously, E freq has is used in the experiments (Figg. 6). The broken rotor bars fault
very small values over the signal length. Thiss confirms that the is introduced by drilling holes in
i the rotor bars close to the end
STLSP can give high frequency resolution using only small ring.
number of data samples.

214
 REFERENCES
[1] P. Zhang, Y. Du, T. G. Habetler, B. Lu, “A Survey of Condition
Monitoring and Protection Methods for Medium-Voltage Induction
Motors,” IEEE Transactions on Industry Applications, vol.47, no. 1, pp.
34-46, 2011.
[2] A. Bellini, F. Filippetti, C. Tassoni, G. A. Capolino, “Advances in
Diagnostic Techniques for Induction Machines,” IEEE Transactions on
Industrial Electronics, vol. 55, no. 12, pp. 4109-4126, 2008.
[3] S. K. Jain, S. N. Singh, “Harmonics estimation in emerging power
system: Key issues and challenges,” Elec. Pow. Sys. Res, vol. 81, pp.
1754-1766, 2011.
[4] J. Zarader, M. Garnier, and M. Nicollet, “New frequency-filtering
Fig. 6. Experimental test bench. zoom,” Int. J. Adapt. Control Signal Process., vol. 6, pp. 547-560, 1992.
[5] A. Bellini, G. Franceschini, and C. Tassoni, “Monitoring of induction
Since to rotor bar fault cannot be introduced suddenly, the machines by maximum covariance method for frequency tracking,”
motor was tested first at healthy state and then with two broken IEEE Trans. Ind. Appl., vol. 42, no. 1, pp. 69-78, Jan./Feb. 2006.
bars. In both cases, 10 s of the three-phase currents are sampled [6] A. Bellini, A. Yazidi, F. Filippetti, C. Rossi and G. A. Capolino, “High
at 10 kHz using a dSpace 1104 card. frequency resolution techniques for rotor fault detection of induction
machines,” IEEE Trans. Ind. Electron., vol. 55, no. 12, pp. 4200-4209,
Dec. 2008.
[7] Y. Wang, Z. He, and Y. Zi, “Enhancement of signal denoising and
multiple fault signatures detecting in rotating machinery using dual-tree
complex wavelet transform,” Mech. Syst. Signal Process., vol. 24, no. 1,
pp. 119-137, Jan. 2010.
[8] A. Garcia-Perez, R. J. Romero-Troncoso, E. Cabal-Yepez, and R. A.
Osornio-Rios, “The application of high-resolution spectral analysis for
identifying multiple combined faults in induction motors,” IEEE Trans.
Ind. Electron., vol. 58, no. 5, pp. 2002-2010, May 2011.
[9] S. H. Kia, H. Henao and G.-A. Capolino, “A High-Resolution Frequency
Estimation Method for Three-Phase Induction Machine Fault
Detection,” IEEE Trans. Ind. Electron., vol. 54, no. 4, pp. 2305-2314,
August 2007.
[10] Y. H. Kim, Y. W. Youn, D. H. Hwang, J. H. Sun, and D. S. Kang,
“High-Resolution Parameter Estimation Method to Identify Broken
Rotor Bar Faults in Induction Motors,” IEEE Trans. Ind. Electron, vol.
60, no. 9, pp.4103-4117, Sep. 2013.
[11] F. F. Costa, A. J. M. Cardoso, “Harmonic and Interharmonic
Identification Based on Improved Prony’s Method,” Proc. 32nd Annual
Fig. 7. Experimental test: Frequency and amplitude tracking for ahealthy Conf. IEEE Industrial Electronics Society, Paris, France, pp.1047-1052,
then with broken rotor bar. Nov. 2006.
[12] D. Potts, M. Tasche, “Parameter estimation for exponential sums by
The acquired signal was preprocessed before being approximate Prony method,” Signal Processing, vol. 90, pp. 1631-1642,
analyzed by the proposed STLSP. For this test, a short sliding 2010.
time-window of only 30 samples is used which represents 30 [13] C. I. Chen, G.W. Chang, “An efficient time-domain approach based on
ms. The order of the algorithm is chosen to P = 7 . The Prony's method for time-varying power system harmonics estimation,”
IEEE Power & Energy Society General Meeting, pp. 1-6, 2009.
obtained results are displayed in Fig. 7. As can be clearly seen,
[14] A. Bracale, P. Caramia,G. Carpinelli, “Adaptive Prony method for
these results demonstrate the capability of the STLSP to deal waveform distortion detection in power systems,” Elec. Pow. Sys. Res,
with practical noisy signals. Besides, since the STLSP track vol.29, pp. 371-379, 2007.
continuously the sideband amplitudes, the broken rotor bars [15] S. Chen, R. Zivanovic, “Estimation of frequency components in stator
can be detected easily. current for the detection of broken rotor bars in induction machines,”
Measurement, vol. 43, pp. 887-900, 2010.
VI. CONCLUSION [16] T. Lobos, I. Hejke, J. Rezmer, T. Sikorski, P. Kostyla, “Parametric
algorithms for frequency estimation in PWM converter systems,” Elec.
This paper introduces the Short Time Least Square Prony’s Pow. Sys. Res., vol.79, pp. 1568-1573, 2009.
(STLSP) method which is an improvement of the classical [17] T. Lobos, J. Rezmer, P. Janik, H. Amarís, M. Alonso, C.
Prony's Analysis. The STLSP method is not only able to Álvarez,“Application of wavelets and Prony method for disturbance
estimate amplitudes and frequencies of the searched detection in fixed speed wind farms,” Elec. Pow. Ener. Sys., vol.31, pp.
429-436, 2009.
harmonics, but also track the amplitudes and frequencies even
[18] F. F. Costa, A. J. M. Cardoso, D. A. Fernandes, “Harmonic Analysis
for time-varying signals with high distortion due to the Based on Kalman Filtering and Prony’s Method,” Proc. Int. Conf. Power
occurrence of broken rotor bars, for example. Improving the Eng., Energy and Electrical Drives, Setubal, Portugal, pp. 696-701,
accuracy of this method for a variety of types of analyzed April 2007.
signals will be the subject of other works, but these very [19] A. Ghoggal, S. E. Zouzou, H. Razik, M. Sahraoui and A. Khezzar, “An
preliminary investigations show the opportunity for use of the improved model of induction motors for diagnosis purposes – Slot
skewing effect and air–gap eccentricity faults,” Energy conversion and
proposed method especially in those systems which have a Management, vol. 50, no. 5, pp. 1336-1347, 2009.
low stability of harmonics.

215