4200 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO.

12, DECEMBER 2008

High Frequency Resolution Techniques for Rotor

Fault Detection of Induction Machines
Alberto Bellini, Member, IEEE, Amine Yazidi, Associate Member, IEEE, Fiorenzo Filippetti, Member, IEEE,
Claudio Rossi, Member, IEEE, and Gérard-André Capolino, Fellow, IEEE

Abstract—Motor current signature analysis (MCSA) is the Identical considerations hold, relying on the more efficient fast
reference method for the diagnosis of medium–large machines in Fourier transform (FFT) algorithm, which is only faster in terms
industrial applications. However, MCSA is still an open research of computational cost. The requirement of long data sequences
topic, as some signatures may be created by different phenomena,
wherein it may become sensitive to load and inertia variations, and is not always possible due to hardware and software limitations.
with respect to an oscillating load torque, although suitable data On the other hand, if a sequence is recorded over a long time,
normalization can be applied. Recently, the topic of diagnostic the steady-state condition can be lost, and both DFT and FFT
techniques for drives and low to medium size machines is be- become inefficient.
coming attractive, as the procedure can be embedded in the drive Motor current signature analysis (MCSA) is the reference
at no additional thanks to a dedicated firmware, provided that a
suitable computational cost is available. In this paper, statistical method for the diagnosis of medium–large machines in indus-
time-domain techniques are used to track grid frequency and trial applications. However, MCSA is still an open research
machine slip. In this way, either a lower computational cost or topic, as some signatures may be created by different phenom-
a higher accuracy than traditional discrete Fourier transform ena, wherein it may become sensitive to load and inertia vari-
techniques can be obtained. Then, the knowledge of both grid ations, and with respect to an oscillating load torque, although
frequency and machine slip is used to tune the parameters of the
zoom fast Fourier transform algorithm that either increases the suitable data normalization can be applied [1].
frequency resolution, keeping constant the computational cost, or Moreover, the wide frequency spreading of inverter-fed in-
reduces the computational cost, keeping constant the frequency duction machines enforces the requirement of high frequency
resolution. The proposed technique is validated for rotor faults. spectrum analysis, while the monitoring of electric faults re-
Index Terms—Current monitoring, diagnostic techniques, quires a high resolution spectrum analysis. The signal is time
Fourier transforms, frequency domain analysis, induction varying, and it may be difficult to obtain steady-state conditions
machines, spectrum resolution, statistical analysis. for a suitable time acquisition period. On the other hand, it is
not always possible to take a large number of data points due to
I. I NTRODUCTION the limitation of the digital system memory size. The reduction
of the time acquisition period in FFT-based methods is not

I N THE LITERATURE, several techniques have been pre-

sented in order to perform accurate tracking of instantaneous
frequencies in complex signals for different purposes. As far
always possible because both frequency and time resolution are
mandatory, and the frequency resolution depends on the inverse
of the time acquisition period. The short time Fourier trans-
as the fault diagnosis of three-phase squirrel-cage induction form (STFT), which is a linear-time frequency representation,
machines is concerned, the accurate knowledge of supply has been widely used due to its robustness and simplicity to
frequency and machine slip is required. Specifically, a high overcome the limitations of FFT for time-varying signals [2].
resolution frequency analysis is necessary to separate different However, the STFT is dependent on the type and the length of
frequency lines that are linked to different phenomena. As an the window, so that it has to be chosen as a tradeoff between
example, broken rotor bar faults can be efficiently detected time and frequency resolutions. Specifically, in STFT, a signal
by checking the sideband harmonics around the fundamental is divided into small time windows, and each is analyzed using
frequency of a stator phase current. Traditional discrete Fourier the Fourier transform. This formulation provides localization
transform (DFT) techniques require a very long data sequence in time while simultaneously capturing frequency information.
and high computational cost to get high frequency resolution. The time–frequency tiling for the STFT is uniform across time
and frequency. In the implementation of the STFT, a design
Manuscript received January 27, 2008; revised September 12, 2008. First tradeoff must be made between time and frequency resolution.
published October 31, 2008; current version published December 2, 2008. A short duration window provides good time resolution at the
A. Bellini is with the Department of Engineering Sciences and Methods,
University of Modena and Reggio Emilia, 42100 Reggio Emilia, Italy (e-mail:
expense of poor frequency resolution, whereas a long duration
[email protected]). window provides good frequency resolution at the expense
A. Yazidi and G.-A. Capolino are with the Department of Electrical Engi- of reduced time resolution. The windowing and overlaps of
neering, University of Picardie Jules Verne, 80039 Amiens, France.
F. Filippetti and C. Rossi are with the Department of Electrical Engineer- adjacent time windows can be used to solve partially these
ing, University of Bologna, 40136 Bologna, Italy (e-mail: fiorenzo.filippetti@ issues. In summary, with short time frequency transform, the
unibo.it; [email protected]). computational cost is reduced together with the time acquisition
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. period. Hence, better time resolution and better assumption of
Digital Object Identifier 10.1109/TIE.2008.2007004 stationarity are achieved; however, the frequency resolution is

0278-0046/$25.00 © 2008 IEEE

BELLINI et al.: RESOLUTION TECHNIQUES FOR ROTOR FAULT DETECTION OF INDUCTION MACHINES 4201

poorer. Thus, STFT is considered inadequate for diagnostic This paper presents a solution to the aforementioned prob-
purposes, [3] and the proposed method clearly outperforms it. lems based on the combined use of maximum covariance
The Goertzel algorithm was used in [4] as an efficient way method for frequency tracking (MCFT) [34] and ZFFT algo-
to compute a single component in the frequency spectrum and rithm [15], [35] with special reference to rotor fault detection.
in [5] for speed estimation in sensorless induction machines. It The former is used to track efficiently both the instantaneous
does not solve the issue of computing accurately the frequency grid or power converter fundamental frequency and the induc-
of the fault signature. Moreover, it does not solve the fixed con- tion machine slip; then, the latter is used to perform accurate
straint between the number of samples, i.e., the time acquisition frequency analysis around the sideband harmonics of interest.
period and the obtained resolution. MCFT allows one to use a very small frequency range, thus
In the literature, many other spectral estimation techniques reducing the required number of samples and the computa-
have been presented to cope with these drawbacks [6]–[9]. tional cost. The combined use of the aforementioned techniques
FFT-based estimation techniques are frequency interpolation allows one to overcome the constraint of reduced available
techniques, which cannot seek for more frequencies than the memory or reduced time acquisition period of the signal. On
actual available bins [10], [11]. Recently, novel estimation the other hand, relying on traditional DFT-based methods, these
techniques have been developed with the underlining idea that constraints would be fixed.
not the whole Nyquist spectrum requires the full resolution. The
zoom-FFT (ZFFT) and the chirp z transform methods embody II. B ASIC T HEORY OF MCFT AND ZFFT
the aforementioned concept. In ZFFT technique [12]–[15], the
computational time and the accuracy in a specified frequency A. MCFT
range are increased. However, as expected, the frequency reso- MCFT [34] allows the sharp estimation of the fundamental
lution is still affected by the time acquisition period. Similarly, frequency and the initial phase displacement of a signal even
the chirp z transform is powerful in detecting frequencies in a under time-varying conditions. MCFT is a time-domain-based
specified bandwidth. It is based on the z transform of sample method that allows one to track the fundamental frequency of a
data points into a circle in the z plane, where the frequency line current signal and the slip of an induction machine relying
range is an arc which can be divided in any arbitrary number on rotor slotting. This method can be seen as an extension of the
[16]. This process reproduces the interpolation in the frequency Fourier series expansion, where the base frequency is variable.
domain, and it is time consuming. In [17], a critical comparison In the Fourier series expansion, the frequency of the kernel is
of signal-processing-based techniques is presented with special changed with discrete steps that are multiple harmonics of the
reference to the detection of broken bars and bearing deteriora- base frequency f0 = 1/T0 , where T0 is the period of the signal.
tion in induction motors. On the contrary, in the MCFT, the frequency of the sinusoidal
Another mean to tackle the resolution problem is the use of kernel can be changed with a step as small as the accuracy
time–frequency methods, which have been also investigated. of the digital system could allow. Only the largest harmonic
They are transformation methods that rely on two independent component in the prefixed seeking bandwidth [fmin , fmax ] is
variables, related to time and frequency. Several papers have kept, and the phase is tuned in order to achieve the maximum
deeply analyzed the application of wavelets [18]–[21] and the cross correlation.
application of Wigner–Ville distributions with their variants The ratio between the computational costs of MCFT and FFT
[22], [23]. for fixed frequency resolution Δf and sampling frequency Fs
Artificial intelligence techniques were also applied to the for typical parameters is (1), as proved in [34]. Relationship (1)
same topic in order to find a nondeterministic solution holds in case the seek bandwidth (fmax − fmin ) is around 10%
[24]–[30]. Recently, other spectrum estimation techniques were of the target frequency ftarget . The gain increases, decreasing
proposed in the literature in order to track accurately rotor speed the seek bandwidth
[31]–[33].  
The computational cost of all these solutions is very high, 1 Fs
costgain = log2 . (1)
and floating point processors with a large storage capability 5 Δf
are required to compute them. The aforementioned items result
in two main drawbacks. At first, a low-cost realization would It turns out that, for the same resolution, the computational
require a simple algorithm that can be implemented in a fixed- cost of MCFT is typically three times lower than that of
point processor, typically shared with the drive control. Then, the computational cost of FFT for frequency resolution lower
the requirements of the computational cost for time–frequency that 1 Hz. This ratio increases with the sampling frequency.
distributions are mitigated, storing a large number of sam- Specifically, the computational cost for MCFT is
ples in a buffer and thus resulting in a remarkable latency Fs fmax − fmin
between the actual faults and their effective recognition. The costMCFT = 50 .
ftarget Δf
previous drawbacks make the investigation of reduced cost
algorithms for monitoring of electric drives in industrial appli- For MCFT, there is no fixed constraint between resolution
cations attractive, where cost and reliability are major concerns. and required time acquisition period, and between frequency
As an example, appliance and automotive industries feature resolution and computational cost, as it is for FFT. A very high
both highly time-varying signals and tight cost and reliability frequency resolution Δf can be achieved, reducing the seek
requirements. bandwidth.
4202 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 12, DECEMBER 2008

sequence by a complex unit vector with a frequency fi , which

is the lower frequency of the band.
Let us consider N samples of a digital signal xin [k], sampled
at frequency Fs . An N -point FFT would achieve a frequency
resolution of Δf = Fs /N . Owing to a complex mixer, the
signal is demodulated, translating the frequency spectrum by fi .
Then, the signal is low-pass filtered with a cutoff frequency that
is equal to Bp , where Bp is the band of interest. The filter used
has to be as flat and as narrow as possible in the transition band
[15]. The filtered signal xf [k] is down sampled at frequency
Fs = Fs /C, where

N
C= (3)
2B
Fig. 1. MCFT flowchart for grid frequency and machine slip detection.
and B is the number of samples in the frequency band
In fact, with MCFT, the frequency resolution Δf is equal to [fi , fi + Bp ]
the step of spanning of the reference tone used to compute the
maximum cross correlation. In this way, Δf is independent of B = (Bp N )/Fs . (4)
both the sampling frequency and the time acquisition period
Tw . On the contrary, with traditional DFT-based techniques, Hence, C = (Fs /2 Bp ). The number of samples is reduced
the frequency resolution is directly related to the sampling to A = (N/C). Then, applying an A-point FFT to the down-
frequency sampled signal xsec [k], the frequency resolution of the obtained
spectrum is the same as before
1 Fs
Δf = = (2)
Tw N
Fs Fs
ΔfZFFT = = . (5)
where Fs is the sampling frequency, N is the number of A N
samples and the number of points of the DFT, i.e., the length
of the input/output vector. The gain in computational cost achieved, owing to ZFFT,
MCFT is used to track the supply frequency and the machine is related to the down-sampling factor C. Specifically, the
slip, relying on rotor slotting frequencies, as shown in Fig. 1. traditional FFT requires N log2 N cycles to achieve a frequency
resolution of Δf , while ZFFT requires A log2 A cycles to
achieve the same frequency resolution. Then, the computational
B. Basic Theory of ZFFT cost is reduced by
The ZFFT is a useful technique for zooming (i.e., extending
the accuracy) in a narrow frequency band for spectral analysis log2 N
costgain = C . (6)
purposes. The frequency resolution of the FFT is only related log2 N − log2 C
to the total length of time spanned by the samples regardless
of the sample rate. At high sample rates, a large number of Consequently, the latency is reduced, allowing for a more
samples have to be taken to span a given time and to achieve effective frequency tracking in a complex spectrum. Moreover,
a given resolution (2). Thus, for high resolution with high the proposed technique is consistent with the time-varying
sample rate, where only a narrowband is of interest, FFT sizes signals, as a few tenths of periods are sufficient to achieve a
are inconveniently large and wasteful. On the other hand, the nice tracking of instantaneous frequency. On the other hand, if
computation time required by FFT increases linearly with the N C samples are available at frequency Fs , keeping the same
number of samples. For real-time implementation on modern computational cost, the frequency resolution is decreased to
DSP hardware, the size of FFT is limited by the memory avail-
able on traditional boards. The ZFFT uses a direct reduction Δf
ΔfZFFT = . (7)
of sample rate to decrease the number of samples, keeping C
the same time acquisition period and frequency resolution. It
turns out that the computational cost is accordingly reduced. Of course, this last operation requires a higher time ac-
The technique is closely related to multirate filtering, and quisition period. In this way, the computation time decreases
specifically to single sideband modulation techniques [14], while the time acquisition period and size increase, and this
[36]. A narrow span centered on any frequency allows zooming has to be considered when the algorithm is implemented in the
around it, reducing the required number of samples keeping digital system. By using ZFFT, it is more interesting to have
the same resolution. The frequency translation in the input a large memory size than a high speed DSP. Moreover, the
signal sequence can be obtained using a digital complex mixer. same processor used for drive control may be used to perform
This element consists of two multipliers that multiply the input frequency analysis at no additional cost.
BELLINI et al.: RESOLUTION TECHNIQUES FOR ROTOR FAULT DETECTION OF INDUCTION MACHINES 4203

III. A PPLICATION TO I NDUCTION

M ACHINE F AULT D IAGNOSIS
A. Methodology
The combined use of MCFT and ZFFT allows an efficient
and reliable diagnosis of faults for three-phase squirrel-cage
induction machines. As an example, the case of rotor faults will
be analyzed.
At first, MCFT is used to estimate accurately the supply
frequency fsupply from either the grid or the power converter
and the machine slip s. The machine slip is estimated from
rotor slotting frequency with an approximation error Δs 
Δfsupply /fsupply [34]. This relative error is equal to 0.004%
with fsupply = 50 Hz and Δfsupply = 2 mHz. The 20 periods
of the sampled stator current are enough to achieve these good
estimation results.
Then, ZFFT is used to check (1 ± 2 s)fsupply sideband
frequencies around the supply frequency. Specifically, the trans-
lation frequency fi and Bp are chosen so that

fi = fsupply − 3 s fsupply (8)

Bp = 6 s fsupply . (9)

Such a small zoom window is guaranteed by the accuracy of Fig. 2. MCFT and ZFFT flowchart for accurate fault detection.
MCFT. In fact, the estimation error ΔBp is remarkably lower
than the safety window taken

ΔBp = 6(fsupply Δs + sΔfsupply ) = 6Δfsupply (s + 1). (10)

Replacing (9) into (4)

6sfsupply N
B= . (11)
Fs
Hence
Fs
C= . (12)
12 sfsupply

C is about 670 in case Fs = 10 kHz, s = 0.025, fsupply =

50 Hz, and costgain = 1540, in case N = 100.000 (i.e., Tw =
10 s). Hence, with the previous parameters, it is possible to
achieve a reduction of the computational cost of 1540 or an
increase of resolution of 670, if and only if a time acquisition Fig. 3. Normalized cross-correlation function between the stator current and
period of 670 × Tw is available. Keeping the same resolution, a reference tone around a grid frequency of 50 Hz (1.7-MW 72-rotor-slot
the latency is reduced according to costgain = 1540. Specifi- induction machine).
cally, the reduced number of operations allows decreasing
the time slice required for polling operations. In this way, a even with very small values and a low number of samples.
processor with reduced processing power and then reduced cost The machine nameplate is 1700 kW, three phase, Nb = 74,
may be used to perform all the tasks. and P = 2. A phase current was sampled at 4096 Hz for
A block diagram of the procedure including all the main tasks Tw = 4 s. The grid frequency was estimated with a resolution
performed by the global frequency tracking algorithms is shown of Δfsupply = 2 mHz, and the higher rotor slotting frequency
in Fig. 2. A dedicated test bed is used to sample stator currents, was estimated with a different resolution of Δfslot = 20 mHz.
the MCFT is used to estimate fsupply and s, and then fi and A grid frequency of 50.8220 Hz, a higher rotor slotting
Bp are computed by (8) and (9). The ZFFT algorithm with frequency of 1296.88, and a slip of 0.602367% were obtained.
the aforementioned parameters allows one either to increase the The corresponding cross-correlation functions are reported
frequency resolution or to reduce the computational cost. for the grid (Fig. 3) and the rotor slotting frequencies (Fig. 4).
As an example, a large three-phase induction machine oper- The resolution of a few millihertz would be impossible with
ated as a fan has been used to verify the effectiveness of MCFT, traditional FFT techniques for which the frequency resolution
4204 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 12, DECEMBER 2008

Fig. 4. Normalized cross-correlation function between the stator current and Fig. 6. Normalized cross-correlation function between the stator current and
a reference tone around the rotor slotting frequency (1.7-MW 72-rotor-slot a reference tone around a grid frequency of 50 Hz (18.5-kW 40-rotor-slot
induction machine). induction machine in faulty conditions at a rated load).

used to define the reference spectrum while the latter has one
artificially built incipient rotor fault in one bar.
The four induction machines are stiffly connected on the
same shaft with a magnetic brake in order to simulate the
load, and each induction machine is tested separately from
the other one. Each induction machine has been operated at
constant volts-per-hertz ratio at a rated load. A power converter,
operating at a switching frequency of 10 kHz, generates the
stator voltages.
At first, the experiments referred to the large machine are
reported. The experiments are referred to a supply frequency of
50 Hz, and the corresponding slip is about 2.5%.
One induction machine phase current is measured by means
of an insulated sensor connected to the machine phase. The
sensor output is used as input of the signal conditioning and
the data acquisition board integrated into a personal computer
(PC). The current probe is built with simple Rogowski coil
Fig. 5. FFT spectrum of the stator current around the grid frequency for with a typical frequency bandwidth of 50 kHz. For the simple
a healthy machine (1.7-MW 72-rotor-slot induction machine; a resolution of spectrum computation using FFT, the stator current has been
0.25 Hz).
sampled at Fs = 10 kHz with a time acquisition period Tw =
10 s (frequency resolution: 0.1 Hz).
with a time acquisition period of 4 s would be bound to 0.25 Hz MCFT was used to track the supply frequency and the
(Fig. 5). induction machine slip. The algorithm is implemented on the
PC connected to the acquisition board. It can be implemented
offline with MATLAB or it can be implemented on the PC with
B. Experimental Results
a proper real-time acquisition and computation system. Here,
Two types of machines were used for the experiments. The the signal processing including the ZFFT has been implemented
first type of machine is an 18.5-kW, 50-Hz, 220-V/380-V, four- by using the LabVIEW environment, because it is well adapted
pole, Nb = 40 rotor bars, and three-phase squirrel-cage induc- to the data acquisition board.
tion machine. In order to check the rotor faults, two machines of The corresponding cross-correlation functions have been
this type were used, the former is healthy and used to define the computed for the induction machine in faulty conditions (three
reference spectrum while the latter has three artificially drilled broken bars) for the grid (Fig. 6) and the rotor slotting (Fig. 7)
contiguous broken rotor bars. frequencies. A spanning step of 2 mHz has been used for
The second type is a 1.8-kW, 50-Hz, 220-V/380-V, four-pole, the supply frequency, and this step has been fixed to 20 mHz
Nb = 26 rotor bars, and three-phase squirrel-cage induction for the rotor slotting frequency. As far as the healthy ma-
machine. This is a special machine with closed slot, and it can chine is concerned, a grid frequency of 50.002 Hz, a rotor
be used with two or three pole pairs. As before, two rotors are slotting frequency of 926.158 Hz, and a slip of 2.387% have
available to check the rotor faults. The former is healthy and been obtained. As for faulty conditions, a grid frequency of
BELLINI et al.: RESOLUTION TECHNIQUES FOR ROTOR FAULT DETECTION OF INDUCTION MACHINES 4205

Fig. 7. Normalized cross-correlation function between the stator current and Fig. 9. ZFFT spectrum of the stator current around the grid frequency for
a reference tone around the rotor slotting frequency (18.5-kW 40-rotor-slot the healthy machine (18.5-kW 40-rotor-slot induction machine; a resolution of
induction machine in faulty conditions at a rated load). 0.002 Hz).

Fig. 10. ZFFT spectrum of the stator current around the grid frequency for the
Fig. 8. Current waveforms before (solid line) and after (dashed line) the faulty machine (18.5-kW 40-rotor-slot induction machine, three broken bars,
demodulation procedure of ZFFT for the healthy machine. and a resolution of 0.002 Hz).

Fig. 9 shows the ZFFT spectrum of the processed stator

49.992 Hz, a rotor slotting frequency of 924.58 Hz, and a slip current for the machine in healthy conditions, and Fig. 10 shows
of 2.526% have been obtained. Of course, the changes are not the ZFFT spectrum for the machine in faulty conditions. The
significant, and they are due the small nonstationary conditions results of Fig. 10 are obtained by keeping the same frequency
even when any load is connected to the power grid. resolution of traditional FFT (i.e., the same acquisition period
It can be noted that, choosing a step Δf = 2 mHz for Tw = 10 s) and reducing the computational cost, as costgain is
the grid frequency, the MCFT states a resolution of 2 mHz about 1540.
independently of the adopted setup and the chosen acquisition On the other hand, keeping constant the computational cost
period. On the other hand, using DFT-based techniques, the and increasing the time acquisition period of a factor 670, the
same resolution can be achieved with a time acquisition period frequency resolution can be increased of a factor 670 as well.
of 1/2 mHz = 500 s. Then, the ZFFT is used to zoom in the Figs. 11 and 12 show the ZFFT spectrum of the processed
bandwidth around 50 Hz and to check the sideband frequency stator current in this case for the machine in healthy and faulty
(1 ± 2 s)fsupply for possible rotor faults. conditions, respectively.
Fig. 8 shows the current waveforms before and after Then, the experiments referred to the small machine are
the demodulation procedure of the ZFFT algorithm with reported. As stated before, this is a special machine, where
fi = 46.4214 Hz and Bp = 7.1613 Hz for the induction the number of poles can be changed and features closed slots.
machine in healthy conditions. For the faulty machine, the Hence, it is difficult to retrieve rotor slotting frequencies from
parameters are fi = 46.2036 Hz and Bp = 7.5768 Hz. the noise. However, the proposed procedure is still effective,
4206 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 12, DECEMBER 2008

Fig. 13. Normalized cross-correlation function between the stator current

and a reference tone around a grid frequency of 50 Hz (1.8-kW 26-rotor-slot
induction machine in faulty conditions at a rated load).

Fig. 11. ZFFT spectrum of the stator current around the grid frequency for
the healthy machine (18.5-kW 40-rotor-slot induction machine; a resolution of
0.002/670 Hz).

Fig. 14. Normalized cross-correlation function between the stator current

and a reference tone around the rotor slotting frequency (1.8-kW 26-rotor-slot
induction machine in faulty conditions at a rated load).

in healthy conditions. In the latter case, a grid frequency of

Fig. 12. ZFFT spectrum of the stator current around the grid frequency for the 50.000 Hz, a rotor slotting frequency of 582.6 Hz, and a slip
faulty machine (18.5-kW 40-rotor-slot induction machine, three broken bars, of 2.68% have been obtained.
and a resolution of 0.002/670 Hz).
Then, ZFFT was used to state the amplitude of the side-
band components. The parameters for the ZFFT algorithm are
and it can be applied using the nominal machine slip in fi = 45.9 Hz and Bp = 8.23 Hz for the induction machine
relationship (10). in faulty conditions, and fi = 45.98 Hz and Bp = 8.04 Hz in
The tests were made, relying on the aforementioned detailed healthy conditions.
test bed. The stator currents have been sampled at Fs = 10 kHz Fig. 15 shows the ZFFT spectrum of the processed stator
with a time acquisition period Tw = 10 s. The corresponding current for the machine in healthy conditions, and Fig. 16 shows
cross-correlation functions have been computed for the induc- the ZFFT spectrum for the machine in faulty conditions. The
tion machine in faulty conditions (incipient rotor fault) for results of Figs. 15 and 16 are obtained by keeping the same
the grid (Fig. 13) and the rotor slotting frequencies (Fig. 14). frequency resolution of traditional FFT (i.e., the same acquisi-
A spanning step of 2 mHz has been used for the supply tion period Tw = 10 s) and reducing the computational cost, as
frequency, and this step has been fixed to 20 mHz for the rotor costgain is about 1540. On the other hand, keeping constant the
slotting frequency. As far as the faulty machine is concerned, computational cost and increasing the time acquisition period
a grid frequency of 50.010 Hz, a rotor slotting frequency of of a factor 670, the frequency resolution can be increased of a
582.32 Hz, and a slip of 2.74% have been obtained. A rated load factor 670 as well. It must be noted that traditional FFT would
was used for the experiments. The same procedure was used not be able to state, with enough accuracy, the amplitudes of
to estimate grid frequency and machine slip for the machine the sideband components, as only a resolution of 0.1 Hz would
BELLINI et al.: RESOLUTION TECHNIQUES FOR ROTOR FAULT DETECTION OF INDUCTION MACHINES 4207

can be performed in real time, owing to the reduced computa-

tional cost achieved by ZFFT. Therefore, a reliable and efficient
monitoring method for induction machine stator currents can be
obtained even under weak time-varying conditions. The high
resolution achieved can be used to search exactly the fault
signatures in order to diagnose them even in the presence of
other nonfault-related harmonics, e.g., mechanical faults, with
similar frequencies.
The proposed method exploits MCFT to compute accurately
the frequency of the fault signature, without the fixed stationary
constraints of Fourier transform and at a reduced cost with the
same frequency resolution. Then, ZFFT computes the spectrum
in a narrow frequency bandwidth, optimizing the ratio between
performances and computational cost.
These features can be used in the monitoring of electrical
drives in industrial applications, where cost and reliability are
major concerns. The proposed method achieves a remarkable
reduction of computational cost and storage requirements, al-
Fig. 15. ZFFT spectrum of the stator current around the grid frequency for lowing an efficient implementation on low-cost fixed-point
the healthy machine (1.8-kW 26-rotor-slot induction machine; a resolution of digital signal processor that may be shared with the overall
0.1 Hz).
drive control system. Eventually, the latency of the monitoring
procedure can be reduced in order to obtain early warning to
prevent faults in electrical drives.

R EFERENCES
[1] M. Haji and H. Toliyat, “Pattern recognition—A technique for induction
machines rotor broken bar detection,” IEEE Trans. Energy Convers.,
vol. 16, no. 4, pp. 312–317, Dec. 2001.
[2] B. Yazici and G. Kliman, “An adaptive statistical time–frequency method
for detection of broken bars and bearing faults in motors using sta-
tor current,” IEEE Trans. Ind. Appl., vol. 35, no. 2, pp. 442–452,
Mar./Apr. 1999.
[3] S. H. Kia, A. M. Mabwe, H. Henao, and G.-A. Capolino, “Wavelet based
instantaneous power analysis for induction machine fault diagnosis,” in
Proc. 32nd Annu. Conf. IEEE Ind. Electron., IECON, Paris, France,
Nov. 2006, pp. 1229–1234.
[4] F. Briz, M. W. Degner, A. B. Diez, and J. M. Guerrero, “Online diag-
nostics in inverter-fed induction machines using high-frequency signal
injection,” IEEE Trans. Ind. Appl., vol. 40, no. 4, pp. 1153–1161,
Jul./Aug. 2004.
[5] T. Zhao and D. S. Zinger, “Induction motor speed detection by applying
Goertzel’s algorithm to current harmonics,” in Proc. 15th Annu. IEEE
Fig. 16. ZFFT spectrum of the stator current around the grid frequency for the APEC, New Orleans, LA, 2000, vol. 1, pp. 140–143.
faulty machine (1.8-kW 26-rotor-slot induction machine, incipient rotor fault, [6] H. Henao, C. Demian, and G. Capolino, “A frequency-domain detec-
and a resolution of 0.1 Hz). tion of stator winding faults in induction machines using an external
flux sensor,” IEEE Trans. Ind. Appl., vol. 39, no. 5, pp. 1272–1276,
Sep./Oct. 2003.
be available. This limitation would be enhanced in the case of [7] C. Demian, A. Mpanda-Mabwe, H. Henao, and G. Capolino, “Detection
reduced load, i.e., reduced machine slip values. of induction machines rotor faults at standstill using signals injection,”
IEEE Trans. Ind. Appl., vol. 40, no. 6, pp. 1550–1559, Nov./Dec. 2004.
[8] F. Cupertino, A. D. Vanna, L. Salvatore, and S. Stasi, “Analysis tech-
niques for detection of induction motor broken rotor bars after supply
IV. C ONCLUSION disconnection,” IEEE Trans. Ind. Appl., vol. 40, no. 2, pp. 526–533,
Mar./Apr. 2004.
The combined use of MCFT and ZFFT algorithms for the [9] D. Thomas and M. Woolfson, “Evaluation of frequency tracking
processing of induction machine stator currents has been pre- methods,” IEEE Trans. Power Del., vol. 16, no. 3, pp. 367–371,
Jul. 2001.
sented. It is shown how the joint use of the aforementioned [10] K. D. Hurst and T. G. Habetler, “Sensorless speed measurement using
methods can overcome the constraints of limited time or fre- current harmonic spectral estimation in induction machine drives,” IEEE
quency resolutions of traditional DFT-based method. After a Trans. Power Electron., vol. 11, no. 1, pp. 66–73, Jan. 1996.
[11] K. Hurst and T. Habetler, “A comparison of spectrum estimation tech-
short presentation of elements to help in the hardware imple- niques for sensor less speed detection in induction machine,” IEEE Trans.
mentation, the experiments confirm that the induction machine Ind. Appl., vol. 33, no. 4, pp. 898–905, Jul./Aug. 1997.
slip and the grid or power converter supply main frequency [12] P. Yip, “Some aspects of the zoom transform,” IEEE Trans. Comput.,
vol. C-25, no. 3, pp. 287–296, Mar. 1976.
can be tracked with high accuracy even with a very small time [13] E. Hoyer and R. Stork, “The zoom FFT using complex modulation,” in
acquisition period. Moreover, an efficient diagnostic procedure Proc. IEEE ICASSP, May 1977, vol. 2, pp. 78–81.
4208 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 12, DECEMBER 2008

[14] J. Zarader, M. Garnier, and M. Nicollet, “New frequency-filtering Alberto Bellini (S’96–A’98–M’03) was born in Italy
zoom,” Int. J. Adapt. Control Signal Process., vol. 6, pp. 547–560, in 1969. He received the Laurea degree (M.S.) in
Jun. 1992. electronic engineering and the Ph.D. degree in com-
[15] A. Yazidi, H. Henao, G. Capolino, M. Artioli, and F. Filippetti, “Improve- puter science and electronic engineering from the
ment of frequency resolution for three-phase induction machine fault University of Bologna, Bologna, Italy, in 1994 and
diagnosis,” in Conf. Rec. IEEE IAS Annu. Meeting, Oct. 2005, vol. 1, 1998, respectively.
pp. 20–25. From 1999 to 2004, he was with the University of
[16] M. Aiello, A. Cataliotti, and S. Nuccio, “An induction motor speed mea- Parma, Parma, Italy. He was an Honorary Scholar at
surement based on current harmonic analysis with chirp-z transform,” in the University of Wisconsin, Madison, during 2000.
Proc. IEEE Instrum. Meas. Technol. Conf., Budapest, Hungary, May 2001, Since 2004, he has been with the Department of
vol. 1, pp. 578–582. Engineering Sciences and Methods, University of
[17] M. E. H. Benbouzid and G. B. Kliman, “What stator current processing- Modena and Reggio Emilia, Reggio Emilia, Italy, where he is currently an
based technique to use for induction motor rotor faults diagno- Assistant Professor of electric machines and drives. His research interests in-
sis?” IEEE Trans. Energy Convers., vol. 18, no. 2, pp. 238–244, clude power electronics, signal processing for audio and industrial applications,
Jun. 2003. and electric drive design and diagnosis. He is the author or coauthor of more
[18] K. Kim and A. G. Parlos, “Induction motor fault diagnosis based on neu- than 80 papers and one textbook, and he is the holder of three industrial patents.
ropredictors and wavelet signal processing,” IEEE/ASME Trans. Mecha- Dr. Bellini was the recipient of the First Prize Paper Award from the Electric
tronics, vol. 7, no. 2, pp. 201–219, Jun. 2002. Machines Committee of the IEEE Industry Applications Society in 2001. He
[19] C. Combastel, S. Lesecq, S. Petropol, and S. Gentil, “Model-based and serves as an Associate Editor of the IEEE TRANSACTIONS ON INDUSTRIAL
wavelet approaches to induction motor on-line fault detection,” Control ELECTRONICS.
Eng. Pract., vol. 10, no. 5, pp. 493–509, May 2002.
[20] H. Douglas, P. Pillay, and A. K. Ziarani, “A new algorithm for transient
motor current signature analysis using wavelets,” IEEE Trans. Ind. Appl.,
vol. 40, no. 5, pp. 1361–1368, Sep./Oct. 2004.
[21] L. Eren and M. J. Devaney, “Motor bearing damage detection via wavelet
analysis of the starting current transient,” in Proc. IEEE IMTC, 2001,
pp. 1797–1800.
[22] S. Rajagopalan, J. Aller, J. Restrepo, T. Habetler, and R. Harley, “Detec-
tion of rotor faults in brushless DC motors operating under nonstation- Amine Yazidi (A’06) received the M.S. degree in
ary conditions,” IEEE Trans. Ind. Appl., vol. 42, no. 6, pp. 1464–1477, electrical engineering and industrial informatics and
Nov./Dec. 2006. the Ph.D. degree in electrical engineering from the
[23] S. Rajagopalan, T. G. Habetler, R. Harley, and J. Aller, “Non- University of Picardie Jules Verne, Amiens, France,
stationary motor fault detection using recent quadratic time–frequency in 2003 and 2006, respectively, and the M.S. degree
representations,” in Conf. Rec. IEEE IAS Annu. Meeting, Oct. 2006, in perception and information analysis from the Uni-
pp. 2333–2339. versity of Rouen, Rouen, France.
[24] F. Filippetti, G. Franceschini, C. Tassoni, and P. Vas, “Recent develop- Since 2007, he has been with the University of
ments of induction motor drives fault diagnosis using AI techniques,” Picardie Jules Verne as an Associate Professor in the
IEEE Trans. Ind. Electron., vol. 47, no. 5, pp. 994–1004, Oct. 2000. Department of Electrical Engineering. His research
[25] B. Ayhan, M.-Y. Chow, and M.-H. Song, “Multiple discriminant analysis interests are diagnosis, modeling, and remote moni-
and neural-network-based monolith and partition fault-detection schemes toring of electrical machines, drives, and, specially, wind turbine generators,
for broken rotor bar in induction motors,” IEEE Trans. Ind. Electron., for which he is developing methods to detect electrical faults using noninvasive
vol. 53, no. 4, pp. 1298–1308, Jun. 2006. sensors. He has worked in collaboration with the Department of Electrical
[26] C. Lo, Y. Wong, and A. Rad, “Intelligent system for process supervi- Engineering, University of Bologna, Bologna, Italy, since 2002.
sion and fault diagnosis in dynamic physical systems,” IEEE Trans. Ind. Prof. Yazidi was the President of the Jules Verne IEEE student branch. He
Electron., vol. 53, no. 2, pp. 581–592, Apr. 2006. has been a member of the Scientific Council of the University of Picardie Jules
[27] W. W. Tan and H. Huo, “A generic neurofuzzy model-based approach for Verne since 2004.
detecting faults in induction motors,” IEEE Trans. Ind. Electron., vol. 52,
no. 5, pp. 1420–1427, Oct. 2005.
[28] H. Su and K. T. Chong, “Induction machine condition monitoring using
neural network modeling,” IEEE Trans. Ind. Electron., vol. 54, no. 1,
pp. 241–249, Feb. 2007.
[29] M. Ballal, Z. Khan, H. Suryawanshi, and R. Sonolikar, “Adaptive neural
fuzzy inference system for the detection of inter-turn insulation and bear-
ing wear faults in induction motor,” IEEE Trans. Ind. Electron., vol. 54,
no. 1, pp. 250–258, Feb. 2007. Fiorenzo Filippetti (M’00) was born in Fano, Italy,
[30] J. Martins, V. F. Pires, and A. Pires, “Unsupervised neural-network- in 1945. He received the M.S. degree in elec-
based algorithm for an on-line diagnosis of three-phase induction motor trical engineering from the University of Bologna,
stator fault,” IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 259–264, Bologna, Italy, in 1970.
Feb. 2007. In 1976, he was with the Department of Electrical
[31] J.-H. Jung, J.-J. Lee, and B.-H. Kwon, “Online diagnosis of induction Engineering, University of Bologna, as an Assistant
motors using MCSA,” IEEE Trans. Ind. Electron., vol. 53, no. 6, Professor, where he is currently a Full Professor
pp. 1842–1852, Dec. 2006. of electrical drives. From 1993 to 2002, he was an
[32] A. Mohanty and C. Kar, “Fault detection in a multistage gearbox by Adjoint Professor of electrotechnics and electrical
demodulation of motor current waveform,” IEEE Trans. Ind. Electron., drives with the University of Parma, Parma, Italy. He
vol. 53, no. 3, pp. 1285–1297, Jun. 2006. spent several visiting periods at the Centre de Gnie
[33] S. Bachir, S. Tnani, J.-C. Trigeassou, and G. Champenois, “Diagnosis Electrique de Lyon (CEGELY), University Claude Bernard, Lyon, France, and
by parameter estimation of stator and rotor faults occurring in induc- at the University of Picardie Jules Verne, Amiens, France. During 1998, he
tion machines,” IEEE Trans. Ind. Electron., vol. 53, no. 3, pp. 963–973, held a position at the University Claude Bernard as a Member of the Scientific
Jun. 2006. Council of CEGELY. He is a Lecturer for the European Master in Advanced
[34] A. Bellini, G. Franceschini, and C. Tassoni, “Monitoring of induction Power Electrical Engineering recognized by the European Commission in 2004.
machines by maximum covariance method for frequency tracking,” IEEE He has been the Advisor of various Ph.D. and M.S. students. He is involved
Trans. Ind. Appl., vol. 42, no. 1, pp. 69–78, Jan./Feb. 2006. in national and European research projects. He has authored or coauthored
[35] A. Bellini, F. Filippetti, D. Casadei, A. Yazidi, and G. Capolino, “Monitor- more than 160 scientific papers published in scientific journals and conference
ing of induction machines currents by high frequency resolution analysis,” proceedings since 1976, and one textbook. He is the holder of an industrial
in Conf. Rec. IEEE IAS Annu. Meeting, Oct. 2006, pp. 2320–2325. patent. His main research interests include the simulation and modeling of
[36] A. V. Oppenheim, R. Schafer, and J. Buck, Discrete-Time Signal electric circuits and systems, and the study and application of condition
Processing. Englewood Cliffs, NJ: Prentice-Hall, 1999. monitoring and fault detection techniques for ac electrical machinery.
BELLINI et al.: RESOLUTION TECHNIQUES FOR ROTOR FAULT DETECTION OF INDUCTION MACHINES 4209

Claudio Rossi (M’05) was born in Forlì, Italy, in Gérard-André Capolino (A’77–M’82–SM’89–
1971. He received the M.Sc. degree in electrical en- F’02) was born in Marseille, France. He received
gineering and the Ph.D. degree from the Department the B.Sc. degree in electrical engineering from
of Electrical Engineering, University of Bologna, Ecole Centrale de Marseille, Marseille, in 1974,
Bologna, Italy, in 1997 and 2001, respectively. the M.Sc. degree from Ecole Suprieure d’Electricit,
Since 2000, he has been with the Department Paris, France, in 1975, the Ph.D. degree from the
of Electrical Engineering, University of Bologna, University Aix-Marseille I, Marseille, in 1978, and
where he currently is an Assistant Professor the D.Sc. degree from the Institut National Poly-
of electrical machines, drives, and power elec- technique de Grenoble, Grenoble, France, in 1987.
tronics. His current research activity is devoted to Since 1994, he has been with the University of
power electronics and drives for electric vehicles and Picardie “Jules Verne,” Amiens, France, where he is
renewable energy systems. a Full Professor in the Department of Electrical Engineering and currently the
Dr. Rossi is a Registered Professional Engineer in Italy. Director of the European Master in Advanced Power Electrical Engineering
recognized by the European Commission in 2004. His research interests are
mainly on modeling and control of induction machines and on the condi-
tion monitoring and fault detection of ac electrical machinery. He has pub-
lished more than 380 papers in scientific journals and conference proceedings
since 1975.
Dr. Capolino is the recipient of the IEEE-IES Dr. Ing. Eugene Mittlemann
Achievement Award for 2008. He has been involved in many IEEE activities in
the France Section (Past Chair), the Region 8 (Chapter Subcommittee Chair),
Power Electronics Society (Associate Editor; Symposium on Diagnostics for
Electric Machines, Power Electronics and Drives founder), and Industrial
Electronics Society (Associate Editor, Vice-President for Membership, and
Distinguished Lecturer). He is the acting Chair for the steering committee of
the International Conferences on Electrical Machines.