Methods to Identify Intermittent Electrical and
Mechanical Faults in Permanent Magnet
AC Drives Based on Wavelet Analysis
Wesley G. Zanardelli and Elias G. Strangas
Department of Electrical and Computer Engineering
Michigan State University
East Lansing, MI 48824, USA
e-mail: zanardel@egr.msu.edu
Abstract Prognosis of failures of electric drives can be
achieved through the detection of non-catastrophic faults. As
the frequency and severity of these faults increase, the expected
working life of the drive decreases, leading to eventual failure. In
this work, a method is presented to identify developing electrical
and mechanical faults based on analysis of the Undecimated
Discrete Wavelet Transform of the field oriented currents in
PMAC drives. Linear discriminant analysis is used to classify
between the fault types.
I. I NTRODUCTION
Because of the small size, high power density and low
audible noise of PMAC machines, they are strong candidates
for use in X-by-wire and other systems in vehicles. As these
systems replace more critical vehicle functions, improved
methods to predict drive failures will be required. Prior to
failure, the drive can intermittently show signs of a developing
fault, while it is operating normally.
Methods to detect a variety of electrical drive faults were
discussed in [1][8]. These include open and shorted switches
[1], open phase, shorting of adjacent commutator bars, and
worn brushes [2], turn-turn shorts [3], [8], voltage unbalance
between phases [4], [8], broken rotor bars and end rings [5],
[6], [8], and increased stator temperature [7].
Methods to detect mechanical faults were discussed in [7]
[11]. These include increased friction [7], static and dynamic
eccentricities [8], [9], gear faults [10], and fracture of a tool
coupled to the machine [11].
A number of signal processing techniques are applied to
the machine voltages, currents, speed, and vibration including
filtering [3], Fourier analysis [5], [9], and wavelet analysis
[10]. Categorization techniques rely on thresholding [3], [5],
[9], [11], Artificial Neural Networks (ANN) [2], [4], [10],
model-based techniques [2], [7], artificial intelligence [6], and
Support Vector Machine (SVM) analysis [8]. Some of the
above work uses Finite Element Analysis (FEA) to collect data
for algorithm design [6], to simulate diagnostic techniques [8],
and to determine machine parameters for simulation [9].
In our previous work, a method for failure prognosis based
on wavelet analysis was developed for brush DC motors
[12]. The modulus maxima of the Discrete Wavelet Transform
(DWT) of the DC current was analyzed. An algorithm to
detect the presence of a fault was based on thresholding of
the analysis coefficients. If the detection criterion was met, the
coefficients were passed to a classification algorithm. Three
classification algorithms were evaluated. The first was based
on a decision tree, the second was based on the nearest
neighbor rule, and the third was based on linear discriminant
analysis. Classification was possible between DC currents in
machines with the following faults: increased coil resistance,
increased friction, faulty brush springs, rotor misalignment,
and damaged commutator face.
Next, methods for failure prognosis method based on both
Fourier and wavelet analysis were developed for PMAC drives
[13], [14]. Two electrical faults were investigated. The first
was a momentary increased resistance in one phase due to a
bad connection between the motor and the controller and the
second was a turn-to-phase short in the stator windings. Analy-
sis was performed on the bandlimited Short-Time Fourier
Transform (STFT) and the Undecimated Discrete Wavelet
Transform (UDWT) of the torque producing component of
the field oriented stator currents. First, a detection algorithm
was based on thresholding of the energy in each sample. If the
detection criterion was met, the coefficients were passed to a
classification algorithm based on linear discriminant analysis
to discriminate between the two faults.
The present work is a continuation and extension to the
methods developed in [13], [14] for mechanical faults. While
our previous techniques were able to recognize the incep-
tion and clearing of electrical faults separately, the method
presented here is designed to recognize a fault event in its
entirety. While it was not the case for the electrical faults
explored, most of the fault information for the mechanical
faults explored lies between the inception and clearing of the
fault. For electrical faults, this technique is best suited for
faults with a known time duration.
Methods to identify intermittent stator winding and gear
faults in PMAC drives are presented. The faults of interest are
non-catastrophic, meaning they allow for continued operation
of the drive, but increase the likelihood of failure. Early
detection of these faults can alert the operator to schedule
maintenance of the drive before failure occurs. h1 ( - n ) 2 dj
Two types of electrical stator faults and one type of mechan-
c j +1 h1 ( - n ) 2 d j- 1
ical gear fault with varying degrees of severity are explored.
The first electrical fault is a momentary increased resistance in h0 ( - n ) 2 cj
one phase due to a bad connection between the motor and the h0 ( - n ) 2 c j- 1
controller. The second electrical fault is a turn-to-phase short,
simulating an insulation failure in the stator windings of the Fig. 1. Two-Stage DWT Filter Bank Analysis Tree
motor. The mechanical fault is missing gear teeth in a gearbox
coupled to the motor. The methods introduced are designed to Scale
detect the short transients, which over time, will develop into
n: h (0) h (1) h (2) h (3)
a fault, e.g. momentary increased resistance. i i i i

The algorithm is based on analysis of the UDWT of the

torque producing component of the field oriented stator cur- n-1: h (0) 0 h (1) 0 h (2) 0 h (3)
i i i i

rents. Unlike the STFT or DWT, the UDWT is shift-invariant

i.e., the UDWT of a signal is independent of the time that
sampling begins. Thresholding on the energy in the UDWT is n-2: h (0)
i
0 0 0 h (1)
i
0 0 0 h (2)
i
0 0 0 h (3)
i

used to detect a fault in the machine, and linear discriminant

analysis is used to classify between the fault types.
A collection of observed data is used to train the detection
and classification components of the algorithm. Results ob-
tained using both FEA and simulation were considered for use
in algorithm development, however, while these techniques can
be used to approximate the effect of a fault, it is not possible
to model all of the dynamics in an electric drive coupled to a
load. Results obtained from laboratory experiments proved to
be more useful.
II. T HEORETICAL BASIS
A. Undecimated Discrete Wavelet Transform
The Fourier Transform gives the spectrum of a signal.
It is best suited for the analysis of stationary signals, or
signals whose spectrum remains constant. The FFT is used
to determine the spectrum of discrete-time signals. In the case
of the FFT, the signal is divided into several frequency bands,
however no information is present on the time axis. The faults
studied in this work manifest themselves as short transients
superimposed on the stator currents. Analysis of these short
transients, however, requires information in both frequency
and time. The inability to provide time localization of a signal
is a fundamental limitation of the FFT.
The STFT [15] is an extension of the FFT, allowing for the
analysis of non-stationary signals. Here, the signal is divided
into small time sections, and each is analyzed using the FFT.
The results for of the STFT are intuitive and easy to correlate
with the original signal. In the case of the STFT, it can be
seen how the spectrum of a signal changes with time. In the
implementation of the STFT, a design tradeoff must be made
between time and frequency resolution. This is due to the
uncertainty principle, which limits the lower bound on the
time-bandwidth product (1).
1 also contributes to additional end effects.
TB (1)
2 An example to show the shift-invariant property of the
Wavelet analysis [16] is also suitable for non-stationary
signals. The DWT has greater flexibility than the STFT.
Different basis functions, or mother wavelets, may be used in
Fig. 2. UDWT Filters Modified by the Algorithme a Trous
Wavelet analysis while the basis function for Fourier analysis
is always the sinusoid. Unlike sinusoids, wavelets have finite
energy concentrated around a point in time. One can choose,
or design a wavelet to achieve the best results for a specific
application.
Unlike Fourier methods, the DWT allows for both good time
resolution of high frequency components, and good frequency
resolution of low frequency components in the same analysis.
A block diagram for the DWT algorithm using filter banks
is shown in Fig. 1, where c and d are the scaling and wavelet
function coefficients, and h1 and h0 are the decomposition
highpass and lowpass filters.
One of the drawbacks of the DWT is that it is not a
shift-invariant transformation. This makes pattern recognition
problems based on the DWT more difficult, since the DWT
coefficients resulting from decomposition of a signal and a
shifted version of the signal can be very different. Only in
special cases, where the signal is shifted by specific powers
of 2 will the outputs be shifted versions of one another.
The UDWT adds the property of shift-invariance to the
DWT. Here, the downsampling step is omitted from the
DWT algorithm, and zeros are inserted between the filter
coefficients at each successive scale, as shown in Fig. 2. This is
known as the Algorithme a Trous [17], [18]. While circular
convolution is used in [17], linear convolution is used here,
since the signals are not periodic. The coefficients at each
end of the convolution where the filter and the signal are not
completely overlapping, known as end effects, are replaced
by zeros. As these zeros propagate through each scale of the
UDWT, the number of end effect coefficients increases. The
fact that the length of the wavelet filters increase at each scale
UDWT is in Fig. 3. Here, the DWT and UDWT of a signal and
the same signal shifted by 6 samples are compared. The first
row shows the non-shifted and shifted signals. The second and
 28
Original Signal
28
Original Signal (Shifted by 6 Samples)
knowledge of a probability distribution among the sample
26 26 points is assumed. The space is divided into K regions, each
Amplitude

Amplitude
24 24
having its own weighting coefficients. In this work, we use
22 22
20 20
linear discriminant functions (2) [19],
18 18
Dk (x) = x1 1k + x2 2k +, . . . , +xN N k + N +1,k
DWT of Original Signal DWT of Original Signal (Shifted by 6 Samples) (2)
4 4
k = 1, 2, . . . , K
3 3
where x is the N -dimensional sample vector and are the
Scale

Scale
2 2
normalized weighting coefficients for the k-th class. Linear
1 1
discriminant functions were chosen for the algorithm since
UDWT of Original Signal UDWT of Original Signal (Shifted by 6 Samples) they are the most computationally efficient form. A sample
4 4 vector belongs to a particular class if its discriminant function
3 3 is greater for that class than for any other class, i.e., xi belongs
Scale

Scale

2 2 to class Cj if
1 1

100 120 140 160 100 120 140 160 Dj (x) > Dk (x) for every k 6= j.
Sample Number Sample Number

Fig. 3. Example of Shift-Invariance of the UDWT The weighting coefficients are adjusted from their initial
guess through a training procedure using sample vectors which
UDWT the proper classification is known. The algorithm for this
d6,1 d6,2 d6,N procedure makes adjustments to the weighting coefficients
d5,1 d5,2 d5,N until each sample vector is correctly classified.
el d4,1 d4,2 d4,N Young and Calvert [19] show that this training algorithm
ac
Sd d d3,N will converge in a finite number of steps. When a sample
3,1 3,2
d2,1 d2,2 d2,N vector is correctly classified, no adjustment to the weighting
d1,1 d1,2 d1,N coefficients is made. When a sample vectors is incorrectly
classified, or
Time (samples)
Dj (x) Dl (x),

where
d6,1 d5,1 d4,1 d3,1 d2,1 d1,1 d6,2 d5,2 d4,2 d3,2 d2,2 d1,2 d6,N d5,N d4,N d3,N d2,N d1,N Dl (x) = max [D1 (x), . . . , DK (x)] ,
l6=j

Fig. 4. Vectorization of N Samples of the UDWT adjustments are made to j (3) and l (4) only,

j (i + 1) = j (i) + axi (3)

third rows show the DWT and UDWT of these signals. The
value of the wavelet coefficients is denoted by shading, with l (i + 1) = l (i) axi , (4)
the darkest coefficients having the highest amplitude. In the
case of the DWT, it can be seen that the distribution of energy where a is a gain constant.
in the wavelet coefficients is different for the non-shifted and
shifted signals. For the UDWT, the distribution of energy is
III. FAULTS E XPLORED
consistent for both signals, with the output coefficients shifted
by the same number of samples as the input signal. The test machine used in this analysis is a six-pole surface
mounted PMAC machine for an automotive application. The
B. Pattern Recognition machine is operated in a vector drive with the torque angle set
Once the UDWT coefficients have been determined, the to /2 [20]. This mode of operation minimizes losses in the
presence of a fault is detected based on thresholding of the machine, and is suitable for operating speeds up to the base
energy in the coefficients. The next step in the fault prediction speed. The torque-producing component of the stator current
process is to vectorize the coefficients, allowing for the use command was iqs = 0.3pu, and the flux-weakening compo-
of discriminant analysis for classification. The vectorization nent ids = 0; the speed is controlled by the dynamometer and
process combines columns of the coefficients into a vector. is set to 1/10 the no-load speed.
Vectorization of the data begins when the criterion for detec- The electrical faults imposed in this work are not periodic.
tion is met. An example of vectorization using N time samples The time interval between successive faults is random. In
is shown in Fig. 4. the case of the missing gear teeth, the fault occurs once per
In the implementation of discriminant functions, no prior mechanical revolution.
A. Series Resistance
The first fault explored in this work is an intermittently
increased contact resistance between the motor and the con-
troller. An experiment was designed to simulate this. The fault
is achieved by adding the parallel combination of a normally
closed switch, and a resistance in series with one of the motor
phases. The fault is initiated by opening the switch for a short
time interval, causing current to flow through the resistance.
The value of the series resistance is approximately ten
times the value of the stator resistance, Rs . The fault is
initiated as the phase current command rises to 95% of its
peak amplitude. Tests with fault durations of 5ms and 10ms
have been performed to investigate invariance of the results
with respect to this parameter.
B. Turn-To-Phase Short
The second fault explored in this work is an insulation
failure in the stator windings of the motor. The machine used
in this experiment has multiple parallel stranded-wire windings
per phase, each with several coils in series. To simulate this
fault in the experimental setup, at one point in a single strand
of one of the windings, the insulation was removed, and a
normally open switch was added between this point and the
corresponding phase terminal of the motor. The fault was
initiated by momentarily closing the switch, causing current
to be split between the intended path and the switch.
The fault is initiated as the phase current command rises to
95% of its peak amplitude. Tests with fault durations of 5ms
and 10ms have been performed to investigate invariance of
the results with respect to this parameter.
C. Missing Gear Teeth
The third fault explored in this work is missing gear teeth in
a gearbox coupled to the motor. The gearbox has two identical
spur gears supported by a set of bearings. One of the gears
is free of faults, and the following conditions of the mating
gear have been tested: Healthy; one tooth missing; one tooth
missing and its adjacent tooth severely worn; and two adjacent
teeth missing. Tests have been performed to recognize the fault
and determine the number of missing teeth.
IV. FAULT D ETECTION AND C LASSIFICATION
The detection and classification of the faults described in
Section III are based on analysis of the stator currents of the
machine. Rather than analyzing the three phase stator currents
independently of each other, the field oriented currents iqs
and ids are used. This has the advantage that the fundamental
electrical frequency is not present. Consequently, rotor speed
has little effect on the spectrum of these currents, allowing
for invariance in the algorithm to rotor speed. Together, iqs
and ids are a complete representation of the stator currents,
however, it has been determined experimentally that through
analysis of iqs only, accurate fault detection and classification
can be achieved.
The input to the detection and classification algorithm is
the UDWT of the measured q-axis current, iqs . For this
analysis, the Daubechies D4 mother wavelet was used and
decomposition was performed to 6 levels. In the case of the
electrical faults, 20ms of data is analyzed corresponding to
400 samples of the UDWT. In the case of the mechanical
faults, 6.4ms of data is analyzed corresponding to 128 samples
of the UDWT.
The algorithm has two parts; a detection phase and a
classification phase. The detection phase of the algorithm is
based on thresholding of the energy in the UDWT. In the case
of the electrical faults, the energy is weighted, giving more
emphasis to the high frequency information in the current as
defined in (5).
6
X
E(x) = dn,x 2 2n1 (5)
n=1
The threshold was set at 12% greater than the largest which
was observed in all samples from the healthy machine data.
For the mechanical faults, the energy is not weighted, and
is defined in (6).
X6
E(x) = dn,x 2 (6)
n=1
The threshold was set at 80% greater than the largest which
was observed in all samples from healthy gears or with
one missing tooth. Based on the contact ratio of the gears
chosen for this project, the difference in the current between
the healthy case and the case with one missing tooth was
insignificant. If the energy in new test data exceeds these
thresholds, a fault is considered to exist.
The classification phase was based on linear discriminant
analysis and is implemented when the criterion for detection
is met. For the electrical fault experiments, data from 32
experiments with both 5ms and 10ms switch durations were
used to train the algorithms for each fault. For the mechanical
fault experiments data from 6 experiments, corresponding to
16 fault events, were used.
Following the training of the weighting coefficients, for the
electrical experiments, two new data sets from each of the fol-
lowing operating conditions were tested: Healthy, 5ms series
resistance, 10ms series resistance, 5ms turn-to-phase short,
and 10ms turn-to-phase short. The faulty operating conditions
are discussed in Section III, and 5ms and 10ms denotes the
duration of the fault. For the mechanical experiments, one
new data set from each of the following operating conditions
were tested: One tooth missing and its adjacent tooth severely
worn; and two adjacent teeth missing. Each of these data sets
contained three fault events.
FEA results using a commercial package, Flux2D, were
considered for use in training the detection and classification
algorithms, however the spectrum of the preliminary FEA
results was significantly different than that of the experimental
results. The use of FEA would allow for increased flexibility,
giving more training samples, for the faults explored. Fault
detection approaches based on results from FEA were intro-
duced in [6], [9], however the faults of interest were limited
to low-frequency phenomena (i.e. broken rotor bars, rotor
 Xilinx (QEP, A/D,
PWM, D/A, DIO)
2.8
RT-Linux GHz
PIV
Safety
Switch Inverter Encoder
12V
Automobile Electronic
Battery Switch
Fig. 5. Experimental Setup Block Diagram
Mounted Bearings Gears
with Set Screw
Shaft
to Motor
Solid Coupling
Fig. 6. Gearbox Drawing
eccentricities, and defective magnets), compared to the faults
studied in this work.
V. E XPERIMENTAL S ETUP
A block diagram for the experimental setup is shown in Fig.
5. A PC running RT-Linux was used as the controller for this
project connected via a custom Xilinx FPGA based I/O board
to the inverter and sensors. Two phase currents were measured
using current transducers with rated accuracy of 0.45% and
bandwidth of 0 200kHz. A single line-line voltage for use
in the initial position calibration was measured as well.
A quadrature encoder with 1024 counts per revolution (4096
for quadrature) and an index pulse was used to measure the
rotor position.
A bi-directionally conducting electronic switch was de-
signed to initiate electrical faults in the stator. This switch,
as shown by the dotted lines in Fig. 5, can be configured to
initiate either the series resistance fault or the turn-to-phase
short. The resistance of the switch is approximately one-third
the resistance of a single turn of a single strand in one of the
stator coils. The switch can be controlled using digital outputs
or PWM to create a resistance profile.
A gearbox was constructed to implement gear tooth faults.
The gearbox has two identical spur gears supported by a set of
bearings. The gears have the following parameters: Pressure
angle=14.5; number of teeth=36; pitch diameter=2.25; outer
diameter=2.38; and bore=0.5. The resultant contact ratio is
8.007. One of the gears is free of faults, and the mating gear
can be replaced with gears having varying numbers of missing
teeth. A drawing of the gearbox is shown in Fig. 6.
VI. R ESULTS
The results from two typical test cases for the electrical
faults are shown in Fig. 7. On the left side of Fig. 7 is the
UDWT for each fault and on the right side are the output
of the detection and classification algorithm with 0=healthy,
PMAC Drive
1=series resistance fault, and 2=turn-to-phase short. The ten
tests described in Section IV resulted in no false positives.
Flexible
Mechanical Each fault was identified correctly in all eight tests.
Coupling
The results from two test cases for the mechanical faults
PMAC PMAC are shown in Fig. 8. On the left side of Fig. 8 is the UDWT
Motor Dyno
for each fault and on the right side are the output of the
detection and classification algorithm with 0=healthy / one
Current Transducer:
tooth missing, 1=one tooth missing and its adjacent tooth
Voltage Transducer: severely worn, and 2=two adjacent teeth missing. The two
tests described in Section IV resulted in no false positives.
All six fault events were identified correctly in both tests.
to Dynamometer VII. C ONCLUSIONS
Solid Coupling
An algorithm capable of identifying intermittent electrical
and mechanical faults in PMAC drives has been developed. It
can be used to give a prognosis for failure of the drive. It is
based on the detection and classification of small transients in
the stator currents corresponding to non-catastrophic faults.
Unlike prior methods, the method proposed here is capable
of identifying both electrical and mechanical faults. This is
made possible by analysis of detected fault events in their
entirety. Early detection of these faults can give indication
when maintenance or mitigation is required, minimizing the
likelihood of system failure. This work addresses the limitation
of existing methods which can only detect permanent faults in
drives.
The UDWT, as used in this work, is shift-invariant and
offers good time resolution of high frequency components
and good time and frequency resolution of low frequency
components in the same analysis. The UDWT allows for
increased flexibility and resolution when compared to the more
traditional FFT and STFT, and overcomes the drawbacks of the
DWT. No other related work has been found which overcomes
the non-shift-invariant nature of Fourier and standard wavelet
methods.
The categorization algorithm uses a linear discriminant
function and is trained using a set of operating conditions
which include healthy drives and samples of faulted drives.
An exhaustive set of such conditions is necessary to develop
a robust algorithm.
Although the algorithms in this work were used offline, they
can be added to an existing motor controller enabling them
to run close to real-time. The vector currents are typically
calculated as part of the control. Minimal, if any, additional
CPU speed and memory capacity would be required, ensuring
a low-cost system. The training phase of the algorithms would
remain offline.
ACKNOWLEDGMENT
The authors thank Drs. Tomy Sebastian, Sayeed Mir, Mo-
hammad Islam, and Avoki Omekanda from Delphi Corpora-
tion for their continued guidance and support in this research
project.
 Classification (0=Healthy, 1=Series
Series Resistance (10ms) Resistance, 2=TurnPhase Short)

6 2
Level 5
4
1
3
2
1 0
Classification (0=Healthy, 1=Series
TurntoPhase Short (10ms) Resistance, 2=TurnPhase Short)

6 2
5
Level

4
1
3
2
1 0
0.1s 0.15s 0.2s 0.25s 0.3s 0.1s 0.15s 0.2s 0.25s 0.3s
Fig. 7. Typical Results for Electrical Faults

Classification (0=Healthy / 1 Missing Tooth,

1.5 Missing Gear Teeth 1=1.5 Missing Teeth, 2=2 Missing Teeth)

6 2
5
Level

4
1
3
2
1 0
Classification (0=Healthy / 1 Missing Tooth,
2 Missing Gear Teeth 1=1.5 Missing Teeth, 2=2 Missing Teeth)

6 2
5
Level

4
1
3
2
1 0
2.1s 2.2s 2.3s 2.4s 2.1s 2.2s 2.3s 2.4s
Fig. 8. Results for Mechanical Faults

R EFERENCES using MSCA, IEEE Transactions on Industry Applications, vol. 38,

[1] N. Bianchi, S. Bolognani, and M. Zigliotto, Analysis of PM syn-
chronous motor drive failures during flux weakening operation, in
27th Annual IEEE Power Electronics Specialists Conference, vol. 2,
pp. 15421548, Jun 1996.
[2] X. Q. Liu, H. Y. Zhang, J. Liu, and J. Yang, Fault detection and
diagnosis of permanent-magnet DC motor based on parameter estimation
and neural network, IEEE Transactions on Industrial Electronics,
vol. 47, pp. 10211030, Oct 2000.
[3] A. Cash, T. G. Habetler, and G. B. Kliman, Insulation failure prediction
in AC machines using line-neutral voltages, IEEE Transactions on
Industry Applications, vol. 34, pp. 12341239, Nov.Dec. 1998.
[4] H. Nejjari and M. E. Benbouzid, Monitoring and diagnosis of induction
motors electrical faults using a current parks vector pattern learning ap-
proach, IEEE Transactions on Industry Applications, vol. 36, pp. 730
735, MayJun. 2000.
[5] A. Bellini, F. Filippetti, G. Franceschini, C. Tassoni, R. Passaglia,
M. Saottini, G. Tontini, M. Giovannini, and A. Rossi, On-field ex-
perience with online diagnosis of large induction motors cage failures
pp. 10451053, JulAug 2002.
[6] C. C. Yeh, B. Mirafzal, R. J. Povinelli, and N. A. O. Demerdash,
A condition monitoring vector database approach for broken bar fault
diagnostics of induction machines, in International Electric Machines
and Drives Conference, pp. 2934, May 2005.
[7] O. Moseler and R. Isermann, Application of model-based fault detection
to a brushless DC motor, IEEE Transactions on Industrial Electronics,
vol. 47, pp. 10151020, Oct 2000.
[8] S. Poyhonen, M. Negrea, P. Jover, A. Arkkio, and H. Hyotyniemi,
Numerical magnetic field analysis and signal processing for fault
diagnostics of electrical machines, in International Conference on
Electrical Machines, Aug. 2002.
[9] W. le Roux, R. G. Harley, and T. G. Habetler, Rotor fault analysis of a
permanent magnet synchronous machine, in International Conference
on Electrical Machines, 2002.
[10] B. A. Paya, I. I. Esat, and M. N. M. Badi, Artificial neural network
based fault diagnostics of rotating machinery using wavelet transforms
as a preprocessor, Mechanical Systems and Signal Processing, vol. 11,
pp. 751765, Sep. 1997.
[11] H. S. Liu, B. Y. Lee, and Y. S. Tarng, Monitoring of drill fracture from
the current measurement of a three-phase induction motor, International
Journal of Machine Tools and Manufacture, vol. 36, pp. 729738, Jun
1996.
[12] W. G. Zanardelli, E. G. Strangas, H. K. Khalil, and J. M. Miller,
Wavelet-based methods for the prognosis of mechanical and electrical
failures in electric motors, Mechanical Systems and Signal Processing,
pp. 411426, Mar. 2005.
[13] W. G. Zanardelli and E. G. Strangas, Failure prognosis for permanent
magnet AC machines based on time-frequency analysis, in International
Conference on Electrical Machines, Sep. 2004.
[14] W. G. Zanardelli, E. G. Strangas, and S. Aviyente, Failure prognosis
for permanent magnet AC drives based on wavelet analysis, in Interna-
tional Electric Machines and Drives Conference, pp. 6470, May 2005.
[15] L. Cohen, Time-Frequency Analysis. Prentice Hall, 1995.
[16] C. S. Burrus, R. A. Gopinath, and H. Guo, Introduction to Wavelets and
Wavelet Transforms, A Primer. Prentice Hall, 1998.
[17] S. Mallat, A Wavelet Tour of Signal Processing. Academic Press, 2nd ed.,
1999.
[18] P. Dutilleux, An implementation of the algorithme a trous to compute
the wavelet transform, in Wavelets: Time-Frequency Methods and Phase
Space, Proceedings of the International Conference, Springer-Verlag,
1987.
[19] T. Y. Young and T. W. Calvert, Classification, Estimation and Pattern
Recognition. American Elsevier Publishing Co., Inc., 1974.
[20] R. Krishnan, Electric Motor Drives: Modeling, Analysis, and Control.
Prentice Hall, 2001.