CEAI, Vol. 9, No. 3;4, pp.

51-58, 2007

Printed in Romania

FAULT ISOLATION BASED ON WAVELET TRANSFORM

S. Lesecq, S. Gentil, I. Fagarasan*

Control Systems Department, Gipsa-lab (CNRS-INPG-UJF UMR 5216),

ENSIEG - BP 46, F-38402 Saint Martin dHres Cedex
*University "POLITEHNICA" of Bucharest, Faculty of Control and Computers
Splaiul Independentei 313, 060042 Bucharest, Romania
{suzanne.lesecq, sylviane.gentil}@inpg.fr}
{[email protected]}

Abstract: This paper evaluates how wavelet transform can be used to detect and isolate particular
faults. The diagnostic method that is proposed is based on the stationary wavelet transform. The
wavelet coefficients allow analysing the signal changes over different scales. Therefore, fault
detection can be performed. Each scale is related to a particular frequency band. Thus if various
faults are known to affect different frequency bands, the wavelet coefficients can be used to isolate
the faults. Fuzzyfication of the wavelet coefficients is first applied, followed by the fuzzy
aggregation of the fuzzyfied coefficients to make the isolation decision easy to compute and
gradual. Academic examples are discussed to show the efficiency of the isolation method
presented here.
Keywords: Fault detection, Fault isolation, Stationary wavelet transform Fuzzy aggregation.

1. INTRODUCTION
Fast detection of faults in dynamic systems is
very important for safe operation. Moreover,
fault isolation allows determining which
physical component is failing. Fault Detection
and Isolation (FDI) is necessary not only for
maintenance, but also for making online correct
protection decision. Signal processing is a well
known tool to deal with fault diagnosis. It is
used to analyse directly the signals measured
online, avoiding system modelling.
Its main drawback is that a change in some
signal feature, e.g. the signal mean or the

frequency contents, must be distinctive for each

fault that may occur on the system. Despite this
difficulty, signal processing is widely used in
industry, for instance when vibration monitoring
is considered.
Detection tests that aim at detecting a change in
the mean or the standard deviation of a signal
are now very common [1, 2]. Frequency
representations are particularly useful for
studying rotating machines. Indeed, extra
frequency contents may appear under the
influence of a particular fault. For instance, [3]
deeply studies faults in a three-phase induction
machine. The spectral analysis of electric and

52

electromagnetic signals shows that mechanical

abnormalities such as broken rotor bars generate
characteristic frequency contents in the signals.
Unfortunately, the Fourier Transform is unable
to accurately analyse and represent a signal with
non periodic features, for instance a transient
signal. To study non stationary signals, timefrequency methods replace traditional spectral
analysis [4, 5]. The Short Time Fourier
Transform (STFT) interpretation is close to a
local Fast Fourier Transform analysis. The
signal to analyse is multiplied by a sliding
window (for instance rectangular, Hamming,
Blackman, etc.) with finite duration. Thus the
spectrum is computed in real time and its
variation contents are used to detect faults. This
method has been applied for instance in the
metallurgical industry [6]. Actually, rise in
productivity in modern rolling mill plants
induces an increase of the rolling speed. This
also increases the potential vibrations of the
system. Different vibrations appear that
correspond to particular faults [7]. Thus
monitoring the frequency contents can help to
localise the faults.
The main drawback of STFT method is due to
the
constant
time-frequency
resolution,
according to the Heisenberg-Gabor uncertainty
principle. Indeed, there is a trade-off between
time and frequency resolutions because an
accurate time resolution requires a short
analysis window while an accurate frequency
resolution involves a long analysis window,
which introduces an extra detection delay.
In order to obtain a variable time and frequency
resolution (their product been constant), the
Wavelet Transform (WT) has been introduced.
Moreover, Wavelet analysis does not require
stationarity hypothesis and it is well adapted to
the analyses of signals with temporary changes.
It has been investigated for monitoring and
diagnosis in various industrial areas. The case of
arc tracking is typical of its use [8]: no model of
the physical phenomenon is available and arcing
appears as random discontinuities in the current
signal. Article [9] reports the use of wavelet
analysis to detect faults in a high voltage direct
current line (HVDC). Line faults, commutation
failures in the converter and single phase short
circuits at the AC side are studied and shown to
produce time varying transients. For this
application, isolating faults is very important
because the safety procedures are very different

CONTROL ENGINEERING AND APPLIED INFORMATICS

depending on the type of fault. In [10], the

surface faults of a compact disc like scratches
and fingerprints are detected and handled with
dedicated filters.
Most of the reported applications deal with fault
detection. How to process wavelet coefficients
to cope with fault isolation is investigated in this
paper. The paper is organised as follows. In
section 2, the Stationary Wavelet Transform
(SWT) is introduced and the wavelet coefficients
thresholding method is explained. Section 3 is
devoted to the description of the isolation
method proposed. It is based on the
fuzzyfication of the wavelet coefficients,
interpreted as partial criteria to be aggregated in
order to make a decision. Isolation results
achieved with the proposed method are
discussed in section 4.
2. STATIONARY WAVELET
TRANSFORM AND THRESHOLDING
A detection method based upon multiresolution
analysis (MRA) has been previously proposed
[8]. It has been applied to vibration monitoring
in [11]. In this paper, the Stationary Wavelet
Transform (SWT) is used instead of the MRA for
the detection of low frequency vibrations. The
main advantage of the SWT [7, 12] is its timeinvariance property: the SWT coefficients of a
delayed signal are just a time-shifted version of
the original ones. A quick overview of the SWT
is now given and the thresholding technique of
the wavelet coefficients is summarised.
2.1. Wavelets Transform and Stationary
Wavelets Transform
The Continuous Wavelet Transform (CWT)
projects a signal x(t) on a family of zero-mean
functions a,b (the wavelets) deduced from an
elementary function (the mother wavelet) by
means of translations and dilatations:
CWTa ,b ( x ) =

x (t ) * ( t ) dt
a ,b

where * stands for conjugate, a is the scaling

parameter (taking |a| > 1 dilates the function ),
b
is
the
translation
parameter
and
a ,b ( t ) = 1/ a ( ( t b ) / a )
[13]. The

redundancy introduced by the CWT can be

reduced by the discretization of parameters a
and b, leading to the Discrete Wavelet
Transform (DWT):

CONTROL ENGINEERING AND APPLIED INFORMATICS

DWT j , k ( x) =

x(t ) j , k ( t ) dt

(1)

a ( a t kb0 ) .

where j , k ( t ) = 1

j
0

j
0

The choice a0 = 2, b0 = 1 corresponds to the

dyadic sampling of the time-frequency plane
(i.e. one set of coefficients per octave). Thanks
to this particular sampling, it is possible to
obtain for the set j,k an orthonormal basis with
a wavelet well localized both in time and
frequency (see Fig. 1 for the shape of a few
wavelets).
At level j = 1 : J, the approximations akj and
details d kj (Fig. 2) are recursively computed
using digital filters, respectively a lowpass filter
h and a highpass one g . Note that the
computation of the coefficients lead to the
evaluation of dot products which is very
attractive for the implementation on Digital
Signal
Processors
(fused
add-multiply
operation).
1

0.8
0.6

0.5

0.4
0.2
0

-0.2
-0.4

-0.5

-0.6
-0.8

-1
0

0.2
D

0.4
iti

0.6
hi h

0.8
filt

10
15
iti
hi h

filt

20

Fig. 1: Wavelets (Haar, Daubechies12 db12)

High freq.

Low freq.
Level 3

a3 d3

Level 2

a2

d1

Freq.

This recursive algorithm is initialized by x(k ) ,

2 k l l

~
h

g~

d k1

~
h
g~

ak2

d k2

Fig. 3. SWT decomposition algorithm

2.2. Wavelet coefficient thresholding

The singularity occurrence in a signal is

revealed by the size of the wavelet coefficients
[11, 16]. Thus, the objective is to find when the
noisy data:

x ( ti ) = g ( ti ) + ( ti ) , i = 1: n

(3)

changes its behaviour, without any hypothesis

about the parametric form of g which contains
this change. In (3), is usually supposed to be a
zero mean and 2 variance independent
normally distributed noise, which allows
interesting theoretical results about the
optimality of thresholding [14]. The aim is to
distinguish the wavelet coefficients containing
information about the singularity from the other
ones which belong to the normal signal
behaviour. In [15] Donoho proposes to extract
these
significant
coefficients
by
soft
thresholding:
dkj + j
dkj < j

kj = 0
0 dkj j
(4)
j
j
d j j
d

>
k
k
given in (2) and j is the threshold value.

Fig. 2: Frequency domain split

i.e. ak0 = x(k ) :

aj =
h a j 1 ;

ak0

1
k

where kj is the thresholded coefficient, d kj is

d2
a1

Level 1

53

d kj = l g 2 k l alj 1

(2)

As stated above, the main advantage of the SWT

(Fig. 3) is its time invariance. This property is
fundamental for diagnosis purpose to provide a
symptom that is time-invariant, i.e. its value
does not depend on the time the fault occurs.

The coefficients below their threshold are set to

zero (they are assumed to represent the normal
behaviour), while exceeding coefficients
indicate the occurrence of a signal abnormal
behaviour (Fig. 4). Note that the occurrence of a
singularity may affect only a few levels.
Therefore, j is level-dependant.
The threshold choice is tricky. Several methods
may be used, and a bibliographical study reveals
many possibilities. The optimal choice requires
knowledge (or at least hypotheses) about the
analysed signal [11, 16]. In this paper, the
minimax threshold has been chosen. Its main
property is that the Risk function:

54

CONTROL ENGINEERING AND APPLIED INFORMATICS

R ( g , g ) =

1
E
n

( g ( t ) g ( t ) )
n

i =1

(5)

Wavelet db12 (blue) and Center frequency based approximation

Wavelet db12
Center Freq. Approx

0.8
0.6

is minimum, being the reconstructed signal.

For each level j, the threshold j is given by:

0.4
0.2
0

=
j

j =

if n j 64

*
nj

2 log n j

( )

-0.2

(6)

if n j < 64

-0.4
-0.6
-0.8

where j is the standard deviation of d kj , nj is

the number of coefficients d kj that are used for
the threshold computation and n*j has been
tabulated and can be found in [14].

-1

10
15
Period: 1.4375; Cent. Freq: 0.69565

20

25

Fig. 5. Centre frequency approximation of db12

3. FAULT DETECTION AND ISOLATION

j
k

3.1. Principle
d kj

j
0

Fig. 4. SWT coefficient soft thresholding

2.3. Choice of the Wavelet

When time-scale methods are used, the

relationship between scale and frequency is
expressed through the pseudo-frequency fa (in
Hz) corresponding to a given scale a. fa is
computed thanks to the normalised centre
frequency fc of the wavelet [17]:

fa =

fc
a Te

where Te = 1 f e

Wavelet decomposition can be implemented for

diagnostic purpose when a fault occurrence is
revealed by a signal singularity. The proposed
detection method analyses the changes that
appear over the different decomposition levels
to detect the singularity. The hypothesis for fault
isolation is that different faults induce different
effects on the wavelet coefficients over the
decomposition levels. The isolation method
proposed in this paper analyses the modification
of the wavelet coefficients over the different
levels of decomposition to deduce which fault is
present.
3.2. Detection

(7)

The underlying idea is to associate with a given

wavelet a purely periodic signal of frequency fc
that maximizes the Fast Fourier Transform
(FFT) of the wavelet modulus. In Fig. 5, the
db12 wavelet [18] and its associated purely
periodic signal are drawn. The pseudofrequency fa depends on the wavelet and the
decomposition level j. Thus, it can suggest the
choice of the analysing function, and the number
of decomposition levels J, depending on the
frequency contents that reveal the appearance of
a particular fault.
Another criterion for the choice of the wavelet is
related to the kind of singularity that must be
detected, i.e. the singularity that appears in the
signal when a fault occurs. This choice is
directly connected to the regularity of the
wavelet [18]. The reader can refer to [13] for a
few examples on this subject.

The detection procedure works in three steps.

The first step transforms the signal into wavelets
coefficients. It decomposes the signal on J
scales. This step also allows characterizing the
frequency contents that define the normal
behaviour of the system.
The second step corresponds to the wavelet
coefficient thresholding, where the thresholds
are computed as explained in section 2.
Nevertheless, the classical thresholding method
is a crisp one, while a gradual indicator is
generally more interesting than a Boolean one.
Indeed, it allows focusing attention on a
component before a fault is completely installed.
Therefore, a fuzzyfication of the thresholded
coefficients is implemented:

CONTROL ENGINEERING AND APPLIED INFORMATICS

1
kj 2 j j

kj
j
j
( k ) =
0 < kj < 2 j j (8)
2 j j

kj = 0
0
where j is a parameter that defines the
membership function of the coefficients (Fig. 6).

j
k

1
0.5

2 j j

j j

j j

2 j j

kj

appearance, and the singularity may not be

present over all scales Thus, (10) is proposed for
the detection decision:
D 1k = max ( kj ); j = 1 : J
j

(10)

From a practical point of view, it can be

observed that the wavelet coefficients may be
very small, during a very short time, even when
there is a singularity in the signal. Thus (11)
may be preferred to (10), to favour a clearer
decision:
D2k = max{max(kjl ); l = 0: N 1}; j = 1: J

Fig. 6.Coefficient fuzzyfication

(11)

The third step corresponds to the detection

decision. In order to give a unique indicator, the
various fuzzy coefficients are considered as
partial criteria and the detection problem is
regarded as a fuzzy decision making one with
partial criteria. Fuzzy decision making allows
formal modelling of decision-making for
imprecise and uncertain conditions. The
decision (here the detection decision) is
considered as a fuzzy set described by its
membership function d that is computed using
the membership functions of the various partial
points of view on the final decision ci (d ) :
d = h ( c1 ( d ) , c2 ( d ) ,..., c p ( d ) )

55

(9)

where h is a fuzzy set operator to be determined

in function of the properties that are required for
the decision.
Consider the most common operators,
conjunctive or disjunctive ones and a small
example. Suppose that the faulty state of a
heat exchanger has to be diagnosed using two
criteria, a high temperature and a small
flow. A conjunctive operator h states that all the
criteria must be met simultaneously (the state is
faulty if the temperature is high and the flow
is small at the same time). It can be expressed
mathematically by a min function, for instance.
A disjunctive operator h states that a single
criterion is sufficient for the decision to be made
(the state is faulty if either the temperature is
high or the flow is small) which is
expressed for instance by a max function. A
mean function expresses a compromise operator.
When a singularity occurs in the signal, at least
one level of decomposition must reveal its

where N is a small time window. [19] proposes a

comparison of different aggregation operators to
detect extra vibrations (considered as faults) in a
rolling mill.
3.3. Isolation

For fault isolation, the singularity appearance

must modify differently the various levels of
decomposition, depending on the considered
fault. A learning phase shows which levels are
modified by a specific fault. For example,
consider a signal that is decomposed over 5
levels. Moreover, suppose that the wavelet
coefficients on levels i and j are modified by the
fault, while the coefficients on levels k, l, m are
not modified. This situation can occur for
instance when the fault gives rise to oscillations
in a specific frequency range as reported by [3]
for electrical drives or [7] for rolling mills. The
isolation decision for this specific fault can be
given by:
ki , kj , (1 kk ),

D 3 k = min

l
m

(1 k ), (1 k )

(12)

(12) expresses that the coefficients on levels i

and j must be high at the same time, and the
other coefficients must be small, to decide that
this fault is present.
4. EXPERIMENTAL RESULTS

In this section, different scenarios are discussed

in order to evaluate the SWT capabilities to
detect and isolate particular faults. For a wide
range of applications, particular additive
frequency contents are related to the occurrence

56

CONTROL ENGINEERING AND APPLIED INFORMATICS

of a particular fault (e.g. faults in rolling mill

process or abnormalities such a broken rotor
bars in induction motor). In other applications,
the signals recorded on the process exhibits
impulses in amplitude or a pseudo frequency
occurrence when a fault occurs. All these
situations can be handled with SWT. Table 1
gives three academic examples that mimic these
situations. These simulated signals will show the
powerful of the isolation method proposed in
this paper.
Table 1. Scenario number and its simulated signal

0 xref (t ) = sin ( 2 ft ) + (t )

x1 (t ) = xref (t ) + 0.7(sin ( 2 f1t )

+ sin ( 2 f 2 t ))(u (t 1 ) u (t 2 ))
1
+ 0.7sin ( 2 f 3t ) (u (t 3 ) u (t 4 ))
n

k +1
2 x2 (t ) = xref (t ) + 40 (1) (t k 1 )
k =1

x3 (t ) = xref (t ) +
+2e 0.25(2 f1t ) sin(2 f1t ) ( u (t 1 ) u (t 2 ) )

The parameters are f = 50Hz, f1 = 20Hz, f2 =

350Hz, f3 = 175Hz. Scenario 0 corresponds to
the reference signal (i.e. normal behaviour): it
corresponds to a noisy sinusoidal signal. is a
Gaussian white noise with zero mean and
variance chosen such that the Signal-to-Noise
Ratio is SNR 10dB .
In scenario 1, extra frequency contents f1 and f2
occur at time t = 1 during a time interval
2 1 and another additive frequency f3 occurs
at instant t = 3 during a time interval 4 3
( 1 < 2 < 3 < 4 ). Scenario 2 corresponds to a
fault characterised by the appearance of periodic
impulses while scenario 3 deals with the
appearance at time t = 1 of a pseudo frequency
of duration 2 1 > 0 .
In order to detect and isolate the faults described
in scenarios 1 to 3, some parameters of the SWT
must be discussed. The sampling frequency fe of
the signal and the number of decomposition
level of the wavelet transform are related to the
frequency that must be detected through
equation (7). The sensitivity of the FDI method
proposed here is increased when fi, i = 1:3,
satisfies (7). This remark should guide the
choice of the mother wavelet. Actually, the SWT
can be performed with different wavelets based
on Matlab function swt. For instance, the Mallat

wavelet is used in [9] for detection and

identification of faults in HVDC systems. The
Morlet wavelet has been used in the literature
for the analysis of vibration signals recorded on
rotating machineries [20]. This is due to the fact
that the Morlet wavelet is able to pick up
impulses generated by the rotating elements.
Other wavelets are used in the literature but the
Daubechies wavelets [18] are used in a wide
range of applications [21, 22]. This is certainly
due to their nice properties (compact support,
number of vanishing moments, orthogonality,
etc.).
For the examples in Table 1, a wavelet
decomposition over 5 levels (J = 5) is sufficient
to ensure a good detection. The sampling
frequency is equal to 1 kHz. The Daubechies 12
db12 wavelet has been used because it is able
to highlight the faulty extra frequency
contents. The thresholds j have been computed
with the reference signal xref thanks to equation
(4). x1, its SWT decomposition and the
thresholds j are given in Fig. 7.
Wavelet decomposition of the analysed signal (SNR = 9.1957 dB)
2

Signal0
-2

a5 [0,15.62] Hz

1
0
-1

d5 [15.62, 31.25] Hz

5
0
-5

d4 [31.25, 62.5] Hz

20
0
-20

d3 [62.5, 125] Hz

5
0
-5

d2 [125, 250] Hz

5
0
-5
2
0
-2
0

d1 [250, 500] Hz

0.5

1.5

2
Time[s]

2.5

3.5

Fig. 7. x1 and its SWT decomposition

Membership functions
2
Signal0
-2

10.5

0.5
0

0.5

0.5
0
1

50.5

0
0

0.5

1.5

2
Time [s]

2.5

3.5

Fig. 8. Fuzzyfication of the thresholded coefficients

The SWT coefficients d5 and d1 clearly exhibit

the extra frequency contents f1 = 20Hz and f2 =

CONTROL ENGINEERING AND APPLIED INFORMATICS

350Hz. This can be explained by the dyadic split

of the frequency domain (see Fig. 2). The other
extra frequency content is characterized by
f3 = 175Hz . It is exhibited in the coefficients d2
on the second level of decomposition. The
thresholded coefficients are fuzzyfied with the
membership functions j , j = 1: 5 calculated
with (8). The result is shown in Fig. 8. The
abnormality in each frequency band is clearly
exhibited.
The fault detection indicator FD is computed
with (11). It measures the appearance of an
abnormal behaviour over all the levels of
decomposition. When fault isolation is
considered, specific aggregation operators must
be defined. These new operators take into
account some knowledge on the kind of
singularity that appears when a particular fault
occurs. Thus, the fault isolation decisions that
are defined are given by:

57

decision rule that takes into account all the

decomposition levels is:
FI F 3 = min {1 , 2 , 3 , 4 , 5 }

(15)

This can be explained by the fact that impulses

induce wavelet coefficients that exceed their
thresholds over several levels of decomposition.
As expected, the fault (occurrence of impulses)
is detected and localised. The results obtained
for the 3rd scenario (signal x3 in Table 1) are
given in Fig. 11. Note that the pseudo frequency
is equal to f1. Thus, the wavelet coefficients on
level 5 are sensitive to the occurrence of this
extra frequency content. To isolate this fault, the
decision rule must mainly focus on 5 :
(1 1 ) , (1 2 ) ,

FI F 4 = min
(1 3 ) , (1 4 ) , 5 (16)

As expected, the fault is detected and isolated.

Analysed signal and FDI

, (1 2 ) , (1 3 ) ,
FI F 1 = min 1
(13)
(1 4 ) , 5

(1 1 ) , 2 , (1 3 ) ,

FI F 2 = min
(1 4 ) , (1 5 ) (14)

20
Signal 0
-20
-40

1 -> fault detection; 0 -> unfaulty

1
FD 0.5

Results are shown in Fig. 9.

0.5

1.5

2.5

3.5

Analysed signal and FDI

1 -> identification fault F3

2
FIF3

Signal 0

0.5

-2

0
0

0.5

1.5

2
Time [s]

2.5

3.5

1
FD

Fig. 10. Signal x2 and FD, FIF3

1 -> fault detection; 0 -> unfaulty

0.5
0
0

0.5

1.5

2.5

3.5

Analysed signal and FDI

2

1
FIF1

0.5

Signal

1 -> identification fault F1

0
0

-1

0.5

1.5

2.5

3.5

-2

1
FIF2

1 -> identification fault F2

0.5

FD

0.5

1.5

2
Time [s]

2.5

1 -> fault detection; 0 -> unfaulty

0
3

3.5

0.5

4
0

Fig. 9. Signal x1 and FD, FIF1, FIF2

It can be observed that the isolation decision

FIF1 that is devoted to the detection of
frequencies f1 and f2 clearly identifies this fault.
Identically FIF2 is able to detect the fault
characterized by f3.
The results achieved for the second scenario
(signal x2 in Table 1) are given in Fig. 10. The
fault detection and isolation procedure is similar
to the one presented for scenario 1 but the

0.5

1.5

2.5

1 -> identification fault F4

1
FIF4

3.5

0.5

0
0

0.5

1.5

2
Time [s]

2.5

3.5

Fig. 11. Signal x3 and FD, FIF4

5. CONCLUSIONS

In this paper, the capability for the stationary

wavelet transform to deal with different faults

58

CONTROL ENGINEERING AND APPLIED INFORMATICS

for fault detection and isolation has been

investigated. A detection procedure based upon
the thresholds of the wavelet coefficients has
been considered. These coefficients are
fuzzyfied and aggregated in order to provide a
symptom. The tuning parameters of this
procedure are the wavelet itself, the number of
decomposition levels, the thresholds and the
decision method. The wavelet choice depends
on the features that must be detected in the
signal under analysis. This selection is
sometimes not unique. For detection purpose,
the final choice is made in order to maximize the
symptom sensitivity.
REFERENCES
[1] Basseville
[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

M., Nikiforov L. (1993),

Detection of abrupt changes, Prentice Hall.
Isermann, R. (2006), Fault Diagnosis
Systems: an introduction from fault
detection to fault tolerance, Springer-Verlag
Berlin-Heidelberg.
Eltabach M., Charara A., Zein I. (2004), A
comparison of external and internal methods
of spectral analysis for broken rotor bars
detection in induction motor, IEEE
Transactions on Industrial Electronics, Vol.
51(1), 107-121.
Flandrin P. (1999), Time-frequency/timescale analysis, Academic Press Ed., San
Diego, California.
Boashash B. (edited by) (2003), Time
frequency signal analysis and processing, a
comprehensive reference, Elsevier.
Fagarasan, I., S. Lesecq, S. Taleb, S. Gentil,
R. Stuecher, (2004), Signal-based diagnosis
of torsional vibration of a main drive of a
hot rolling mill, in Proc. IFAC Symposium
on automation in Mining, Mineral and Metal
processing MMM04, Nancy, France.
Seelinger A., Mackel J., Georges D. (2002),
Measurement and diagnosis of preprocessdisturbing oscillations in high-speed rolling
plants, in Proc. IMEKO 2002, Tampere,
Finland.
Lesecq S., Barraud A. (2003), Arcing fault
detection using wavelet transform, in Proc.
IFAC Safeprocess 2003, Washington D.C.,
USA
Shang L., Herold G., Jaeger J., Krebs R.,
Kumar
A.
(2001),
Analysis
and
identification of HVDC system faults using
wavelet modulus maxima, in Proc. IEE

Conference AC-DC Power Transmission,

315-320.
[10] Odgaard P. F., Stoutstrup J., Wickerhauser
M. V. (2006), Wavelet packet based
detection of surface faults on compact discs,
in Proc. IFAC Symposium Safeprocess,
Beijing, China.
[11] Taleb S., Lesecq S., Stuecher R. (2004),
Torsional vibration detection using torque
measurement of a main drive of a hot
rolling mill with wavelet transform, in
Proc. IFAC MMM04, Nancy, France
[12] Coifman R.R, Donoho D. (1995),
Translation Invariant De-noising, series
Lecture Notes in Statistics, New York
Springer-Verlag, Vol. 103, 125-150
[13] Strang G., Nguyen T. (1996), Wavelets and
Filter Banks, Wellesley-Cambridge Press.
[14] Donoho D., Johnstone I. (1994), Ideal
Spatial Adaptation by Wavelet Shrinkage,
Biometrika, Vol. 81.
[15] Donoho D. (1995), De-Noising by softThresholding, IEEE Trans. on Information
Theory, Vol. 41(3), 613-627.
[16] Mallat S., Zhong S. (1992), Characterization
of Signals from Multiscale Edges, IEEE
Trans. on Pattern Anal. and Machine
Intelligence, Vol. 14, 710-732.
[17] Abry P. (1997), Ondelettes et Turbulence.
Multirsolutions,
Algorithmes
de
Dcomposition, Invariance d'chelles,
Diderot Ed., Paris.
[18] Daubechies I. (1998) Orthonormal Bases of
Compactly
Supported
Wavelets,
Communications on Pure and Applied
Mathematics, No. XLI, pp. 909996.
[19] Lesecq S., Gentil S., Taleb S. (2006), Fault
detection based on wavelets transform.
Application to a roughing mill, in Proc.
IFAC Symposium Safeprocess 2006,
Beijing, China.
[20] Lin, J (2000). Feature extraction based on
Morlet wavelet and its application for
mechanical fault diagnosis, Journal of
Sound and Vibration, 234 (1), 135-148.
[21] Li, C., Song Z., Li P. (2004), Bearing fault
detection via wavelet packet transform and
rough set theory, in Proc. 5th World
Congress on Intelligent Control and
Automation, Hangzhou, China, 1663-1666.
[22] Li, W.J, and Li Y.C. (2005), Arc fault
detection based on wavelet packet, in Proc.
4th International Conference on Machine
Learning and Cybernetics, Guangzhou,
China, 1783-1788.