35

Chapter 4

Wavelet Transform and Denoising





4.1 Why wavelet

Fourier transform based spectral analysis is the dominant analytical tool for frequency
domain analysis. However, Fourier transform cannot provide any information of the
spectrum changes with respect to time. Fourier transform assumes the signal is
stationary, but PD signal is always non-stationary. To overcome this deficiency, a
modified method-short time Fourier transform allows to represent the signal in both time
and frequency domain through time windowing function [29]. The window length
determines a constant time and frequency resolution. Thus, a shorter time windowing is
used in order to capture the transient behavior of a signal; we sacrifice the frequency
resolution. The nature of the real PD signals is nonperiodic and transient as shown in Fig
18, Fig 19; such signals cannot easily be analyzed by conventional transforms. So, an
alternative mathematical tool- wavelet transform must be selected to extract the relevant

36
time-amplitude information from a signal. In the meantime, we can improve the signal to
noise ratio based on prior knowledge of the signal characteristics.

In this work, we stated only some keys equations and concepts of wavelet transform,
more rigorous mathematical treatment of this subject can be found in [30-35]. A
continuous-time wavelet transform of f(t) is defined as:

= = dt
a
b t
t f a a b W b a f CWT
f
) ( ) ( ) , ( ) , (
*
2
1

(19)

Here a, b R , 0 a and they are dilating and translating coefficients, respectively. The
asterisk denotes a complex conjugate. This multiplication of
2
1

a is for energy
normalization purposes so that the transformed signal will have the same energy at every
scale. The analysis function ) (t , the so-called mother wavelet, is scaled by a, so a
wavelet analysis is often called a time-scale analysis rather than a time-frequency
analysis. The wavelet transform decomposes the signal into different scales with different
levels of resolution by dilating a single prototype function, the mother wavelet.
Furthermore, a mother wavelet has to satisfy that it has a zero net area, which suggest that
the transformation kernel of the wavelet transform is a compactly support function
(localized in time), thereby offering the potential to capture the PD spikes which
normally occur in a short period of time [36].

4.2 Discrete Wavelet transform and Multiresolution Analysis

One drawback of the CWT is that the representation of the signal is often redundant,
since a and b are continuous over R (the real number). The original signal can be
completely reconstructed by a sample version of ) , ( a b W
f
. Typically, we sample
) , ( a b W
f
in dyadic grid, i.e.,
m
a

= 2
and
m
n b

= 2 (20)

37
Z n m , , and Z is the set of positive integers. Substituting (19) into (18), we have

= dt t t f n m f DWT
n m
) ( ) ( ) , (
*
,

where ) 2 ( 2 ) (
,
n t t
m m
n m
=

is the dilated and translated version of the mother
wavelet ) (t .

The family of dilated mother wavelets of selected a and b constitute an orthonormal basis
of ) (
2
R L . In addition, we sample ) , ( a b W
f
in dyadic grid, this wavelet transform is also
called dyadic-orthonormal wavelet transform. Due to the orthonormal properties, there is
no information redundancy in the discrete wavelet transform. In addition, with this choice
of a and b, there exists the multiresolution analysis (MRA) algorithm, which decompose
a signal into scales with different time and frequency resolution. MRA is designed to give
good time resolution and poor frequency resolution at high frequencies and good
frequency resolution and poor time resolution at low frequencies.

The fundamental concept involved in MRA is to find the average features and the details
of the signal via scalar products with scaling signals and wavelets. In the PD signals we
have seen, sharp spikes are observed when PD occurs. The spikes are typically of high
frequency and we are able to discriminate the PD spikes with other noises through the
decomposition of MRA into different levels.

The differences between different mother wavelet functions (e.g. Haar, Daubechies,
Coiflets, Symlet, Biorthogonal and etc.) consist in how these scaling signals and the
wavelets are defined. The choice of wavelet determines the final waveform shape;
likewise, for Fourier transform, the decomposed waveforms are always sinusoid. To
have a unique reconstructed signal from wavelet transform, we need to select the
orthogonal wavelets to perform the transforms.


38
The wavelet decomposition results in levels of approximated and detailed coefficients.
The algorithm of wavelet signal decomposition is illustrated in Fig 22. Reconstruction
of the signal from the wavelet transform and post processing, the algorithm is shown in
Fig 23. This multi-resolution analysis enables us to analyze the signal in different
frequency bands; therefore, we could observe any transient in time domain as well as in
frequency domain.


A
0
(0 - f
s
/2)
h
g
2
D
1
(t)
(f
s
/4 - f
s
/2)
A
1
(t)
(0 - f
s
/4)
2
Level 1 Original
Signal
h
g
2
D
2
(t)
(f
s
/8 - f
s
/4)
A
2
(t)
(0 - f
s
/8)
2
Level 2

Fig 22. Multi-resolution wavelet decomposition. h = low-pass decomposition filter; g = high-pass
decomposition filter;
2
= down-sampling operation. A
1
(t), A
2
(t)are the approximated coefficient of
the original signal at levels 1, 2 etc. D
1
(t), D
2
(t) are the detailed coefficient at levels 1,2.


A
0
(t)'
h
g
D
1
(t)
Level 1
Reconstructed
Signal
h
g
2
D
2
(t)
A
2
(t)
2
Level 2
2 A
1
(t)
2

Fig 23. Multi-resolution wavelet reconstruction. h = low pass reconstruction filter; g = high-pass
reconstruction filter;
2
= up-sampling operation. A
1
(t), A
2
(t) are the processed or non-processed
approximated coefficient of the original signal at levels 1, 2 etc. D
1
(t), D
2
(t) are the processed or non-
processed detailed coefficient at levels 1,2.

39

The relation between the low-pass and high-pass filter and the scalar function ) (t and
the wavelet ) (t can be states as following:

=
=
k
k
k t k g t
k t k h t
] 2 [ ] [ ) (
] 2 [ ] [ ) (




The relation between the low-pass filter and high-pass filter is not independent to each
other, they are related by:
] [ ) 1 ( ] 1 [ n h n L g
n
=

where g[n] is the high-pass, h[n] is the low-pass filter, L is the filter length (total number
of points). Filters satisfying this condition are commonly used in signal processing, and
they are known as the Quadrature Mirror Filters (QMF). The two filtering and down-
sampling operation can be expressed by:

=
=

n
i i
n
i i
n k g t A k D
n k h t A k A
] 2 [ ) ( ] [
] 2 [ ) ( ] [
1
1

The reconstruction in this case is very easy since the halfband filters form the
orthonormal bases. The above procedure is followed in reverse order for the
reconstruction. The signals at every level are upsampled by two, passed through the
synthesis filters g[n], and h[n] (highpass and lowpass, respectively), and then added.
The interesting point here is that the analysis and synthesis filters are identical to each
other, except for a time reversal. Therefore, the reconstruction formula becomes (for each
layer)
]) 2 [ ] [ ] 2 [ ] [ ( ] [
1 1
k n h k A k n g k D k A
i
k
i i
+ + + =
+

=
+

40
4.3 Wavelet-based Denoising

The general wavelet denosing procedure is as follows [29, 37, 38]:
Apply wavelet transform to the noisy signal to produce the noisy wavelet
coefficients to the level which we can properly distinguish the PD occurrence.
Select appropriate threshold limit at each level and threshold method (hard or soft
thresholding) to best remove the noises.
Inverse wavelet transform of the thresholded wavelet coefficients to obtain a
denoised signal.

4.3.1 Wavelet selection


To best characterize the PD spikes in a noisy signal, such as Fig 18, and Fig 20, we
should select our mother wavelet carefully to better approximate and capture the
transient spikes of the original signal. Mother wavelet will not only determine how
well we estimate the original signal in terms of the shape of the PD spikes, but also, it
will affect the frequency spectrum of the denoised signal. The choice of mother wavelet
can be based on eyeball inspection of the PD spikes [29], or it can be selected based on
correlation (21) between the signal of interest and the wavelet-denoised signal [38], or
based on the cumulative energy (22) over some interval where PD spikes occur.

=
2 2
) ( ) (
) )( (
Y Y X X
Y Y X X
( 21 )

where
X
and
Y
are the mean value of set
X
and
Y
, respectively.

=
2
X E ( 22 )

where E is the energy and X is the signal vector.


41
We choose to select the mother wavelet based on the last two methods: correlation
between two signals and cumulative energy over some interval of PD spike occurrence.
We found that the two methods give us a very similar outcome.

4.3.2 Threshold limits

Many methods for setting the threshold have been proposed. The most time-consuming
way is to set the threshold limit on a case-by-case basis. The limit is selected such that
satisfactory noise removal is achieved. For a Gaussian noise [39]; if we apply orthogonal
wavelet transform to the noise signal, the transformed signal will preserve the Gaussian
nature of the noise, which the histogram of the noise will be a symmetrical bell-shaped
curve about its mean value. From theory, four times the standard deviation would cover
99.99% of the noise. Therefore, we could set the threshold be 4.5 times of the standard
deviation of the wavelet-transformed signal to remove the Gaussian noise in the signal.

We have found that for the fiber optic signals, we could simply apply the standard
deviation methods, since the signal is mostly white noises (see Fig 18 and Appendix B),
however for the PZT signals, we should set the threshold case-by-case to best denoise the
signals.

Two rules are generally used for thresholding the wavelet coefficients (soft/hard
thresholding). Hard thresholding sets zeros for all wavelet coefficients whose absolute
value is less than the specified threshold limit. It has shown that hard thresholding
provides an improved signal to noise ratio [29]. In this study, we adopt the hard
thresholding method.

4.3.3 Level of Decomposition

From the previous section, we have known that the wavelet transform is constituted by
different levels. The maximum level to apply the wavelet transform depends on how
many data points contain in a data set, since there is a down-sampling by 2 operation
from one level to the next one. In our experience, one factor that affects the number of

42
level we can reach to achieve the satisfactory noise removal results is the signal-to-noise
ratio (SNR) in the original signal. Generally, the measured signals from the PZT sensors
have higher SNR than that of the measured signals from fiber optic sensors. So to
process the PZT data, we need more level of wavelet transform (e.g. 12) to remove most
of its noise. For the fiber optic sensor data, we could only go up to 4 or 5 level otherwise
we would remove much of the PD signal, therefore the PD spikes wouldnt be captured.