Wavelet Transform Analysis

2.0 Motivation and Introduction

In this chapter the mathematical tools that are the focus of this dissertation are introduced.
Most discussions on the wavelet transform (WT) begin with a note that describes how WT is
different from Fourier transform (FT). This is because most of the researchers in signal
procession are familiar with the FT. The FT is a powerful tool that has been used extensively in
various signal processing applications. Nonetheless, it does have some limitations. The FT
provides a good representation for stationary signals that are made up of a combination of
pure sinusoids. However, it has been problematic, if the signal to be represented contains
sudden sharp transitions (spikes or bursts). These “transients” require an infinite number of
sinusoidal components to be accurately described using the FT. Thus this type of behaviour is
distributed over the entire frequency axis, and therefore is hard to isolate in the frequency
domain.
In order to overcome this type of problem a method, called short-time Fourier Transform
(STFT), is proposed where it is allowed to window the signal before taking FT. Essentially, the
STFT divides a signal into segments and takes the FT of each resulting section. The size of the
window to use relies on what type of features require to be identified. For example, to locate
short-time high frequency behaviour, narrow time windows are needed, while to resolve low
frequency behaviour, wider time windows are required. Therefore, time localization and
frequency resolution cannot simultaneously be determined to arbitrary precision. The “narrow
window” STFT has involved in detection and localization of the delta function at the coast of
losing frequency resolution of the sinusoid. So, to analyse various types of behaviour in a signal
the STFT would have to be taken several times, each time with a different size window function.
This is problematic when the location and/or duration of transient behaviour is unknown,
which is typical the case for real earth complex signal. The need of hour is to devise a mapping
framework that involves one initial window function and varies its size automatically so that all
frequency behaviour can be analysed with just one pass (narrow windows at high frequencies
and wide windows at low frequencies). This is where the wavelet transform (WT) comes in.
2.1 Wavelet Transform (WT)

1|Page
 Wavelets are a set of window functions that have compact support in time (i.e., decays to
zero quickly) and are band-limited in the frequency domain. These tiny waves that grow and
decay in short periods of time. Therefore they are called as “wavelets”. The wavelet analysis,
based on wavelet transform, is initiated by geophysicists in seismic signal analysis (Morlet et
al. 1982). A wavelet transform is an alternative way to decompose a signal to identify its
frequency distribution through time. Wavelet transform allows us to perform multi-scale
analysis, which is vital for analysing non-stationary records. It provides a good localization
properties in both depth and frequency domains (Perez- Munoz et al., 2013). Moreover, such
transformations are associated with different shapes and sizes of functions (a family of
functions) called wavelet (Fig. 2.1).

Figure 2.1: Shape and nature of the mother wavelet used in the present study (a) Haar, Gaus
1, Gaus 3, Morlet and Db2.

A wavelet function is defined as (Goupillaud et al., 1985; Perez- Munoz et al., 2013)

1  z u
 u ,s ( z)   , u  0, s  R .................. (2.1)
( s)  s 

where the function  is called the mother wavelet, s is the scale factor which determines the

wavelength and u denotes the shift of the wavelet (Goupillaud et al., 1985). In the wavelet

transform, the signal analysed is convolved with mother wavelet and the transformation is

computed for different segments of the data by varying both s and u . The continuous wavelet

transform (CWT) of signal f (z ) that assumes only the continuous values of scale and shift

parameters can be defined as


1  z u
CWT f (u, s)   f ( z) * 
| s | 
dz
s 
.................. (2.2)

where *=represents the complex conjugate. In CWT analysis, choosing scale is very important
part because it gives fine control over scale levels. The CWT analysis allows carrying out space-

2|Page
scale analysis instead of space-frequency analysis. However, proper scale to frequency
transformation allows the analysis which is very close to space-frequency analysis. So, by
reducing the scaling parameter s , the support of the wavelet reduces in space and hence

covers higher frequency ranges and vice versa. It is worth making a note of the fact that CWT
needs a large amount of computation time and storage in comparison to discrete wavelet
transform (DWT) to implement. In the DWT, variations are expressed by dyadic increase of
scale and shift parameter (based on power of two). Accordingly, the scale and shift parameters

are discretized as s  s0 and u  nu0 s0 . The resulting wavelets are defined as

m m

 z  nu0 s0 m 
 m,n ( z )  u 0  
m / 2
m

 ................... (2.3)
 s 0 
where m and n are integer values. The DWT is defined as

DWT f (m, n) 

 f ( z) m,n ( z )dz ................... (2.4)

The matrix of the wavelet coefficients, CWT f (u, s) and DWT f (m, n) is represented by

scalogram which indicates the frequency localization to different scales and shifts.

2.1.1 Wavelet Decompositions

A fast DWT algorithm developed by Mallat (1989) consists of decomposition of signal
into one or more levels of resolution by four filters including decomposition low-pass,
decomposition high-pass, reconstruction low-pass and reconstruction high-pass. The low pass

filter, associated with scaling function  j,k produces gross structure of the signal while high

pass filter, associated with wavelet function  j,k produces detail structure of the signal. The
scaling and wavelet functions are define as (Mallat, 1989)

 j ,k (n)  2 j / 2  (2 j n  k ) .................... (2.5)

 j ,k (n)  2 j / 2 (2 j n  k ) .................... (2.6)

where j represents scale level and k represents translation in that scale level. The scaling

function provides the flexibility to approximate a given signal with desired level of accuracy.
The DWT plays a significant role in dividing a complicated signal into simpler ones and analyze

3|Page
them separately. The DWT For discrete signal f (n ) defined in [0, N-1] with N data points,

wavelet coefficients are estimated by the following equations (Perez- Munoz et al., 2013)

1
W ( j , k ) 
N
 f (n)
n
j ,k ( n) ................... (2.7)

1
W ( j , k ) 
N
 f (n)
n
j ,k ( n) ................... (2.8)

where W ( j, k ) and W ( j, k ) are the approximation coefficients and detail coefficients.

The original signal f (z ) is reconstructed by using the inverse wavelet transform of the

approximation and detail coefficients as follows (Misiti et al., 2000).

f ( z)  W l  W i .................... (2.9)

i l

The original signal can be obtained from the sum of these two coefficients. By using DWT, it
is obtained detail coefficient which is used to reconstruct the high-frequency signal.
A multilevel decomposition process (Fig. 2.2) can be proposed, where the original signal
is broken down into lower resolution components (Catalão et al., 2011). There are many types
of wavelets, including Daubechies, Symmlet, Coiflet, Meyer, Gaussian, Mexican hat, Morlet and
Shanon wavelets, and so on, which have been extensively used for wavelet-based time/space
series analysis (Minuet al., 2010). For discrete wavelet analysis, orthogonal wavelets
(Daubechies’ extremal-phase and least symmetric wavelets), B-spline biorthogonal and Haar
wavelets have been commonly used because they meet the orthogonal properties.

Figure 2.2: General architecture of discrete wavelet transform.

For continuous wavelet analysis, Morlet, Meyer, Gaussian and Paul wavelets have been
used (Mathworks, 2014b). The extremal-phase wavelets including Haar and DbN wavelets,

4|Page
where N refers to the number of vanishing moments (Constantine and Percival, 2013), are also
called Daublets or Daubechies wavelets. Daubechies wavelets are one of the most commonly
used wavelets. They represent a collection of orthogonal mother wavelets with compact
support, characterized by a maximal number of vanishing moments for some given length of
the support. For example, Daubechies wavelet of order 1 (Db1), which also called the Haar
wavelet, has one vanishing moment, Db2 has two vanishing moments and so on. The Haar
wavelet has the shortest support among all orthogonal wavelets. The Haar wavelet produces
orthogonal wavelets by translations and dilations. Note that it is the single symmetric
orthogonal wavelet (Stolojescu, 2012). The least-asymmetric wavelets are also known as
Symmlets. The Symmlets are proposed to improve symmetry as a modification to Daubechies
wavelets. The Symmlets are compact supported, orthogonal, continuous, but only nearly
symmetric mother wavelets. Their construction is very similar to the construction of
Daubechies wavelets, but the symmetry of Symmlets is stronger than that of Daubechies
wavelets (Seo et al., 2015). Coifman wavelets, which also called Coiflets, are discrete wavelets
which are compactly supported wavelets and designed to be more symmetrical than
Daubechies wavelets (Stolojescu, 2012). Although, there is a large number of wavelets which
are available, the most commonly used wavelets (such as Haar, Gaus 1, Gaus 3, Morlet and
Db2) are considered in the present dissertation.

5|Page