Wavelet Transform for

time-frequency analysis
in transient turbulent flow
Alessandro Bellofiore
 Fourier Analysis Figure 1.7

Let's look at another example. Figure 1.8 plots a signal with four different frequency components
at four different time intervals, hence a non-stationary signal. The interval 0 to 300 ms has a 100
Hz sinusoid, the interval 300 to 600 ms has a 50 Hz sinusoid, the interval 600 to 800 ms has a 25
Hz sinusoid, and finally the interval 800 to 1000 ms has a 10 Hz sinusoid.

Figure 1.5

And the following is its FT:

Figure 1.5

And the following is its FT:

Figure 1.5 Figure 1.8

And the following is its FT:

And the following is its FT:

Figure 1.6
Figure 1.9
The top plot in Figure 1.6 is the (half of the symmetric) frequency spectrum of the signal in Figure

•
Do not worry about the little ripples at this time; they are due to sudden changes from one

The FT yields reliable results when the analyzed signal is stationary

1.5. The bottom plot is the zoomed version of the top plot, showing only the range offrequency
frequenciescomponent to another, which have no significance in this text. Note that the amplitudes
that are of interest to us. Note the four spectral components corresponding to the frequencies 10,frequency components are higher than those of the lower frequency ones. This is due to
of higher
fact that higher frequencies last longer (300 ms each) than the lower frequency components (200
25, 50 and 100 Hz. ms each). (The exact value of the amplitudes are not important).

• When the frequency content of the signal is time-dependent, then a

Other than those ripples, everything seems to be right. The FT has four peaks, corresponding to
Contrary to the signal in Figure 1.5, the following signal is not stationary. Figure 1.7 plots a signal
four frequencies with reasonable amplitudes... Right
whose frequency constantly changes in time. This signal is known as the "chirp" signal. This is a
WRONG (!)

time-frequency representation is preferable.

non-stationary signal.
Well, not exactly wrong, but not exactly right either...
Here is why:
Figure 1.6
For the first signal, plotted in Figure 1.5, consider the following question:

he top plot in Figure 1.6 is the (half of the symmetric) frequency spectrum of theAtsignal in Figure
what times (or time intervals), do these frequency components occur?
Time-frequency analysis: the Wavelet Transform

• The Wavelet Transform is viable but pointless for stationary signals.

• The Wavelet Transform is not the unique solution for time-frequency

analysis, but it is usually preferred to the others (e.g. Short Time
Fourier Transform).

• The reason is connected to the fact that, according to the uncertainty

principle, in signal analysis the higher is the frequency resolution, the lower
the time resolution.

• STFT uses a fixed resolution, whereas WT gives a variable resolution:

• it yields a higher time resolution (and poorer frequency resolution) in the range
of high frequency, while in the low frequency range we get higher frequency
(and poorer time) resolution.
 Chapter 3 Signal Processing with Continuous Wavelets

The resolution problem

Figure 3-7 shows the tiling of the STFT-based time-frequency
representation.

The STFT spectrogram has uniform

Chapter 3 Signal Processing with Continuous Waveletstime-frequency resolution across the
whole time-frequency domain.
Frequency

Frequency
One can balance the time-frequency
Figure 3-9 shows the tiling of the AWT-based time-frequency
resolution by adjusting the window
representation that provides fine frequency resolution at low frequen
length.
and fine time resolution at high frequencies.
Time Time

Figure 3-7. Tiling of STFT-Based Time-Frequency Representation

In Figure 3-7, you can see that the STFT spectrogram has uniform
Thetime-frequency
WT-based resolution across the whole time-frequency domain. You
time-frequency representation
can balance the time-frequency resolution by adjusting the window length.
provides fine
Thefrequency resolution
left tiling diagram at low
provides better frequencies
frequency resolution in the STFT

Frequency
andSpectrogram
fine time(Window Length at
resolution = 256)
high frequencies.
graph of Figure 3-6. The right
tiling diagram shows better time resolution in the STFT Spectrogram
(Window Length = 64) graph of Figure 3-6. However, you cannot achieve
However, highone
timecannot
resolution achieve high
and frequency time simultaneously.
resolution resolution
and3-8
Figure frequency resolution
shows the AWT-based simultaneously.
time-frequency representation of the
HypChirps signal. In the Scalogram graph, you can distinguish the two
frequency components at both low and high frequencies. Time

Figure 3-9. Tiling of Wavelet-Based Time-Frequency Representation

g Wavelet Signal
2 Wavelet and Wavelet Transform
The Continuous Wavelet Transform Just as the Fourier transform decomposes a signal into a family of co
sinusoids, the wavelet transform decomposes a signal into a family
wavelets. Unlike sinusoids, which are symmetric, smooth, and regu

•
wavelets can be either symmetric or asymmetric, sharp or smooth, r
As well as FT, the Wavelet Transform performs
or irregular. Figure 2-1an inner
shows a sineproduct. Inwavelet,
wave, the db02 this and th
case the inner product involves awavelet.
family of wavelet, which are signals
localized in both time and frequency domains.
apter introduces wavelets and the wavelet transform and describes

•
efits of wavelet signal processing in detail.
The family of wavelet is generated by continuous stretching and shifting
of a given mother wavelet.
let Transform
• Dilatation
he Fourier transform decomposesis aconnected to offrequency
signal into a family complex
detection.
ds, the wavelet transform decomposes a signal into a family of

•
s. Unlike sinusoids, which are symmetric, smooth, and regular,
s can be eitherTranslation is usedsharp
symmetric or asymmetric, to orevaluate time
smooth, regular
localization.
ular. Figure 2-1 shows a sine wave, the db02 wavelet, and the FBI
.

Figure 2-1. Sine Wave versus Wavelets

Fourier Transform Mother wavelets for WT
The Continuous Wavelet Transform

the blue area represents

the region of influence
of the wavelet
3.3
Figure 3.5
3.4

• The signal is scanned both in time and frequency by inner product

In Figure 3.3, the signal and the wavelet function are shown for four different values of tau . The

to
with the family of wavelets.
signal is3.4
Figures
thethe
a truncated
lowestwidth
window
version of
and 3.5 illustrate
scale,changes
or highest
thethe
with
signal
same
frequency.
increasing
shownforinthe
process
Note how
scale
Figure 3.1.
scales s=5Theandscale
compact itfrequency).
(decreasing
value
s=20,
is (the blueAs
is 1 , corresponding
respectively.
window).
the window
Note how
It should
widthbe

• as narrow as
increases, thethe highest starts
transform frequency component
picking that exists
up the lower in thecomponents.
frequency
Smaller scales give good time resolution, but intercept a portion of
signal. Four distinct locations of
the wavelet function are shown in the figure at to=2 , to=40, to=90, and to=140 . At every
As a result,
it isformultiplied
every scale
by and for every time (interval), one point of the only
time-scale
whereplane is
the signal too small to achieve good frequency resolution.
location,
computed.
falls in the The
region computations
the signal.
of support of at the
Obviously,
onewavelet,
scale construct
the product
and it isthe rows
zero
is nonzero
of the time-scale
elsewhere. plane,
By shifting
the signal
and the in
the wavelet
computations
time, the signalatisdifferent scales
localized construct
in time, and bythe columnsthe
changing ofvalue
the time-scale
of s , the plane.
signal is localized in scale
Potential for biomedical investigation

Downloaded By: [National University of Ireland Galway] At: 12:06 20 June 2008
•
54 S. M. Debbal and F. Bereksi-Reguig
Usual biomedical signalsauscultation.
present a “leading” low-frequency
Therefore, clinic capabilities to diagnose heart
component,
appear. Hence, heart sound analysis by ausc
which defines the basic sounds
structure of the signal.
are limited. The components of the sound S1 are
M1 and T1, one due to closure of the mitral valve and the
primary test conducted by physicians t
condition of the heart. Yet, heart sound

•
other due to closure of the tricuspid valve. For the second auscultation as well as analysis of the PC
Precise time localization of this part of the frequency spectrum is not sound S2 the components are due to the closure of the
aortic valve (A2) and to the closure of the pulmonary
not gained widespread acceptance. This is
many controversies regarding the genesis of t
as relevant as it is in the case of high-frequency (small scale) valve (P2).
Many diseases of the heart cause changes in heart sounds
the lack of quantitative techniques for relia
the signal features. The heart sound signal h
components, which are often connected to punctual events. and additional murmurs before other signs and symptoms information than can be assessed by the hu

• In the latter case precise time

localization and resolution can
be of higher interest than
frequency resolution.
• The example signal is a
phonocardiogram (from
Debbal et al., 2008)

Figure 1. Two cycles of the phonocardiogram signal.

 signal with a set of continuous wavelets∞according to the following
equation:
Formulation and output
!"
# of the CWT
= ! %, ψ " = % ( & )ψ∗
#, $ ( & ) '&
#, $
–∞
#, $

where !" #, $ = ! %, ψ #, $" =

#–∞ % ( & )ψ∗ #, $ ( & ) '&

! & &-----------
– #-' mother
where ψ #, $ = ------- ψ
$ $ $ % wavelet
where
and u = shift factor
WTu, a is the resulting wavelet coefficients. ψu, a denotes a continuous
a = scale
! & & – factor
#'
wavelet,
Time where u is the shift
Time ψfactor
#, $ = andψa$ is$the% scale factor of the wavelet.
------- ------------ Time
ψ∗u, a is the complex conjugate of ψu, a$. For the continuous-time signal s(t),
(mother wavelet)
the scale factor must be a positive real number, whereas the shift factor can
WTbe is the
u, aany realresulting
number. Ifwavelet coefficients.
the continuous wavelet ψu,ψa denotesmeets a continuous
the admissibility
u, a
wavelet,
condition where
1, youu is
canthe shift
use the factor
computedandwavelet
a is thecoefficients
scale factor to of the wavelet.
reconstruct
•ψ∗The
u,the
result of the CWT analysis can be
is the
a originalcomplex conjugate
signal s(t). of ψ . For the continuous-time higher
signal frequency
s(t),
plotted as a scalogram. u, a
the scale factor must be a positive real number, whereas the shift factor can
•beThe exampleyou
However, is the closinguse
seldom sound
the signal integration to compute the CWT
above
ofany real number. al., 2006) wavelet ψu, a meets the admissibility
If theetcontinuous
scale

a SJM valve (from Sugiki

because1, of
condition you thecanfollowing
use thereasons:
computed wavelet coefficients to reconstruct
• • The majority of real-world signals that you encounter are available as
The analysis is capable to detect the high-
the original
frequency signal
sound splits(t).
due to the delayed lower frequency
discrete-time
closure of one of thesamples.
leaflets. The analytical form of the signal s(t) usually is
time
However,
notyou seldomgrayuse
accessible. scalethe
chartabove
cale and grayscale charts of complex integration
more clearly to compute
shows the salient features thanthe
the CWT
 Velocity data in scaled-up MHV
X/D
0 0.05 0.1 0.15 0.2 0.25 0.3

• Flow field in MHV

0
vorticity
(1/s)

•
0.05 20
The flow is characterized by: 16
12
8

•
0.1
turbulence (Re ≈ 6000)
4
0
-4

•
-8
0.15
pulsatile behavior -12
-16
-20

• leaflet dynamics
0.2

0.25

Y/D
• Flow investigation with PIV 0.3
2 m/s

• PIV system provides velocity maps with 64x64 cells spatial resolution.

• The time resolution is given by the sampling rate: 15 Hz.

• PROBLEM: how to measure scales and dynamics of the turbulent

structures developing downstream of the leaflets?
 Velocity data in scaled-up MHV
• If the turbulent flow were “organized and repeatable”, then the
Fourier analysis could be suitable for achieving the pattern of
dominant scales in vortex shedding.

• The results of the FT might be held reliable only if the vortex shedding is
verified to be stationary.

• In a transient pulsatile flow time-frequency analysis is to be pursued.

• The WT may be performed either in time or space domain, providing

(in principle) different results.

• WT over space might yield a minimum detectable length scale of about

250 µm in real valve scale. Nonetheless each velocity map has only 64
points per line.

• WT over time is expected to detect length scale not lower than 1 mm in

real valve scale. The sampling over time can cover a full cardiac cycle, with
a number of samples between 300 and 500 per cycle (depending on heart
frequency and according to the PIV acquisition rate)
 PIV – Wavelet Interactive Program
• A code has been developed in Labview 7.1 environment, in order to
evaluate the Continuous Wavelet Transform of data collected using
the PIV system and the scaled-up MHV model.
• The code is meant as a powerful tool to intestigate the dynamical
behavior of the flow field as measured by PIV system.
• The program requires the selection of a set of velocity (or vorticity)
data, provided as .vec files generated by Tecplot.
• The interface displays:
• the velocity field, in a 2D colour map, at a certain time location, which can
be set as program runs (live update!).
• the velocity signal over time at a certain spatial location, which can be set
as program runs (live update!).
• the scalogram of the CWT.
• Either U- or V- velocity component can be selected as program runs.
• The selected mother wavelet can be set as program runs.
PIV – Wavelet Interactive Program

Screenshot of the Labview program