You are on page 1of 97

Empirical Wavelet Transform

By ARJUN SS
GECT
Roll no 5
March 23, 2014
Gec Thrissur Empirical Wavelet Transform
Table of Contents
Introduction
Empirical Mode Decomposition
Empirical Wavelet Transform Concept
Empirical Wavelets
Empirical Wavelet Transform
Comparing EMD and EWT
Conclusion
Gec Thrissur Empirical Wavelet Transform
Introduction
Fourier transform, STFT and Wavelet transforms
Well known ability of the wavelet transforms to pack the main
signal information into a very small number of wavelet
coecients.
Many wavelets are present to be chosen for a specic signal.
Basis are designed independent of signal.
Gec Thrissur Empirical Wavelet Transform
Cont..
Signals which are non stationary and non-linear require
adaptive ways.
Adaptive methods to construct basis directly from information
in the signal.
Empirical Mode Decomposition(EMD).
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition
Figure:
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition
Figure:
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition
Figure:
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition
Figure:
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition
Figure:
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition
Figure:
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition
Figure:
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition
Figure:
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition
Figure:
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Principle)
Decompose a signal as a (nite) sum of N + 1 Intrinsic Mode
Functions (IMF) such that
f (t) =
N

k=0
f
k
(t)
where IMF is an AM-FM signal
f
k
(t) = F
k
(t)cos(
k
(t))whereF
k
(t),

k
(t) > 0t
Algorithmic method present to extract modes.
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Algorithm)
consider the signal, say f (t).
Figure: f (t)
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Algorithm)
Identify maxima and join using cubic spline interpolation.
Figure: Computing the upper envelope
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Algorithm)
Identify minima and join using cubic spline interpolation.
Figure: Computing the lower envelope
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Algorithm)
Find the mean
Figure: Mean in bold
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Algorithm)
Compute upper and lower envelopes,f
u
(t) and f
l
(t) ,using
cubic spline interpolation.
Compute mean envelope, m(t) =
f
u
(t)+f
l
(t)
2
.
Candidate r
1
(t) = f (t) m(t).
r
1
(t) is the new signal and same process is continued until
rst IMF is obtained, say f
1
(t) = r
n
(t)
same procedure is repeated for signal f (t) f
1
(t).
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Algorithm)
Compute upper and lower envelopes,f
u
(t) and f
l
(t) ,using
cubic spline interpolation.
Compute mean envelope, m(t) =
f
u
(t)+f
l
(t)
2
.
Candidate r
1
(t) = f (t) m(t).
r
1
(t) is the new signal and same process is continued until
rst IMF is obtained, say f
1
(t) = r
n
(t)
same procedure is repeated for signal f (t) f
1
(t).
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Algorithm)
Compute upper and lower envelopes,f
u
(t) and f
l
(t) ,using
cubic spline interpolation.
Compute mean envelope, m(t) =
f
u
(t)+f
l
(t)
2
.
Candidate r
1
(t) = f (t) m(t).
r
1
(t) is the new signal and same process is continued until
rst IMF is obtained, say f
1
(t) = r
n
(t)
same procedure is repeated for signal f (t) f
1
(t).
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Algorithm)
Compute upper and lower envelopes,f
u
(t) and f
l
(t) ,using
cubic spline interpolation.
Compute mean envelope, m(t) =
f
u
(t)+f
l
(t)
2
.
Candidate r
1
(t) = f (t) m(t).
r
1
(t) is the new signal and same process is continued until
rst IMF is obtained, say f
1
(t) = r
n
(t)
same procedure is repeated for signal f (t) f
1
(t).
Gec Thrissur Empirical Wavelet Transform
Empirical Mode Decomposition (Algorithm)
Compute upper and lower envelopes,f
u
(t) and f
l
(t) ,using
cubic spline interpolation.
Compute mean envelope, m(t) =
f
u
(t)+f
l
(t)
2
.
Candidate r
1
(t) = f (t) m(t).
r
1
(t) is the new signal and same process is continued until
rst IMF is obtained, say f
1
(t) = r
n
(t)
same procedure is repeated for signal f (t) f
1
(t).
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure: original signal
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure: Finding maxima
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure: join using cubic spline interpolation
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure: Finding minima
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure: Join using cubic spline interpolation
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure: Mean
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure: candidate
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure: r
1
(t)
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Illustration
Figure:
Gec Thrissur Empirical Wavelet Transform
Advantages
Doesnt use any prescribed basis.
Self adapting accordingly to the analyzed signal.
Able to separate stationary and non-stationary components
from a signal.
Locality(between maximas and minimas) and
adaptability(data driven).
Gec Thrissur Empirical Wavelet Transform
Advantages
Doesnt use any prescribed basis.
Self adapting accordingly to the analyzed signal.
Able to separate stationary and non-stationary components
from a signal.
Locality(between maximas and minimas) and
adaptability(data driven).
Gec Thrissur Empirical Wavelet Transform
Advantages
Doesnt use any prescribed basis.
Self adapting accordingly to the analyzed signal.
Able to separate stationary and non-stationary components
from a signal.
Locality(between maximas and minimas) and
adaptability(data driven).
Gec Thrissur Empirical Wavelet Transform
Advantages
Doesnt use any prescribed basis.
Self adapting accordingly to the analyzed signal.
Able to separate stationary and non-stationary components
from a signal.
Locality(between maximas and minimas) and
adaptability(data driven).
Gec Thrissur Empirical Wavelet Transform
Disadvantages
Lack of mathematical theory, not well dened.
Algorithmic approach, dicult to model.
Gec Thrissur Empirical Wavelet Transform
Disadvantages
Lack of mathematical theory, not well dened.
Algorithmic approach, dicult to model.
Gec Thrissur Empirical Wavelet Transform
Empirical Wavelet Transform Concept
EMD
Signal=Fast oscillation + Slow oscillation
Iteration
separation of fast and slow oscillations.
data driven.
local analysis based on extremas.
adaptability but no mathematical theory.
Gec Thrissur Empirical Wavelet Transform
Empirical Wavelet Transform Concept
WAVELETS
Signal=Approximation + Details
Iteration
separation of approximation and details.
based on apriori ltering.
global analysis.
strong mathematical background.
Gec Thrissur Empirical Wavelet Transform
Empirical Wavelet Transform Concept
EWT
EWT = EMD + WAVELETS
Idea is to combine the strength of wavelets formalism and EMDs
adaptability.
Gec Thrissur Empirical Wavelet Transform
Cont..
wavelets are equivalent to lter banks
Figure: dyadic decomposition of Fourier spectrum
EWT gives adaptive decomposition of Fourier spectrum
Figure: adaptive decomposition of Fourier spectrum
Gec Thrissur Empirical Wavelet Transform
Empirical Wavelets
Instead of dyadic decomposition of Fourier spectrum, why not
decompose it adaptively.
Method to build a family of wavelets adapted to a signal is
equivalent to building a set of bandpass lters in the Fourier
domain, such that their support depends on information
present in the signal being analyzed .
Lets see how spectrum can be segmented depending on
information present in signal.
Gec Thrissur Empirical Wavelet Transform
Cont..
Figure: Fourier spectrum of a signal
Gec Thrissur Empirical Wavelet Transform
Cont..
Figure: Fourier spectrum segmented
Gec Thrissur Empirical Wavelet Transform
Cont..
Figure: Fourier spectrum segmented(another signal)
Gec Thrissur Empirical Wavelet Transform
Cont..
Aim is to separate dierent portion of the spectrum which
corresponds to modes.
Assume N no of segments which means N 1 extra
boundaries excluding for 0 and .
Algorithm
detect M local maximas, sort in decreasing order.
M N : the algorithm found enough maxima to dene the wanted
number of segments, keep only the rst N 1 maxima.
M < N : the signal has less modes than expected, keep all the
detected maxima and reset N to the appropriate value.
Gec Thrissur Empirical Wavelet Transform
Filter Bank Construction
assume the Fourier support [0, ] is segmented into N
contiguous segments.
denote
n
to be the limits between each segments (where

0
= 0 and
N
= ) and each segment is denoted as

n
= [
n1
,
n
].
Figure: Filter bank construction

n
=half length of transition phase, in practice,
n
=
n
Gec Thrissur Empirical Wavelet Transform
Filter Bank Construction
Scaling function spectrum

n
() =

1 if || (1 )
n
cos

1
2
n
(|| (1 )
n
)

if(1 )
n
|| (1 +)
n
0 otherwise
wavelet spectrum

n
() =

1 if (1 + )
n
|| (1 )
n+1
cos

1
2
n+1
(|| (1 )
n+1
)

if(1 )
n+1
|| (1 + )
n+1
sin

1
2
n
(|| (1 )
n
)

if(1 )
n
|| (1 + )
n
0 otherwise
Gec Thrissur Empirical Wavelet Transform
Filter Bank Reconstruction
Scaling function spectrum,
n
= 1, = 0.5
Wavelet spectrum,
n
= 1,
n+1
= 2.5and = 0.2
Gec Thrissur Empirical Wavelet Transform
Tight frame and an Example
proposition for a tight frame is < min
n
_

n+1

n

n+1
+
n
_
Examples

n
{0, 1.5, 2, 2.8, } with = 0.05
Figure: Fourier partitioning of an empirical lter bank
Gec Thrissur Empirical Wavelet Transform
Empirical Wavelet Transform
Detail coecient
W

f
(n, t) = f ,
n
=
_

f ()

n
()
_

Approximation coecient
W

f
(0, t) = f ,
1
=
_

f ()

n
()
_

Reconstruction
f (t) = W

f
(0, t)
1
(t) +
N

n=1
W

f
(n, t)
n
(t)
=
_

f
(0, )

1
() +
N

n=1

f
(n, )

n
()
_

Gec Thrissur Empirical Wavelet Transform


Cont..
We have already seen
f (t) =
N

k=0
f
k
(t)
Empirical mode from above reconstructed form f
k
f
0
(t) = W

f
(0, t)
1
(t)
f
k
(t) = W

f
(k, t)
k
(t)
Gec Thrissur Empirical Wavelet Transform
Cont..
Algorithm
Finding Fourier transform of signal, f

f .
Compute the local maxima of

f on [0, ] and nd the set
n
.
Choose < min
n
_

n+1

n

n+1
+
n
_
.
Build lter bank.
Filter the signal to get each component.
Gec Thrissur Empirical Wavelet Transform
Examples
Figure: signal 1
Gec Thrissur Empirical Wavelet Transform
Examples
Figure: signal 2
Gec Thrissur Empirical Wavelet Transform
Examples
Figure: signal 3
Gec Thrissur Empirical Wavelet Transform
Examples
Figure: real ECG signal
Gec Thrissur Empirical Wavelet Transform
Examples
Figure: real ECG signal
Gec Thrissur Empirical Wavelet Transform
Examples
Figure: signal 1 modes by EWT and EMD
Gec Thrissur Empirical Wavelet Transform
Examples
Figure: signal 2 modes by EWT and EMD
Gec Thrissur Empirical Wavelet Transform
Examples
Figure: signal 3 modes by EWT and EMD
Gec Thrissur Empirical Wavelet Transform
Examples
Figure: real ECG signal modes by EWT and EMD
Gec Thrissur Empirical Wavelet Transform
Comparing EMD and EWT
EMD automatically estimates number of modes whereas in
EWT, we set number of modes.
EMD overestimates number of modes but EWT gives
dierent components which are closer to original signal .
Gec Thrissur Empirical Wavelet Transform
Comparing EMD and EWT
EMD automatically estimates number of modes whereas in
EWT, we set number of modes.
EMD overestimates number of modes but EWT gives
dierent components which are closer to original signal .
Gec Thrissur Empirical Wavelet Transform
Conclusion
Wavelets built adapted to signal being analyzed.
Wavelet lter bank based on Fourier support detected from
information.
Dilation factors dont follow a prescribed scheme but are
detected empirically.
Further research can be done to segment spectrum even more
eciently .
A matlab toolbox is available in matlab central website.
Gec Thrissur Empirical Wavelet Transform
Conclusion
Wavelets built adapted to signal being analyzed.
Wavelet lter bank based on Fourier support detected from
information.
Dilation factors dont follow a prescribed scheme but are
detected empirically.
Further research can be done to segment spectrum even more
eciently .
A matlab toolbox is available in matlab central website.
Gec Thrissur Empirical Wavelet Transform
Conclusion
Wavelets built adapted to signal being analyzed.
Wavelet lter bank based on Fourier support detected from
information.
Dilation factors dont follow a prescribed scheme but are
detected empirically.
Further research can be done to segment spectrum even more
eciently .
A matlab toolbox is available in matlab central website.
Gec Thrissur Empirical Wavelet Transform
Conclusion
Wavelets built adapted to signal being analyzed.
Wavelet lter bank based on Fourier support detected from
information.
Dilation factors dont follow a prescribed scheme but are
detected empirically.
Further research can be done to segment spectrum even more
eciently .
A matlab toolbox is available in matlab central website.
Gec Thrissur Empirical Wavelet Transform
Conclusion
Wavelets built adapted to signal being analyzed.
Wavelet lter bank based on Fourier support detected from
information.
Dilation factors dont follow a prescribed scheme but are
detected empirically.
Further research can be done to segment spectrum even more
eciently .
A matlab toolbox is available in matlab central website.
Gec Thrissur Empirical Wavelet Transform
References
J. Gilles Empirical wavelet transform, IEEE Trans. Signal
Process., vol. 61, no. 16, pp.3999 -4010 2013 .
I. Daubechies, J. Lu, and H.-T. Wu, Synchrosqueezed wavelet
transforms: An empirical mode decomposition-like tool, J. Appl.
Computat.Harmon. Anal., vol. 30, no. 2, pp. 243261, 2011.
N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q.
Zheng, N.-C. Yen, C. C. Tung, andH.H.Liu, The empirical mode
decomposition and the Hilbert spectrum for nonlinear and
non-stationary time series analysis, Proc. Roy. Soc. London A, vol.
454, pp. 903995, 1998..
ECE 804 - Spring 2012 - Lecture 005 with Dr. Patrick Flandrin,
CNRS &

Ecole Normale Superieure de Lyon, France - Mar. 16,
2012.
Gec Thrissur Empirical Wavelet Transform
Thanks for your patience
Gec Thrissur Empirical Wavelet Transform
APPENDIX
Gec Thrissur Empirical Wavelet Transform
APPENDIX A (orthogonality of IMF)
By virtue of the decomposition, the elements should all be
locally orthogonal to each other, for each element is obtained
from the dierence between the signal and its local mean
through the maximal and minimal envelopes.
X(t) =
n+1

j =1
C
j
(t)
X
2
(t) =
n+1

j =1
C
2
j
(t) + 2
n+1

j =1
n+1

k=1
C
j
(t)C
k
(t)
Gec Thrissur Empirical Wavelet Transform
APPENDIX B (orthogonality of IMF)
Figure: orthogonal IMFs
Gec Thrissur Empirical Wavelet Transform
APPENDIX
(limit for IMF iteration in EMD)
SD =
T

t=0
_
|h
1(k1)
(t) h
1k
(t)|
2
h
2
1(k1)
(t)
_
SD can be set between 0.2 - 0.3
Gec Thrissur Empirical Wavelet Transform
APPENDIX C (Detecting value for N)
Detect {M
i
}
M
k=1
, set of M detected maximas in magnitude of
spectrum so that, M
1
M
2
... M
M
.
keeping all the maximas larger than threshold
M
M
+(M
1
MM), where is relative amplitude ratio.
around 0.3 - 0.4 gives consistent result.
Gec Thrissur Empirical Wavelet Transform
APPENDIX D (function)
(x) =
_
_
_
0 if x 0and(x) +(1 x) = 1
1 ifx 1
eq : (x) = x
4
(35 84x + 70x
2
20x
3
)
Gec Thrissur Empirical Wavelet Transform
APPENDIX E (time frequency representation)
Hilbert-Huang transform.
H
f
(t) =
1

p.v
_
inf
inf
f ()
t
d
.
for f (t) = F(t)cos((t)), it provides f
a
(t) = F(t) exp
i (t)
where f
a
(t) = f (t) + iH
f
(t), .
extract instantaneous amplitude F(t) and freq

(t).
Gec Thrissur Empirical Wavelet Transform

You might also like