A New Wavelet-Based Mode Decomposition for
Oscillating Signals and Comparison with the
Empirical Mode Decomposition
Adrien DELIÈGE University of Liège, Belgium
ITNG 2016 - 13th International Conference on Information Technology: New Generations Las Vegas, Nevada – April 2016
Joint work with S. NICOLAY
Introduction
Decomposingtime series into severalmodeshas become more and more
popular andusefulin signal analysis.
Methods such as EMD or SSA (among others) have been successfully
applied inmedicine, finance, climatology, ...
Old but gold: Fouriertransform allows to decompose a signal as
f
(
t) ≈
K
∑
k=1
ckcos(
ω
kt+
φ
k).
Problem: oftentoo many componentsin the decomposition.
Idea: Considering theamplitudes and frequencies as functions of tto
decrease the number of terms.
Development of an empirical wavelet mode decomposition(EWMD)
method which can be used in numerous ways.
Edit: EWMD becomesWIME(wavelet-induced mode extraction).
Table of contents
1 Methods: EMD and WIME
2 Reconstruction skills
3 Period detection skills
4 Real-life data
5 Recent improvements
Methods: EMD and WIME
Table of contents
1 Methods: EMD and WIME
2 Reconstruction skills
3 Period detection skills
4 Real-life data
5 Recent improvements
6 Conclusion
Methods: EMD and WIME
Empirical Mode Decomposition (EMD) in a nutshell
1. Given a signal f , compute m the mean of its upper and lowerenvelopes.
2. Compute h
=
f−
m andrepeatsteps 1.and 2.with h instead of funtilh is“stableenough” and becomes an “intrinsic mode component” (IMF), f1.
3. Repeatthe whole process with f
−
f1instead of f .4. The procedure alwaysstopsand the IMFs successively extracted
reconstructf accurately.
Maindrawbacksinclude: lack of solid mathematical theoretical background, high sensitivity to noise, mode-mixing problems.
Methods: EMD and WIME
Empirical Wavelet Mode Decomposition (EWMD) – WIME
Thewavelet usedin this study is the function
ψ
(
t) =
e iΩt 2√
2π
e −(2Ω8tΩ+2π)2 eπΩt+1
with
Ω =
π
p
2/ln 2, which is similar to theMorletwavelet. TheFouriertransform of
ψ
is given byˆ
ψ
(
ω
) =
sinπω
2Ω e−(ω−Ω) 2 2.
Thewavelet transformof the signal is computed as:
Wf
(
a,
t) =
Z f(
x)¯
ψ
x−
t a dx a,
where
ψ
¯
is the complex conjugate ofψ
, t∈ R
stands for the location/timeparameter and a
>
0 denotes thescaleparameter.Methods: EMD and WIME
Empirical Wavelet Mode Decomposition (EWMD) – WIME
a) Perform the continuouswavelet transformWf
(
a,
t)
of f .b) Compute thewavelet spectrum
Λ
associated to f :Λ(
a) =
E|
Wf(
a, .)
|
where E denotes the mean over time. Then look for thescalea∗at which
Λ
reaches its globalmaximum.c) Extractthe component related to a∗:
c1
=
Rψ−1|
Wf(
a∗,
t)|
cos(arg Wf(
a∗,
t))
where Rψ
≈
1.25 is a reconstruction constant.d) Repeatsteps (a) to (d) with f
−
c1instead of f .e) Stopthe process when
Λ(
a∗) <
ε
for a thresholdε
or at your convenience.Reconstruction skills
Table of contents
1 Methods: EMD and WIME
2 Reconstruction skills
3 Period detection skills
4 Real-life data
5 Recent improvements
6 Conclusion
Reconstruction skills Reconstruction skills First example: 1 1024 −1.5 0 1.5 −1 0 1 −1 0 1 −1 0 1 −1 0 1
Reconstruction skills Reconstruction skills time frequency spectrum frequency −1 0 1 time c1 time frequency spectrum frequency −1 0 1 time c2 time frequency spectrum frequency −1 0 1 time c3
First row:
|
Wf(
a,
t)
|
, spectrum of f , first extracted component c1. Second and third rows: the same for f−
c1and f−
c1−
c2.Reconstruction skills
Reconstruction skills
WIME Expected EMD
−1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1
First column: components extracted with WIME. Second column: the real (expected) components. Third column: the IMFs extracted with the EMD. The reconstruction c1
+
c2+
c3has a correlation of 0.992 with the original signal and a RMSE of 0.085.Period detection skills
Table of contents
1 Methods: EMD and WIME
2 Reconstruction skills
3 Period detection skills
4 Real-life data
5 Recent improvements
6 Conclusion
Period detection skills
Period detection
We consider the AM-FM signal f
(
x) =
∑
4i=1fi
(
x)
where f1(
x) =
1+
0.5 cos 2π
200x cos 2π
47x (1) f2(
x) =
ln(x)
14 cos 2π
31x (2) f3(
x) =
√
x 60 cos 2π
65x (3) f4(
x) =
x 2000cos 2π
23+
cos 16002π xx!
.
(4)Period detection skills Period detection 1 200 400 600 800 −2 −1 0 1 2 −1 0 1 −1 0 1 −1 0 1 1 200 400 600 800 −1 0 1
Target periods:
≈
23, 31, 47, 65 units.Period detection skills
Period detection - WIME
time frequency spectrum frequency −1 0 1 time c1 time frequency spectrum frequency −1 0 1 time c2 time frequency spectrum frequency −1 0 1 time c3 time frequency spectrum frequency −1 0 1 time c4
Target periods:
≈
23, 31, 47, 65 units. Detected periods corresponding to a∗:Period detection skills
Period detection - EMD
−1 0 1 −1 0 1 1 200 400 600 800 −1 0 1 1 200 400 600 800 −1 0 1
Target periods:
≈
23, 31, 47, 65 units. Periods extracted from theHilbert-Huang transform: 41, 75, 165, 284 units.
Real-life data
Table of contents
1 Methods: EMD and WIME
2 Reconstruction skills
3 Period detection skills
4 Real-life data
5 Recent improvements
Real-life data
ENSO index
Analyzed data: Niño 3.4 time series, i.e. monthly-sampled sea surface temperature anomalies in the Equatorial Pacific Ocean from Jan 1950 to Dec 2014 (http://www.cpc.ncep.noaa.gov/).
Real-life data ENSO index Niño 3.4 index: 1950 1960 1970 1980 1990 2000 2010 2020 −2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 time (years) Nino 3.4 index
17 El Niño events: SST anomaly above
+0.5
◦C during 5 consecutivemonths.
14 La Niña events: SST anomaly below
−
0.5◦C during 5 consecutiveReal-life data
ENSO index
Floodingin the West coast of South America
Droughtsin Asia and Australia
Fish killsor shifts in locations and types of fish, havingeconomic impacts
in Peru and Chile
Impact on snowfalls andmonsoons, drier/hotter/wetter/cooler than normal
conditions
Impact onhurricanes/typhoonsoccurrences
Links with famines, increase inmosquito-borne diseases(malaria,
dengue, ...), civil conflicts
In Los Angeles, increase in the number of some species of mosquitoes (in 1997 notably).
...
Real-life data ENSO index time frequency spectrum frequency −1 0 1 time c1 time frequency spectrum frequency −1 0 1 time c2 time frequency spectrum frequency −1 0 1 time c3 time frequency spectrum frequency −1 0 1 time c4 time frequency spectrum frequency −1 0 1 time c5
Real-life data ENSO index −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1
Components extracted with WIME and the IMFs from the EMD. Periods: 44.8, 28.6, 17, 65.6, 140.6 months (WIME) and 9.8, 21, 38.6, 75.9, 138.4 (EMD). Those from WIME are more in agreement with some previous works.
Recent improvements
Table of contents
1 Methods: EMD and WIME
2 Reconstruction skills
3 Period detection skills
4 Real-life data
5 Recent improvements
Recent improvements
Recent improvements
Problems with “highly” non-stationary components (classic example of the EMD): 1 2000 −0.5 0 0.5 1 −0.5 0 0.5 −0.5 0 0.5 −0.5 0 0.5 −0.5 0 0.5 −0.5 0 0.5 −0.5 0 0.5
Reconstruction is excellent (RMSE
=
0.08, PCC=
0.97) but components arenot correct.
Recent improvements
Recent improvements
Solution: add flexibility, follow ridges of maxima.
0 0.5 1 20 40 80 160 320 640 time frequency 0 0.2 0.41 0.62 20 40 80 160 320 640 spectrum frequency 0 0.5 1 −1 0 1 time c1 0 0.5 1 −1 0 1 time f1 0 0.5 1 20 40 80 160 320 640 time frequency 0 0.2 0.41 0.62 20 40 80 160 320 640 spectrum frequency 0 0.5 1 −1 0 1 time c2 0 0.5 1 −1 0 1 time f2 0 0.5 1 20 40 80 160 320 640 time frequency 0 0.2 0.41 0.62 20 40 80 160 320 640 spectrum frequency 0 0.5 1 −1 0 1 time c3 0 0.5 1 −1 0 1 time f3
Recent improvements
Recent improvements
Crossings in the TF plane:
f1
(
t) =
1.25 cos−(
t−
3)3+
180t (5) f2(
t) =
cos 1.8t+
20t (6) f3(
t) =
0.75 cos(−
1.6π
t2+
20π
t)
(t<
5) 0.75 cos(4π
t)
(t≥
5) (7) 0 10 20 −3 −2 −1 0 1 2 3 time fRecent improvements Recent improvements WIME 0 10 20 1.25 2.5 5 10 20 40 80 time frequency 0 0.39 0.78 1.17 1.25 2.5 5 10 20 40 80 spectrum frequency 0 10 20 −2 −1 0 1 2 time c1 0 10 20 −2 −1 0 1 2 time f1 0 10 20 1.25 2.5 5 10 20 40 80 time frequency 0 0.39 0.78 1.17 1.25 2.5 5 10 20 40 80 spectrum frequency 0 10 20 −2 −1 0 1 2 time c2 0 10 20 −2 −1 0 1 2 time f2 0 10 20 1.25 2.5 5 10 20 40 80 time frequency 0 0.39 0.78 1.17 1.25 2.5 5 10 20 40 80 spectrum frequency 0 10 20 −2 −1 0 1 2 time c3 0 10 20 −2 −1 0 1 2 time f3
Recent improvements Recent improvements EMD 0 10 20 1.25 2.5 5 10 20 40 80 time frequency 0 10 20 −2 −1 0 1 2 time IMF1 0 10 20 1.25 2.5 5 10 20 40 80 time frequency 0 10 20 −2 −1 0 1 2 time IMF2 0 10 20 1.25 2.5 5 10 20 40 80 time frequency 0 10 20 −2 −1 0 1 2 time IMF3
Conclusion
Table of contents
1 Methods: EMD and WIME
2 Reconstruction skills
3 Period detection skills
4 Real-life data
5 Recent improvements
Conclusion
Conclusion
WIME
is a wavelet-baseddecompositionmethod
hasreconstruction skillscomparable to the EMD
hasperiod detection skillsbetter than the EMD in mode-mixing problems
handlescrossing patternsin the TF plane
gives accurate results withreal-lifesignals
Future works: study the mathematical properties of the method, test its tolerance to noise, limit edge effects,...
Thanks for your attention.
Please refer to the paper for relevant references.