Post on 11-Mar-2018
transcript
5/26/2006 1
Beyond the Fourier Transform :Coping with Nonlinear, Nonstationary Time Series
Norden E. HuangGoddard Institute for Data AnalysisNASA Goddard Space Flight Center
5/26/2006 2
Seminar Announcement, Johns Hopkins University, 1998
5/26/2006 3
Jean-Baptiste-Joseph Fourier
1807 “On the Propagation of Heat in Solid Bodies”
1812 Grand Prize of Paris Institute
“Théorie analytique de la chaleur”
‘... the manner in which the author arrives at these equations is not exempt of difficulties and that his analysis to integrate them still leaves something to be desired on the score of generality and even rigor.’
1817 Elected to Académie des Sciences
1822 Appointed as Secretary of Math Section
paper published
Fourier’s work is a great mathematical poem.Lord Kelvin
5/26/2006 4
Fourier Integral
( ) ( ) ;
1( ) ( )
2
i t
i t
F f t e dt
f t F e d
ω
ω
ω
ω ωπ
∞
−∞
∞−
−∞
=
=
∫
∫
5/26/2006 5
Fourier Spectrum
5/26/2006 6
Fourier Series Expansion:
n Any function f(t) can be expanded in terms of discrete sine or cosine functions as
( )01
1( ) cos sin .
2 n n n nn
f t a a t b tω ω∞
=
= + +∑
5/26/2006 7
Random and Delta Functions
5/26/2006 8
Fourier Components : Random Function
5/26/2006 9
Fourier Components : Delta Function
5/26/2006 10
Fourier Sums : Delta Function
5/26/2006 11
Problems with Fourier Expansion
n Linear and Stationary assumptions.n Trigonometric function with constant
frequency and amplitude over the whole time span
n Superposition holds true limited to linear systems.
n Phase information not fully used.n No difference between delta and random
functions in frequency spectral representation.
5/26/2006 12
Data Analysis is equivalent to Information Extraction
n Data is the only connection between us and the realty.
n All our information is contained in the data.
n Data analysis is the means to extract information form the data.
n Unless we have clear understanding of the underlying processes, data analysis should not be based on a priori basis methods.
n Adaptive basis is the best approach to extract the maximum amount information.
n Hilbert-Huang Transform (HHT) is based on an adaptive approach.
n Data analysis is mechanical; result interpretation is the key to yield information.
5/26/2006 13
The Main Data Analysis Tasks
n Distribution: global properties limited to homogeneous population only; HHT can help extract component with homogeneous scale.
n Filtering: mostly Fourier based in frequency space; HHT is a nonlinear time scale based filter.
n Regression: fit data to an a priori functional; HHT fits adaptively with spline.
n Correlation: need to detrend; HHT offers adaptive detrend.
n Spectral Analysis: time-frequency representation; HHT for data from nonlinear and nonstationary processes.
n Prediction: stationary processes; HHT could help here too by provide band-limited components fro easier prediction.
5/26/2006 14
Motivations for a New Method
n Physical processes are mostly nonstationary
n Physical Processes are mostly nonlinear
n Data from observations are invariably too short
n Physical processes are mostly non-repeatable.
∪ Ensemble mean impossible, and temporal mean might not be meaningful for lack of ergodicity. Traditional methods inadequate.
5/26/2006 15
Available Data Analysis Methodsfor Nonstationary (but Linear) time series
n Various probability distributionsn Spectral analysis and Spectrogramn Wavelet Analysisn Wigner-Ville Distributionsn Empirical Orthogonal Functions aka Singular
Spectral Analysisn Moving meansn Successive differentiations
5/26/2006 16
Available Data Analysis Methodsfor Nonlinear (but Stationary and Deterministic) time series
n Phase space method• Delay reconstruction and embedding• Poincaré surface of section• Self-similarity, attractor geometry &
fractals
n Nonlinear Prediction
n Lyapunov Exponents for stability
5/26/2006 17
The Need for Instantaneous Frequency in Nonstationary and Nonlinear Processes
( )
32
2
2
22
d xx cos t
dt
d xx cos t
dt
Spring with positiondependent cons tan t ,int ra wave frequency m o d ulation;therefore,we need ins tan
x
1
tan eous frequenc
x
y .
γε ω
ε γ ω
+ + =
⇒ + =
⇒
+
−
5/26/2006 18
Duffing Pendulum
2
22( co .) s1
d xx tx
d tε γ ω=++
x
5/26/2006 19
( )
p
2 2 1 / 2 1
i ( t )
For any x( t ) L ,
1 x( )y( t ) d ,
t
then, x ( t )and y( t )arecomplexconjugate :
z( t ) x ( t ) i y ( t ) ,
wherey ( t )
a ( t ) x y and ( t
a(
)
t ) e
tan .x ( t )
θ
τ
ττ
π τ
θ −
∈
= ℘−
= + =
= + =
∫
Hilbert Transform : Definition
5/26/2006 20
Hilbert Transform Fit
5/26/2006 21
The Traditional View of the Hilbert Transform for Data Analysis
5/26/2006 22
Traditional Viewa la Hahn (1995) : Data LOD
5/26/2006 23
Traditional Viewa la Hahn (1995) : Hilbert
5/26/2006 24
Traditional Viewa la Hahn (1995) : Phase Angle
5/26/2006 25
Traditional Viewa la Hahn (1995) : Phase Angle Details
5/26/2006 26
Traditional Viewa la Hahn (1995) : Frequency
5/26/2006 27
Why the traditional view does not work?
5/26/2006 28
Hilbert Transform a cos q + b : Data
5/26/2006 29
Hilbert Transform a cos q + b : Phase Diagram
5/26/2006 30
Hilbert Transform a cos q + b : Phase Angle Details
5/26/2006 31
Hilbert Transform a cos q + b : Frequency
5/26/2006 32
The Empirical Mode Decomposition Method and Hilbert Spectral Analysis
Sifting
5/26/2006 33
Empirical Mode Decomposition: Methodology : Test Data
5/26/2006 34
Empirical Mode Decomposition: Methodology : data and m1
5/26/2006 35
Empirical Mode Decomposition: Methodology : data & h1
5/26/2006 36
Empirical Mode Decomposition: Methodology : h1 & m2
5/26/2006 37
Empirical Mode Decomposition: Methodology : h3 & m4
5/26/2006 38
Empirical Mode Decomposition: Methodology : h4 & m5
5/26/2006 39
Empirical Mode DecompositionSifting : to get one IMF component
1 1
1 2 2
k 1 k k
k 1
x ( t ) m h ,
h m h ,
. . . . .
. . . . .h m h
.h c
.−
− =
− =
− =
=⇒
5/26/2006 40
Two Stoppage Criteria : S and SD
A. The S number : S is defined as the consecutive number of siftings, in which the numbers of zero-crossing and extrema are the same for these S siftings.
B. SD is small than a pre-set value, where
2Tk 1 k
2t 0 k 1
h ( t ) h ( t )SD .
h ( t )−
= −
−= ∑
5/26/2006 41
Empirical Mode Decomposition: Methodology : IMF c1
5/26/2006 42
Definition of the Intrinsic Mode Function (IMF)
Any function having the same numbers ofzero cros sin gs and extrema,and also havingsymmetric envelopesdefined by local m a x i m aand min ima respectively isdefined asanIntrinsic M odeFunction( IMF ).
All IMF enjoys good HilbertTransfo
−
i ( t )
rm :
c( t ) a( t ) e θ⇒ ⇒ =
5/26/2006 43
Empirical Mode DecompositionSifting : to get all the IMF components
1 1
1 2 2
n 1 n n
n
j nj 1
x( t ) c r ,
r c r ,
x( t ) c r
. . .r c r .
.
−
=
− =
− =
−⇒ =
− =
∑
5/26/2006 44
Empirical Mode Decomposition: Methodology : data & r1
5/26/2006 45
Empirical Mode Decomposition: Methodology : IMFs
5/26/2006 46
Definition of Instantaneous Frequency
i ( t )
t
The FourierTransform of the Instrinsic M odeFunnction, c( t ), gives
W ( ) a( t ) e dt
By Stationary phaseapproximation we have
d ( t ),
dt
This isdefined as the Ins tan tan eous Frequency .
θ ωω
θω
−=
=
∫
5/26/2006 47
Comparison between FFT and HHT
j
jt
i tj
j
i ( ) d
jj
1 . F F T :
x ( t ) a e .
2 . H H T :
x ( t ) a ( t ) e .
ω
ω τ τ
= ℜ
∫= ℜ
∑
∑
5/26/2006 48
Comparisons: Fourier, Hilbert & Wavelet
5/26/2006 49
Speech AnalysisHello : Data
5/26/2006 50
Four comparsions D
5/26/2006 51
An Example of Sifting
5/26/2006 52
Length Of Day Data
5/26/2006 53
LOD : IMF
5/26/2006 54
Orthogonality Check
n Pair-wise %
n 0.0003n 0.0001n 0.0215n 0.0117n 0.0022n 0.0031n 0.0026n 0.0083n 0.0042n 0.0369n 0.0400
n Overall %
n 0.0452
5/26/2006 55
LOD : Data & c12
5/26/2006 56
LOD : Data & Sum c11-12
5/26/2006 57
LOD : Data & sum c10-12
5/26/2006 58
LOD : Data & c9 - 12
5/26/2006 59
LOD : Data & c8 - 12
5/26/2006 60
LOD : Detailed Data and Sum c8-c12
5/26/2006 61
LOD : Data & c7 - 12
5/26/2006 62
LOD : Detail Data and Sum IMF c7-c12
5/26/2006 63
LOD : Difference Data – sum all IMFs
5/26/2006 64
Traditional Viewa la Hahn (1995) : Hilbert
5/26/2006 65
Mean Annual Cycle & Envelope: 9 CEI Cases
5/26/2006 66
Hilbert’s View on Nonlinear Data
5/26/2006 67
Duffing Type WaveData: x = cos(wt+0.3 sin2wt)
5/26/2006 68
Duffing Type WavePerturbation Expansion
( )( ) ( )
For 1 , we can have
x( t ) cos t sin2 t
cos t cos sin2 t sin t sin sin2 t
cos t sin t sin2 t ....
1 cos t cos 3 t ....2 2
This isvery similar tothe solutionof Duffingequation .
ε
ω ε ω
ω ε ω ω ε ω
ω ε ω ω
ε εω ω
= +
= −
= − +
= − + +
=
5/26/2006 69
Duffing Type WaveWavelet Spectrum
5/26/2006 70
Duffing Type WaveHilbert Spectrum
5/26/2006 71
Duffing Type WaveMarginal Spectra
5/26/2006 72
Duffing Equation
23
2 .
Solved with for t 0 to 200 with1
0.1
od
0.04 H z
Initial condition :[ x ( o ) ,
d xx x c
x ' ( 0 ) ] [ 1
os t
, 1 ]
3
t
e 2
d
tbε
ε γ ω
γω
== −==
=
+ + =
5/26/2006 73
Duffing Equation : Data
5/26/2006 74
Duffing Equation : IMFs
5/26/2006 75
Duffing Equation : IMFs
5/26/2006 76
Duffing Equation : Hilbert Spectrum
5/26/2006 77
Duffing Equation : Detailed Hilbert Spectrum
5/26/2006 78
Duffing Equation : Wavelet Spectrum
5/26/2006 79
Duffing Equation : Hilbert & Wavelet Spectra
5/26/2006 80
What This Means
n Instantaneous Frequency offers a total different view for nonlinear data: instantaneous frequency with no need for harmonics and unlimited by uncertainty.
n Adaptive basis is indispensable for nonstationary and nonlinear data analysis
n HHT establishes a new paradigm of data analysis
5/26/2006 81
Comparisons
noyesyesHarmonics
noyesyesUncertainty
yesyesnoNon-stationary
yesnonoNonlinear
Energy-time-frequency
Energy-time-frequency
Energy-frequency
Presentation
Differentiation:Local
Convolution: Regional
Convolution: Global
Frequency
Adaptivea prioria prioriBasis
HilbertWaveletFourier
5/26/2006 82
Current Applications
n Non-destructive Evaluation for Structural Health Monitoring n (DOT, NSWC, and DFRC/NASA, KSC/NASA Shuttle)
n Vibration, speech, and acoustic signal analysesn (FBI, MIT, and DARPA)
n Earthquake Engineeringn (DOT)
n Bio-medical applicationsn (Harvard, UCSD, Johns Hopkins, and Southampton, UK)
n Global Primary Productivity Evolution map from LandSat data n (NASA Goddard, NOAA)
n Cosmological Gravity Wave and Planets huntingn (NASA Goddard, and Nicholas Copernicus University, Poland)
n Financial market data analysisn (NASA and HKUST)