sep.2008disp lab @md5311 time-frequency analysis 時頻分析 speaker: wen-fu wang 王文阜 ...
TRANSCRIPT
Sep.2008 DISP Lab @MD531 1
Time-Frequency Analysis
時頻分析 Speaker: Wen-Fu Wang 王文阜 Advisor: Jian-Jiun Ding 丁建均 教授 E-mail: [email protected] Graduate Institute of Communication Engineering National Taiwan University, Taipei, Taiwan, ROC
Sep.2008 DISP Lab @MD531 2
Outline
Introduction Short Time Fourier Transform Gabor Transform Wigner Distribution Gabor-Wigner Transform Cohen’s Class Time-Frequency
Distribution
Sep.2008 DISP Lab @MD531 3
Outline
S Transform Hilbert-Huang Transform Applications of Time-Frequency
Analysis Conclusions References
Sep.2008 DISP Lab @MD531 4
IntroductionFourier Transform (FT)
Example :x(t) = cos( t) when t < 10,x(t) = cos(3 t) when 10 t < 20,x(t) = cos(2 t) when t 20
0 5 10 15 20 25 30-1
-0.5
0
0.5
1
-5 0 5-2
-1
0
1
2f(t)
Fouriertransform
Sep.2008 DISP Lab @MD531 5
Introduction Instantaneous Frequency If
then the instantaneous frequency of f (t) are
If order of >1, then instantaneous frequency varies with time
1
( ) exp( ( ))N
k kk
x t a j t
31 2 '( ) '( )'( ) '( ), , , ,
2 2 2 2Nt tt t
( )k t
Sep.2008 DISP Lab @MD531 6
Introduction Example :
→ chirp function
Instantaneous frequency =
2cot( )2ej
t
cot( )
2
t
Sep.2008 DISP Lab @MD531 7
Introduction Time-Frequency analysis Example :
0 5 10 15 20 25 30
-5
-4
-3
-2
-1
0
1
2
3
4
5
t -axis
f -a
xis
-5 0 5-0.5
0
0.5
f -axis
am
plit
ude
Sep.2008 DISP Lab @MD531 8
Short Time Fourier Transform
The earliest Time-Frequency representation was the short time Fourier transform (STFT)
This scheme divides the temporal signal into a series of small overlapping pieces.
The STFT of a function is defined by1
( , ) ( ) ( )2
jX t x t h t e d
Sep.2008 DISP Lab @MD531 9
Short Time Fourier Transform
is the window function. The principle of STFT
( )h t
time
s(t)h(t)
FT FT FT
time
frequency
STFT
Sep.2008 DISP Lab @MD531 10
Short Time Fourier Transform
Example :
The signal of simulation
0 1 2 3 4 5 6 7 8 9 10 11
sec
-1
-0.5
0
0.5
1
Am
plit
ud
e
STFT
0 1 2 3 4 5 6 7 8 9 10 11
sec
0
2
4
6
8
10
12
14
Hz
cos(2 8 ) 1 4
( ) cos(2 4 ) 4 9
cos(2 5 ) 9 12
t t
x t t t
t t
Sep.2008 DISP Lab @MD531 11
Short Time Fourier Transform
Advantage:
(1) Least computation time for digital implementation compared with other
(2) Its ability to avoid cross-term problem
Sep.2008 DISP Lab @MD531 12
Gabor Transform
In fact, Gabor transform is a special case of STFT.
When the of STFT, it can be rewritten as
Why does it choose the Gaussian function as a window?
2( ) exp( )h t t
2 2 ( )( ) 2( , ) ( )t
j ftxG t f e e x d
Sep.2008 DISP Lab @MD531 13
Gabor Transform
The principle of Gabor Transform
time
s(t)h(t)
FT FT FT
time
frequency
STFT
Sep.2008 DISP Lab @MD531 14
Gabor Transform
Example:
Gabor
0 1 2 3 4 5 6 7 8 9 10 11
sec
0
2
4
6
8
10
12
14
Hz
The signal of simulation
0 1 2 3 4 5 6 7 8 9 10 11
sec
-1
-0.5
0
0.5
1
Am
plit
ud
e
cos(2 8 ) 1 4
( ) cos(2 4 ) 4 9
cos(2 5 ) 9 12
t t
x t t t
t t
Sep.2008 DISP Lab @MD531 15
Gabor Transform
Advantage:(1) Its ability to avoid cross-term problem(2) The resolution is better than STFT
Sep.2008 DISP Lab @MD531 16
Wigner Distribution
The Wigner distribution is defined as
In terms of the spectrum, it is
is the Fourier transform of and * means the complex conjugate.
* 2( , ) ( / 2) ( / 2) j fxW t f x t x t e d
* 2( , ) ( / 2) ( / 2) jxW t f X f X f e d
( )X f ( )x t
Sep.2008 DISP Lab @MD531 17
Wigner Distribution
WD has much better time-frequency resolution
WD is not a linear distribution WD has more computation time If the signal is composed by several
time-frequency components, additional interference will be produced.
Sep.2008 DISP Lab @MD531 18
Wigner Distribution
Example:
0 5 10 15 20 25 30
-6
-4
-2
0
2
4
6
cos( ) 1 10
( ) cos(3 ) 10 20
cos(2 ) 20 33
t t
x t t t
t t
Sep.2008 DISP Lab @MD531 19
Wigner Distribution
The inner interference is caused by interference between positive and negative frequency of the signal itself
In order to reduce to inner interference problem, Wigner Ville Distribution can be used
Sep.2008 DISP Lab @MD531 20
Wigner Distribution
The outer interference is caused by mutual interference of multi-component in signal
In order to reduce to outer interference problem, Modified Wigner Distribution can be used
Sep.2008 DISP Lab @MD531 21
Wigner Ville Distribution
Due the inner interference is caused by interference between positive and negative frequency of the signal itself
We can use analytic version signal to replace the original signal for filtering out negative frequency
Sep.2008 DISP Lab @MD531 22
Wigner Ville Distribution
We denote the analytic signal of real valued signal by
is Hilbert transform of We can redefine WD by analytic
signal
( ) ( ) ( )x t x t jx t
* 2( , ) ( / 2) ( / 2) j f
xW t f x t x t e d
( )x t ( )x t
( )x t ( )x t
Sep.2008 DISP Lab @MD531 23
Modified Wigner Distribution
Due outer interference is caused by mutual interference of multi-component in signal
We can select suitable window function is a way to suppress the outer interference, but retain the sharpness of auto terms.
Sep.2008 DISP Lab @MD531 24
Modified Wigner Distribution
The Modified Wigner Distribution is defined as
When the window function , it also become Wigner distribution.
* 2
*
( , ) ( / 2) ( / 2) ( / 2) ( / 2)
( , / 2) ( , / 2)
j fxW t f w w x t x t e d
Y t f Y t f d
Sep.2008 DISP Lab @MD531 25
Modified Wigner Distribution
Example: WD MWD
0 5 10 15 20 25 30
-6
-4
-2
0
2
4
60 5 10 15 20 25 30
-6
-4
-2
0
2
4
6
Sep.2008 DISP Lab @MD531 26
Modified Wigner Ville Distribution
Combined form of the WVD and MWD The advantage of WVD is filtering out
inner interference The advantage of MWD is suppressing
the outer interference, but retain the sharpness of auto terms
The MWVD can avoid the inner and outer interference at the same time
Sep.2008 DISP Lab @MD531 27
Gabor-Wigner Transform
We have compared the properties of the WD and Gabor transform in previous section
The advantage of WD is its high clarity, and the disadvantage of WD is it’s the cross-term problem
Sep.2008 DISP Lab @MD531 28
Gabor-Wigner Transform
In contrast, the advantage of Gabor transform is its ability to avoid cross-term problem, but its clarity is not as good as that of the WDF
The Gabor-Wigner transform (GWT) can achieve the higher clarity and avoiding cross-term problem at the same time.
-10 -5 0 5 10-10
-5
0
5
10
-10 -5 0 5 10-10
-5
0
5
10(c) (d) -10 -5 0 5 10
-10
-5
0
5
10
-10 -5 0 5 10-10
-5
0
5
10(a) (b)
Sep.2008 DISP Lab @MD531 29
Gabor-Wigner Transform
Combined form of the WDF and the Gabor transform
(a)
(b)
(c)
(d)
( , ) ( , ) ( , )f f fC t G t W t 2
( , ) min( ( , ) , ( , ) )f f fC t G t W t
( , ) ( , ) { ( , ) 0.25}f f fC t W t G t
0.7 2.6( , ) ( , ) ( , )f f fC t W t G t
Sep.2008 DISP Lab @MD531 30
Cohen’s Class Time-Frequency Distribution
The definition of the Cohen class is
is the ambiguity function
How does the Cohen’s class distribution avoid the cross term?
( , ) ( , ) ( , ) exp( 2 ( ))x xC t f A j t f d d
* 2( , ) ( / 2) ( / 2) j txA x t x x e dt
Sep.2008 DISP Lab @MD531 31
For the ambiguity function:(1) The auto term is always near to the origin (2) The cross-term is always far from the origin
AF WD
Cohen’s Class Time-Frequency Distribution
t
f
IFTf FTt
IFTf FTt
IFTf FTt
Sep.2008 DISP Lab @MD531 32
Cohen’s Class Time-Frequency Distribution
Relationship between Wigner distribution and ambiguity function
( , )xWD t f
( , )xC t
( / 2) ( / 2)x t x t ( , )xA
( , )xS f
Sep.2008 DISP Lab @MD531 33
Cohen’s Class Time-Frequency Distribution
How does the Cohen’s class distribution avoid the cross term?
Choi-Williams Distribution
Cone-Shape Distribution
2, exp
2, sin exp 2c
Sep.2008 DISP Lab @MD531 34
Cohen’s Class Time-Frequency Distribution
Example:
Ambiguity Function Choi-Williams Distributiontau (sec)
eta
-15 -10 -5 0 5 10 15
-8
-6
-4
-2
0
2
4
6
8
tau(sec)
eta
-15 -10 -5 0 5 10 15
-8
-6
-4
-2
0
2
4
6
8
Sep.2008 DISP Lab @MD531 35
Cohen’s Class Time-Frequency Distribution
Example:
Ambiguity Function Wigner Distribution
tau(sec)
eta
-15 -10 -5 0 5 10 15
-5
0
5
t(sec)
freq
uenc
y
-8 -6 -4 -2 0 2 4 6 8
-5
0
5
tau(sec)
eta
-15 -10 -5 0 5 10 15
-5
0
5
t(sec)
freq
uenc
y
-8 -6 -4 -2 0 2 4 6 8
-5
0
5
Sep.2008 DISP Lab @MD531 36
Cohen’s Class Time-Frequency Distribution Some popular distributions and their kernels
TFD Kernel Formulation
Page
Levin
Kirkwood
Spectrogram
Wigner-Ville 1
Choi-Williams
Cone shape
( , )
/2je 2
'( ') 't j ts t e dt
t
1cos( )
2 *Re ( ) ( ) js t S e
/2je *( ) ( ) js t S e
2
( ) js h t e d *1 1
( ) ( )2 2
js t s t e d
2 2
e
22sin ( ) ec 221
e ( )t
Sep.2008 DISP Lab @MD531 37
Cohen’s Class Time-Frequency Distribution
Advantage:The Cohen’s class distribution may avoid the cross term and has higher clarity
Disadvantage:It requires more computation time and lacks of well mathematical properties
Sep.2008 DISP Lab @MD531 38
S Transform
Unlike STFT, the width of S transform’s window changes with frequency.
Closely related to the wavelet transform
2 22( )
( , ) ( ) exp[ ]22
i ftf f ts f h t e dt
Sep.2008 DISP Lab @MD531 39
S Transform
Example:
50 100 150 200 250 300
0
50
100
150
cos( ) 1 10
( ) cos(3 ) 10 20
cos(2 ) 20 33
t t
x t t t
t t
Sep.2008 DISP Lab @MD531 40
Hilbert-Huang Transform
Traditional data analysis methods are all based on linear and stationary assumptions
The HHT consists of two parts: empirical mode decomposition (EMD) and Hilbert spectral analysis (HSA).
Sep.2008 DISP Lab @MD531 41
Hilbert-Huang Transform---EMD
Intrinsic Mode Functions (IMF)
Sep.2008 DISP Lab @MD531 42
Hilbert-Huang Transform--EMD
The Sifting Process
10 20 30 40 50 60 70 80 90 100 110 120
-2
-1
0
1
2
IMF 1; iteration 0
10 20 30 40 50 60 70 80 90 100 110 120
-2
-1
0
1
2
IMF 1; iteration 0
Sep.2008 DISP Lab @MD531 43
Hilbert-Huang Transform--EMD
10 20 30 40 50 60 70 80 90 100 110 120
-2
-1
0
1
2
IMF 1; iteration 0
10 20 30 40 50 60 70 80 90 100 110 120
-2
-1
0
1
2
IMF 1; iteration 0
Sep.2008 DISP Lab @MD531 44
Hilbert-Huang Transform--EMD
10 20 30 40 50 60 70 80 90 100 110 120
-2
-1
0
1
2
IMF 1; iteration 0
10 20 30 40 50 60 70 80 90 100 110 120
-1.5
-1
-0.5
0
0.5
1
1.5
residue
Sep.2008 DISP Lab @MD531 45
Hilbert-Huang Transform--EMD
Sep.2008 DISP Lab @MD531 46
Hilbert-Huang Transform--HSA
We have obtained the intrinsic mode function components by EMD process method
Then we will do the Hilbert transform to each IMF component.
Sep.2008 DISP Lab @MD531 47
Applications of Time-Frequency Analysis
(1) Finding the Instantaneous Frequency
(2) Sampling Theory (3) Modulation and Multiplexing (4) Filter Design (5) Signal Representation (6) Random Process Analysis
Sep.2008 DISP Lab @MD531 48
Applications of Time-Frequency Analysis
(7) Acoustics (8) Data Compression (9) Spread Spectrum Analysis (10) Radar Signal Analysis (11) Biomedical Engineering (12) Economic Data Analysis
Sep.2008 DISP Lab @MD531 49
Conclusions
Advantage Disadvantage
STFT
and
Gabor transform
1. Low computation
2. The range of the integration is
limited
3. No cross term
4. Linear operation
1. Complex value
2. Low resolution
Wigner
distribution
function
1. Real
2. High resolution
3. If the time/frequency limited,
time/frequency of the WDF is
limited with the same range
1. High computation
2. Cross term
3. Non-linear operation
Cohen’s
class
distribution
1. Avoid the cross term
2. Higher clarity
1. High computation
2. Lack of well
mathematical
properties
Gabor-Wigner
distribution
function
1. Combine the advantage of the
WDF and the Gabor transform
2. Higher clarity
3. No cross-term
1. High computation
Sep.2008 DISP Lab @MD531 50
Conclusions Compare with Fourier, wavelet and HHT analyses
Fourier Wavelet Hilbert
Basis A priori A priori Adaptive
Frequency Convolution: global Convolution: regional Differentiation: local
Presentation Energy-frequency Energy-time-frequency Energy-time-frequency
Non-linear No No Yes
Non-stationary No Yes Yes
Uncertainty Yes Yes No
Harmonics Yes Yes No
Feature extraction No Discrete: no
Continuous: yes
Yes
Theoretical base Theory complete Theory complete Empirical
Sep.2008 DISP Lab @MD531 51
Conclusions
Advantage: The instantaneous frequency can be
observed Disadvantage:
Higher complexity for computation Which method is better?
Sep.2008 DISP Lab @MD531 52
References
N. E. Huang, Z. Shen and S. R. Long, et al., "The Empirical Mode Decomposition and the Hilbert Spectrum for Nonlinear and Non-Stationary Time Series Analysis", Proc. Royal Society, vol. 454, pp.903-995, London, 1998
N. E. Huang, S. Shen, "Hilbert-Huang Transform and its Applications" , World scientific, Singapore, 2005.
Sep.2008 DISP Lab @MD531 53
References S. C. Pei and J. J. Ding, “
Relations between the fractional operations and the Wigner distribution, ambiguity function,” IEEE Trans. Signal Processing, vol. 49, no. 8, pp. 1638-1655, Aug. 2001.
L. Cohen, "Time-Frequency distributions-A review” Proc. IEEE, Vol. 77, No. 7, pp. 941-981, July 1989.
C. H. Page, “Instantaneous Power Spectra,” National Bureau of Standards, Washington, D. C., 1951
Sep.2008 DISP Lab @MD531 54
References S. C. Pei and J. J. Ding, “Relations between
Gabor transforms and fractional Fourier transforms and their applications for signal processing,” IEEE Trans. Signal Processing, vol. 55, no. 10, pp. 4839-4850, Oct. 2007
R. G. Stockwell, L. Mansinha, and R. P. Lowe, “Localization of the complex spectrum: the S transform,” IEEE Trans. Signal Processing, vol. 44, no. 4, pp. 998–1001, Apr. 1996.