An Introduction to Time-Frequency Analysis Speaker: Po-Hong Wu Advisor: Jian-Jung Ding Digital Image and Signal Processing Lab GICE, National Taiwan University NTU, GICE, MD531, DISP Lab 1 Outline • • • • • • • • • • • • Introduction Short-Time Fourier Transform Gabor Transform Wigner Distribution Function Spectrogram S Tranform Cohen’s Class Time-Frequency Distribution Fractional Fourier Transform Motion on Time-Frequency Distributions Hilbert-Huang Transform Conclusion Reference NTU, GICE, MD531, DISP Lab 2 Introduction Fourier transform (FT) X f x t e j 2 f t dt t varies from ∞~∞ Time-Domain Frequency Domain f(t) 1 2 0.5 Fourier transform 1 0 0 -0.5 -1 [A1] -1 0 5 10 15 20 25 30 -2 -5 0 5 Why do we need time-frequency transform? NTU, GICE, MD531, DISP Lab 3 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 -5 -4 -3 -2 -1 0 1 2 3 [B2] 4 5 0 5 10 15 NTU, GICE, MD531, DISP Lab 20 25 30 4 Short Time Fourier Transform X t , f w t x e j 2 f d w(t): mask function 也稱作 windowed Fourier transform or time-dependent Fourier transform NTU, GICE, MD531, DISP Lab 5 When w(t) is a rectangular function w(t) = 1 for |t| B , w(t) = 0 , otherwise X t, f tB t B x e j 2 f d STFT 14 12 Hz 10 8 6 4 2 0 [B3] 0 1 2 3 4 5 6 7 8 9 10 11 sec NTU, GICE, MD531, DISP Lab 6 Advantage: less computation time Disadvantage: worse representaion Application: deal with large data Ex: real time processing NTU, GICE, MD531, DISP Lab 7 Gabor Transform A specail case of the STFT 2 where w(t ) exp( t ) Gx t , f e ( t ) e j 2 f x d 2 Other definition 1 Gx t , 2 e ( t )2 2 e t j ( ) 2 x d Gabor 14 12 Hz 10 8 6 4 2 0 [B4] 0 1 2 3 4 5 6 7 8 9 10 11 sec NTU, GICE, MD531, DISP Lab 8 Why do we choose the Guassian function? Among all functions of w(t), the Gaussian function has area in time-frequency distribution is minimal than other STFT. Gaussian function is an eigenfunction of Fourier transform, so the Gabor transform has the same properties in time domain and in frequency domain. NTU, GICE, MD531, DISP Lab 9 Approximation of the Gabor Transform Because of e Gx t , f a 2 t 1.9143 t 1.9143 Because of e 0.00001 e a2 / 2 1 Gx ,3 t , 2 ( t ) 2 j 2 f e 0.00001 ( t )2 t 4.7985 2 t 4.7985 when |a|>1.9143 e NTU, GICE, MD531, DISP Lab x d when |a|>4.7985 e t j ( ) 2 x d 10 Generalization of the Gabor Transform Gx t , f e ( t ) 2 j 2 f e x d For larger σ: higher resolution in the time domain but lower resolution in the frequency domain For smaller σ: higher resolution in the frequency domain but lower resolution in the time domain NTU, GICE, MD531, DISP Lab 11 Resolution • Using the generalized Gabor transform with larger σ • Using other time unit instead of second NTU, GICE, MD531, DISP Lab 12 Wigner Distribution Function Wx t , f x t / 2 x* t / 2 e j 2 f d Other definition Wx (t , f ) X ( f / 2) X * ( f / 2)e j 2 d Wigner distribution 6 frequency (Hz) 4 2 0 -2 -4 -6 [B5] 0 5 10 15 20 25 30 time (sec) NTU, GICE, MD531, DISP Lab 13 Signal auto-correlation function Cx t , x t / 2 x t / 2 Spectrum auto-correlation function S x , f X f / 2 X f / 2 Ambiguity function (AF) Ax , x t / 2 x* t / 2 e j 2 t dt IFTf IFTf Wx(t, f ) FTt FTt Cx(t, ) Ax(, ) FTt Sx(, f ) NTU, GICE, MD531, DISP Lab [B6] IFTf 14 Modified Wigner Distribution Wigner Ville Distribution For compressing inner interference Wx (t , f ) x(t / 2) x (t / 2)e j 2 f d * Analytic signal x(t ) x(t ) jx(t ) NTU, GICE, MD531, DISP Lab 15 Pseudo Wigner Distribution For surpressing outer interference Wx (t , f ) w( / 2) w( / 2)x(t / 2) x* (t / 2)e j 2 f d Y (t , f / 2)Y * (t , f / 2) d where Y (t , f ) w( ) x(t )e j 2 f d Pseudo L-Wigner distribution 6 4 frequency (Hz) 2 0 -2 [B7] -4 -6 0 5 10 15 20 NTU, GICE, MD531, DISP Lab time (s) 25 30 16 Gabor-Wigner Distribution C f (t , w) p(G f (t , w),W f (t , w)), (a) 10 (b) 10 5 5 0 0 -5 -5 -10 -10 -5 0 5 10 -10 -10 p( x, y ) xy NTU, GICE, MD531, DISP Lab -5 0 5 [B8] 10 p ( x, y ) min( x , y ) 2 17 Spectrogram SP(t , f ) x( )h(t )e j 2 f d 2 Another form SP(t , f ) Wh ( t , f )Wx ( , )d d Spectrogram 14 12 Hz 10 8 6 4 2 0 0 1 2 3 4 5 6 7 sec NTU, GICE, MD531, DISP Lab 8 9 10 11 [B9] 18 S-Transform Original S-Transform f 2 ( t )2 i 2 ft ST ( , f ) x(t ) exp[ ]e dt 2 2 f f 2 ( t )2 exp[ ] Where w(t)= 2 2 f S Transform 4 3 2 frequency cos( t ) when 1 t 10 x(t ) cos(3 t ) when 10 t 20 cos(2 t) when 20 t 30 1 0 -1 -2 -3 [B10] -4 0 5 10 15 20 25 30 time (sec) NTU, GICE, MD531, DISP Lab 19 Generalized S-Transform GS ( , f , p) x(t )w( t , f , p)ei 2 t dt Another definition GS ( , f , p) X ( f )W ( , f , p)ei 2 d Ristriction w( t , f , p)d 1 NTU, GICE, MD531, DISP Lab 20 Novel S-Transform with the Special Varying Window wS ( t , f ) F ( f )( t ) 2 exp( ) 2 2 F( f ) Restriction ws ( t , f )d 1 2 F ( f ) 1/ When When F ( f ) , it becomes the Gabor transform. f 2 , it becomes the original S-trnasform. NTU, GICE, MD531, DISP Lab 21 Cohen’s Class Time-Frequency Distribution Cx (t , f ) Ax (, )(, )exp( j 2 (t f ))d d Ambiguity function Ax ( , ) x(t / 2) x* ( x / 2)e j 2 t dt Cx (t , ) FTt IFTf FTt IFTf WDx (t , f ) Ax ( , ) x(t / 2) x (t / 2) [B11] FTt S x ( , f ) NTU, GICE, MD531, DISP Lab IFTf 22 For the ambiguity function The auto terms are always near to the origin. The cross terms are always from the origin. [B12] NTU, GICE, MD531, DISP Lab 23 Kernel function ( , ) • Choi-Williams Distribution exp 2 , Choi-Williams distribution 10 8 6 4 eta 2 0 -2 -4 [B13] -6 -8 -10 NTU, GICE, MD531, DISP Lab -15 -10 -5 0 tau (sec) 5 10 15 24 • Cone-Shape Distribution 2 2 (, ) sin c( )e >0 Cone Shape distribution 10 8 6 4 eta 2 0 -2 -4 -6 -8 -10 -15 -10 -5 0 5 10 15 tau (sec) NTU, GICE, MD531, DISP Lab 25 Fractional Fourier Transform FT x t X f FT FT x t x t FT FT FT x t X f IFT f t FT FT FT FT x t x t How to rotate the time-frequency distribution by the angle other than /2, , and 3/2? NTU, GICE, MD531, DISP Lab 26 • Zero rotation: R I 0 • Consistency with Fourier transform: /2 R = FT • Additivity of rotation: R R R • rotation: 2 R I NTU, GICE, MD531, DISP Lab 27 X u 1 j cot 2 1 2 f(t): rectangle 0 -1 -5 2 0 5 j 2 csc u t 2 1 1 0 0 -1 -5 2 e j cot u e 2 0 5 1 1 1 0 0 0 -1 -5 -1 -5 -1 -5 x t dt -1 -5 2 e j cot t 2 0 5 F(w): sinc function 0 5 0 5 0 5 [A3] NTU, GICE, MD531, DISP Lab 28 Application Decomposition in the time-frequency distribution 1.5 1 2 x(t) = signal + noise Fourier transform of x(t) 1 0.5 0 0 -0.5 -10 2 1 (non-separable) -5 0 5 10 -1 -10 0 5 10 5 10 1.5 fractional Fourier transform of x(t) 1 0 (separable) -5 recovered signal 0.5 0 -1 -10 -5 0 5 10 -0.5 -10 NTU, GICE, MD531, DISP Lab -5 0 29 f-axis Signa l noise Signal noise FRFT FRFT t-axis noise Signa l cutoff line cutoff line NTU, GICE, MD531, DISP Lab 30 Modulation and Multiplexing 5 5 unfilled T-F slot 0 -5 -20 0 -5 0 20 -20 (c) WDF of G(u) NTU, GICE, MD531, DISP Lab 0 20 (d) GWT of G(u) 31 • Time domain • Modulation • • • Frequency domain fractional domain Shifting Shifting Modulation Differentiation −j2f Modulation + Shifting Modulation + Shifting j2f Differentiation and j2f Differentiation Differentiation and −j2f NTU, GICE, MD531, DISP Lab 32 Motion on Time-Frequency Distributions Horizontal Shifting x (t t0 ) S x (t t0 , f ) e j 2 f t0 Wx ( t t 0 , f ) ,STFT, Gabor ,Wigner Vertical Shifting e j 2 f 0 t x ( t ) S x ( t , f f0 ) ,STFT,Gabor Wx ( t , f f0 ) ,Wigner NTU, GICE, MD531, DISP Lab 33 Dilation 1 t t x ( ) S x ( , af ) ,STFT,Gabor |a| a a Wx ( t , af ) ,WDF a Shearing 2 x (t ) e j at y (t ) S x (t , f ) S y (t , f at ) ,STFT,Gabor Wx (t , f ) W y (t , f at ) ,WDF 2 j t x (t ) e a y (t ) S x (t , f ) S y (t af , f ) ,STFT,Gabor Wx (t , f ) W y (t af , f ) ,WDF NTU, GICE, MD531, DISP Lab 34 Rotation If F{x(t)}=X(f), then F{X(t)}=x(-f). We can derive: | S X (t , f ) || S x ( f , t ) | G X (t , f ) Gx ( f , t )e W X (t , f ) Wx ( f , t ) ,STFT j 2 ft ,Gabor ,WDF NTU, GICE, MD531, DISP Lab 35 Hilbert-Huang Transform Introduction Most of distribution are designed for stationary and linear signals, but, In the real world, most of signals are non-stationary and non-linear. HHT consists two parts: empirical mode decomposition (EMD) Hilbert spectral analysis (HSA) NTU, GICE, MD531, DISP Lab 36 Empirical decomposition function Any complicated data can be decomposed into a finite and small number of intrinsic mode functions (IMF) by sifting processing. Intrinsic mode function (1)In the whole data set, the number of extrema and the number of zero-crossing must either equal or differ at most by one. (2)At any point, the mean value of the envelope defined by the local maxima and the envelope defined by the local minima is zero. NTU, GICE, MD531, DISP Lab 37 Sifting Process (1) First, find all the local maxima extrema of x(t). IMF 1; iteration 0 2 1 0 -1 -2 10 20 30 40 50 60 70 80 90 100 110 120 (2) Interpolate (cubic spline fitting) between all the maxima extrema ending up with some upper envelope emax (t ) . IMF 1; iteration 0 2 1 0 -1 NTU, GICE, MD531, DISP Lab -2 10 20 30 40 50 60 70 38 80 90 100 110 120 (3) Find all the local minima extrema. IMF 1; iteration 0 2 1 0 -1 -2 10 20 30 40 50 60 70 80 90 100 110 120 (4) Interpolate (cubic spline fitting) between all the minima extrema ending up with some lower envelope emin (t ) . IMF 1; iteration 0 2 1 0 -1 -2 10 20 30 40 50 60 70 NTU, GICE, MD531, DISP Lab 80 90 100 110 120 39 (5) Compute the mean envelope between upper envelope and lower envelope. m(t ) emin (t ) emax (t ) 2 IMF 1; iteration 0 2 1 0 -1 -2 10 20 30 40 50 60 70 80 90 100 110 120 (6) Compute the residue h(t ) x(t ) m(t ) residue 1.5 1 0.5 0 -0.5 -1 -1.5 10 20 30 40 50 60 70 NTU, GICE, MD531, DISP Lab 80 90 100 110 120 40 (7) Repeat the above procedure (step (1) ~ step (6)) on the residue until the residue is a monotonic function or constant. The original signal equals the sum of the various IMFs plus the residual trend. n x(t ) ck (t ) rn (t ) k 1 NTU, GICE, MD531, DISP Lab 41 EX: 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1 0 1 2 3 4 5 6 4 x 10 IMF1 0.1 0.05 0 -0.05 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.5 0.6 0.7 IMF2 0.2 0.1 0 -0.1 -0.2 0 0.1 0.2 0.3 0.4 Time NTU, GICE, MD531, DISP Lab 42 IMF3 0.4 0.2 0 -0.2 -0.4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.5 0.6 0.7 0.5 0.6 0.7 0.5 0.6 0.7 IMF4 0.4 0.2 0 -0.2 -0.4 0 0.1 0.2 0.3 0.4 Time IMF5 0.1 0.05 0 -0.05 -0.1 0 0.1 0.2 0.3 0.4 IMF6 0.04 0.02 0 -0.02 -0.04 0 0.1 0.2 0.3 0.4 Time NTU, GICE, MD531, DISP Lab 43 IMF7 0.01 0.005 0 -0.005 -0.01 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.5 0.6 0.7 0.5 0.6 0.7 0.5 0.6 0.7 IMF8 0.01 0.005 0 -0.005 -0.01 0 0.1 0.2 0.3 -3 10 0.4 Time IMF9 x 10 5 0 -5 0 0.1 0.2 0.3 -3 10 0.4 IMF10 x 10 5 0 -5 0 0.1 0.2 0.3 0.4 Time NTU, GICE, MD531, DISP Lab 44 -3 4 IMF11 x 10 2 0 -2 -4 0 0.1 0.2 0.3 -3 0 0.4 0.5 0.6 0.7 0.5 0.6 0.7 IMF12 x 10 -0.5 -1 -1.5 0 0.1 0.2 0.3 0.4 Time NTU, GICE, MD531, DISP Lab 45 Hilbert Spectral Anaysis 1 1 x( ) y (t ) H {x(t )} x(t ) PV d t t z (t ) x(t ) iy(t ) a(t )ei (t ) y (t ) (t ) arg( z (t )) tan ( ) x(t ) 1 f (t ) 1 1 d (t ) (t ) 2 2 dt NTU, GICE, MD531, DISP Lab 46 Advantage Disadvantage STFT 1. Low computation 1. Complex value and 2. The range of the integration is limited 2. Low resolution Gabor transform 3. No cross term 4. Linear operation Real 1. High computation 2. High resolution 2. Cross term 3. If the time/frequency limited, time/frequency of the 3. Wigner function distribution 1. Non-linear operation WDF is limited with the same range Cohen’s 1. Avoid the cross term 1. High computation class 2. Higher clarity 2. Lack of well mathematical properties distribution Gabor-Wigner distribution 1. function Combine the advantage of the WDF and the Gabor 1. High computation transform 2. Higher clarity 3. No cross-term NTU, GICE, MD531, DISP Lab 47 Conclusion We introduce many distributions here and put most attention on computation time and representations. We can find that the representation with higher clarity cost more computation time for all methods. Resolution Computation time The Hilbert-Huang transform is the most power method to deal with non-linear and non-stationary signals but lacks of physical background. NTU, GICE, MD531, DISP Lab 48 Reference [1][A]J. J. Ding, “Time-Frequency Analysis and Wavelet Transform,” National Taiwan University, 2009. [Online].Available: http://djj.ee.ntu.edu.tw/TFW.htm. [2][B]W. F. Wang, “Time-Frequency Analyses and Their Fast Implementation Algorithm,” Master Thesis, National Taiwan University, June, 2009. [3]Luis B. Almeida, Member, IEEE, “The Fractional Fourier Transform and Time-Frequency Representations,” IEEE Transaction On Signal Processing, vol. 42, no. 11, November 1994. [4]M. R. Spiegel, Mathematical Handbook of Formulas and Tables, McGrawHill, 1990. [5]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. NTU, GICE, MD531, DISP Lab 49