1
Electrical Engineering Department 2
Signal Theory Group
University of Montenegro Ruhr University Bochum
Serbia & Montenegro Germany
E-mail: {igordj,ljubisa}@cg.ac.yu E-mail: {url,boehme}@sth.rub.de
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
GOAL
• Nonparametric estimation of multicomponent signal parameters (estimation of parameters of
components from the mixture) is considered.
• Estimation tools are the time-frequency representations combined with some basic pattern
recognition methods.
PAPER OUTLINE
COMBUSTION KNOCK
• Combustion signal can be modeled as a sum of lowpass signal with several high pass
components.
• In this paper we propose technique for combustion signal parameters estimation based on the
time-frequency representations with the following advantages:
estimation of parameters is performed using single observation;
problem of recognition of close signal components is solved.
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
SIGNAL MODEL
Multicomponent FM signal:
Assumptions:
Ai(t)=0 for t<ti |dAi(t)|<<|dφi(t)|
dφj(t)/dt≠dφj(t)/dt for i≠j
P P
x(t ) = ∑ xi (t ) + ν(t ) = ∑ Ai (t )exp( jφ(t )) + ν(t )
i =1 i =1
S-METHOD AS TF REPRESENTATION
L
SM (t , ω) =| STFT (t , ω) | +2 Re ∑ STFT (t , ω + l ∆ω) STFT * (t , ω − l ∆ω)
2
l =1
time and frequency sampling
Short-time Fourier transform
N / 2 −1
STFT (t , ω) = STFT ( n∆t , k ∆ω) = ∑
m =− N / 2
x((n + m) ∆t ) w(m∆t ) exp(− jkm∆ω∆t )
window function
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
In the region of interest the TFD is highly concentrated comparing to the points outside the
signal region.
Then we decided that as a region-of-interest indication function adopt ratio between signal
energy in narrow region around the considered TF point and in wider region around
considered TF point.
∑ TF (t , ω)
Atω
x
R (t , ω) = 1 if ≥ p ∧ TF (t , ω) ≥ max[qm (t ), r γ ]
∑ TF (t , ω)
x x x
x
Bt ω
We selected:
Atω=[t-TA∆t,t+TA∆t]×[ω-ΩA∆ω, ω+ΩA∆ω] (2TA+1)×(2ΩA+1) size
Btω=[t-TB∆t,t+TB∆t]×[ω-ΩB∆ω, ω+ΩB∆ω] (2TB+1)×(2ΩB+1) size
• Assume that the TF components are well separated (for more than (2ΩB+1) samples)
and that the TF representation of the component is concentrated in a single frequency
point within the considered time-interval.
• Then ratio for TF point belonging to the signal components the following
approximation holds:
∑At ω
TFx (t , ω)
2T + 1
≈ A
∑ TFx (t , ω) 2TB + 1 Bt ω
∑ TF (t , ω) (2T
At ω
x
+ 1)(2Ω A + 1)
≈ A
∑ TF (t , ω) (2T
Bt ω
x B + 1)(2Ω B + 1)
Since the combustion knocks parameters are relatively slow varying compare to the
sampling rate we selected that TA=TB=20 (in a wide range of these parameters selection
doesn't influence algorithm accuracy).
In order to separate close signal components ΩA and ΩB should be selected as small as
possible. In our simulations it is selected ΩA=1 ad ΩB=3:
3
7 < p < 1 we selected p=1/2 (close to lower bound) since measurements was corrupted
with very moderate noise amount and we want to keep weak signal components.
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
Additional criterions are introduced to remove TF points influenced by noise that satisfy
the first criterion. Since we want to keep weak signal components we adopted very small
values of q and r (q=1.5% and r=0.1% in simulations but very similar results are obtained
for any q<2% and r<1%).
• The next step in the algorithm is to separate components from the binary image;
• All points R(τ,θ)=1 such that there is path (τ,θ) connecting (t0,ω0) that passing
through points that belong to the binary image are assigned to the specified
component (fire phase);
• The search for new grass point is performs. Algorithms stops when there is no more
grass points.
Assume that regions are ordered in such a way that Ri(t,ω) contains higher energy than
Ri+1(t,ω):
∑
( t ,ω)∈Ri ( t ,ω )
TFx (t , ω) > ∑
( t , ω)∈Ri +1 ( t ,ω )
TFx (t , ω)
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
PARAMETER ESTIMATION
1. IF ESTIMATION
∑ ωTF (t , ω) R (t , ω)
x i
ω
ˆ i (t ) = arg max TFx (t , ω) Ri (t , ω) or ω
ˆ (t ) = ω
∑ TF (t , ω) R (t , ω)
i
ω
x i
ω
tie
φˆ i (t ) = ∑ ω
ˆ i (t ) ∆t component start
t = tib
i =1
Signal parameters
u(t)=1 for t>0 and 0 elsewhere a=32π t0=0.1
Ai=0.5+i0.1 αi=0.6+0.3i bi=(6-i)32π
i=1,...,5 t ∈ [0,1] ∆t=1/256
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
TFD parameters
Hanning window with N=256 is used
L=1 in the S-method realization
Algorithm parameters
p=0.5 q=1.5% r=0.1%
components with less than 1% of signal energy are removed
(used also for real data)
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
|STFT(t,ω)|2 SM(t,ω)
4
x(t)
2
-2
-4
t
-6
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Knock pressure signal recorded on a VW Passat engine (four cylinder engine) at 1200 rpm.
Signal is high pass filtered with cut-off frequency 3000 Hz.
Sampling rate ∆t=10-2 ms.
|STFT(t,ω)|2 SM(t,ω)
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
R(t,ω)
4.5
3.5
2.5
1.5
0.5
t
0
0 0.002 0.004 0.006 0.008 0.01 0.012
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
500
red line reconstructed signal
blue line original signal
0
-500
t[s]
-1000
0 0.002 0.004 0.006 0.008 0.01 0.012
1000 300
200
500
100
0 0
-100
-500
-200
t[s] t[s]
-1000 -300
2 2.2 2.4 2.6 2.8 3 4 4.2 4.4 4.6 4.8 5
-3 -3
x 10 x 10
EUSIPCO 2004 Instantaneous Frequency and Nonstationary Spectral Estimation
Igor Djurović, Mark Urlaub, LJubiša Stanković, Johann F. Böhme
"Estimation of multicomponent signals by using time-frequency representations with application to knock signal analysis"
NUMERICAL STUDY
The algorithm is tested on 50 pressure and vibration signals recorded on 1200, 1750, 2000,
3000 and 3500 rpm.
|STFT(t,ω)|2
SM(t,ω)
Noise is more emphasized than
in previous case (and numerous
other records). It can be seen
that noise influenced
components are removed. Then
"small "normalized covariance
is in this case desired.
CONLUSION
• Algorithm can be applied (with reviewed setup) for some other real-life signals.