search for bursts with the frequency domain adaptive filter ( fdaf )
DESCRIPTION
Search for bursts with the Frequency Domain Adaptive Filter ( FDAF ). Sabrina D’Antonio Roma II Tor Vergata Sergio Frasca, Pia Astone Roma 1. Outlines: FDAF description Project1a data application Filters performances comparison WSR7 seg. 27 data application. - PowerPoint PPT PresentationTRANSCRIPT
Search for bursts with the Search for bursts with the Frequency Domain Adaptive Filter (Frequency Domain Adaptive Filter (FDAFFDAF ) )
Sabrina D’Antonio Roma II Tor Vergata Sergio Frasca, Pia Astone Roma 1
Outlines:Outlines:
FDAF descriptionFDAF description
Project1a data application Project1a data application
Filters performances comparisonFilters performances comparison
WSR7 seg. 27 data applicationWSR7 seg. 27 data application
Overview on the filter and cluster generation procedure
Three STEPS:1. Filtering procedure:• An Adaptive Wiener Filter (AWF), in frequency
domain, followed by a series (N) of band-pass filters with a Gaussian shape (phase zero) ->(N+1) filtered output
2. Event extraction• An Adaptive threshold algorithm for the selection of
the events applied at each filtered channel CH (N+1).
3. Cluster generation
Events coming from different CH in coincidence in a given time window W are put together: this is one CLUSTER
RAW DATA
Wiener Filter
WF
Power Spectrum SP estimation
Filters bank (N)
With
Gaussian
shape
FFT
1
2 ... ..
4 ... ..
.. … ..
. … ..
N+1
N+1
filtered
output
channels
In the time domain
Filtering procedure
IFFT
IFFT
Hp
Frequency domain
back in
time-domain
Hp
Hp
Power spectra estimation
PS is the estimated Power Spectrum of the noise, evaluated with a first order Auto-Regressive (AR) sum of the periodograms, Pi : from
PSi’ = W ∙ PS’i-1 + Pi
and ni = 1 + W ∙ ni-1
PS=PS’/n in our case:
W=exp(-T/tau)=0.9991
T=3.2768 s (time duration of one data chunk used to obtain the Periodogram) tau=3600 s (Memory time: Simulated data-> stationary noise)
Event extraction: adaptive threshold technique for the events selection (event search
procedure applied at each filtered channel y(i)) Let y(i) the filtered data samples in time domain, we estimate
mi = yi + W∙mi-1 qi = y²i + W∙qi-1
ni = 1 + W∙ni-1
with W = exp(-dt/tau) = 0.9999900 (corresponding to dt= 1/20000 s & tau = 5 s) and Mi = mi /n I
Qi = qi/ni
Si = sqrt(Qi-(Mi)2)
From these we define the Critical Ratio (CRi) of y(i)
CRi=|(yi-Mi)/Si|
Event extractionWe define a “dead time’’, td, as the minimum time between two events, and
we put the threshold, ϑ, on the CR. A two-state ( 0 and 1 ) mechanism “event machine’’ has been used: 1. The “machine’’ starts with state 02. When CR > ϑ, it changes to state 1 and an event begins
3. The state changes to 0 after CR remains below ϑ for a time > td (the event finishes)
T0 = starting time CRmax = Max value of CR
A = amplitudeThe ‘event’ is characterized Tmax = time of max CR
by: L = length (in seconds) (duration of state 1) CH = frequency channel
ϑ=3.9 td=0.2 s
Event cluster
EVENT LIST
of all frequency Channel (N+1)
Ch1 Time CR … … …
Ch4 Time CR … … …
Ch7 Time CR … … …
………………
…………………
… … ………………………….
Ch1 Time CR
All Events coming from different frequency channel Ch in coincidences into a given time window W are put together: this is one CLUSTER.
The time corresponding to the higher CR is the CLUSTER time .
CLUSTER list
Time CR1 CR2 CR3 …. CRN+1
Time CR1 … … … CRN+1
………………….
………………….
Time CR1 … … … CRN+1
CRi=0 if the frequency channel Chi not in time coincidence with other.
Event cluster example (Preliminary Procedure!)
Frequency channels: Mean values of the Gaussian filters
CR value
WF channel
0-2000 Hz
Event list:
Freq. Channel Time cr Ampl length ..
1 (40 Hz) tim1 6.13 … … .. .. … .. ..
2 (90Hz) tim2 6.75 … … .. .. … ..
3 (200Hz) tim3 6.0 … … .. .. … .. ..
10(0-2000Hz) tim10 4.8
6 …
… Time distances tim10-tim1 < W=10ms
They are put together-> one CLUSTER
Cluster list
Time CR1 CR2 CR3 CR4 .. .. .. .. CR10
Time2 6.13 6.75 6.0 0 0 0 0 0 4.8
clust
er
ord
eri
ng n
um
ber
CR
40 90 200 600 1000 1400 0-2000 Hz
tim2 is the time corresponding to the maximum CR ->time2=CLUSTER time
Project1a preliminary results
gr-qc/0701026 A comparison of methods for gravitational wave burst searches
from LIGO and Virgo
3 hours of Virgo simulated noise
Injected signalsINPUT:
3 hours of Virgo (vs=20 kHz) simulated noise
Signals injected with SNR=7, 10
Gaussian signals with σ = 1ms
2 kinds of supernovae signals (from Dimmelmeier-Font-Muller simulations) @ 8.5 kpc: A1B2G1,A2B4G1)
Sine-Gaussian signals with Q = 5 and ν = 235 Hz or ν = 820 HzSine-Gaussian signals with Q = 15 and ν = 820 Hz
Wiener filter (WF) +Band-Pass filters with Gaussian shape:The frequency range 0-2000 Hz is linearly divided into 9 bands (step = 200 Hz, Sigma=100 Hz) .--> 10 different filters
Waveform families of burst sources used in this study: time domain
Waveform families of burst sources used in this study: frequency domain
SGQ15f820: clusters in time coincidences with the injected signals (163) at SNR=7 (frequency domain
characteristic)
Due to the noise!
Not in the expected channel and they don’t change with the SNR of injected signals
SNR=7: CRSNR=7: number of event detected from each channel
clust
er
ord
eri
ng n
um
ber
Event
num
ber
90 200 600 1000 1400 0-2000 Hz
90 200 600 1000 1400 0-2000 Hz
SGQ15f820: clusters in time coincidences with the injected signals (163) at SNR=10 (frequency domain
characteristic)
Due to the noise!
Not in the expected channel and they don’t change with the SNR of injected signals
N
SNR=10: CR
clust
er
ord
eri
ng n
um
ber
90 200 600 1000 1400 0-2000 Hz
SNR=10: number of event detected from each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
SGQ5f820: clusterin time coincidences with the injected signals (178) at SNR=7 (frequency domain
characteristic) SNR=7: CR
clust
er
ord
eri
ng n
um
ber
90 200 600 1000 1400 0-2000 Hz
SNR=7: number of event detected from each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
SGQ5f820: cluster in time coincidences with the injected signals (178) at SNR=10 (frequency domain
characteristic) SNR=10: CR
clust
er
ord
eri
ng n
um
ber
90 200 600 1000 1400 0-2000 Hz
SNR=10: number of event detected from each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
SGQ5f235: clusters in time coincidences with the injected signals (190) at SNR=7 (frequency domain
characteristic) SNR=7: CR
clust
er
ord
eri
ng n
um
ber
90 200 600 1000 1400 0-2000 Hz
SNR=7: number of event detected from each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
SGQ5f235: clusters in time coincidences with the injected signals (190) at SNR=10 (frequency
domain characteristic) SNR=10: CR
clust
er
ord
eri
ng n
um
ber
90 200 600 1000 1400 0-2000 Hz
SNR=10: number of event detected from each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
A1B2G1: clusters in time coincidences with the injected signals (165) at SNR=7 (frequency domain
characteristic) SNR=7: CR
clust
er
ord
eri
ng n
um
ber
90 200 600 1000 1400 0-2000 Hz
SNR=7: number of event detected from each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
A1B2G1: clusters in time coincidences with the injected signals (165) at SNR=10 (frequency domain
characteristic) SNR=10: CR
clust
er
ord
eri
ng n
um
ber
90 200 600 1000 1400 0-2000 Hz
SNR=10: number of event detected from each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
A2B4G1: clusters in time coincidences with the injected signals (170) at SNR=7 (frequency domain
characteristic) SNR=7: CR
clust
er
ord
eri
ng n
um
ber
90 200 600 1000 1400 0-2000 Hz
SNR=7: number of event detected from each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
A2B4G1: clusters in time coincidences with the injected signals (170) at SNR=10 (frequency domain
characteristic) SNR=10: CR
clust
er
ord
eri
ng n
um
ber
90 200 600 1000 1400 0-2000 Hz
SNR=10: number of event detected from each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
To see better the lower frequency region (A2B4G1 & GAUSS1ms)I’ve added another channel at 40 Hz
A2B4G1: clusters in time coincidences with the injected signals (170) at SNR=7 (frequency domain
characteristic) SNR=7: CR
clust
er
ord
eri
ng n
um
ber
40 90 200 600 1000 0-2000 Hz
SNR=7: number of event detected from each channel
Event
nu
mber
40 90 200 600 1000 0-2000 Hz
A2B4G1: clusters in time coincidences with the injected signals (170) at SNR=10 (frequency domain
characteristic) SNR=10: CR
clust
er
ord
eri
ng n
um
ber
SNR=10: number of event detected from each channel
Event
nu
mber
40 90 200 600 1000 0-2000 Hz 40 90 200 600 1000 0-2000 Hz
GAU1ms: clusters in time coincidences with the injected signals (178) at SNR=7 (frequency domain
characteristic) SNR=7: CR
clust
er
ord
eri
ng n
um
ber
40 90 200 600 1000 0-2000 Hz
SNR=7: number of event detected from each channel
Event
nu
mber
40 90 200 600 1000 0-2000 Hz
GAU1ms: clusters in time coincidences with the injected signals (178) at SNR=10 (frequency domain
characteristic) SNR=10: CR
clust
er
ord
eri
ng n
um
ber
40 90 200 600 1000 0-2000 Hz
SNR=10: number of event detected from each channel
Event
nu
mber
40 90 200 600 1000 0-2000 Hz
Trigger due to the noise (no signal injection!)
NOISE: CR
clust
er
ord
eri
ng n
um
ber
NOISE: number of event in each channel
Event
nu
mber
90 200 600 1000 1400 0-2000 Hz
90 200 600 1000 1400 0-2000 Hz
Signal SNR <CR> Std(CR) bias[ms]
Std(DT)[ms]
Eff %
SGQ15f820 10 8.98@(4.2)
1.0@(0.46)
-0.031*35.6%
0.55 100
7 6.42(4.2)
0.94(0.3)
-0.06*20.86%
0.8 100
A1B2G1 10 8.28(4.2)
1.0(0.4)
-0.008*43%
0.05 100
7 5.92(4.2)
0.9(0.3)
-0.0024*32.5%
0.1 98.2
SGQ5f820 10 8.3(4.2)
1.0(0.4)
-0.008*55%
0.26 100
7 5.99(4.2)
0.9(0.3)
0.0*34.3%
0.39 97.8
*: percentage of CLUSTERS detected at the exact sample (DT=0.0)
@: obtained over all CLUSTERS (due to the noise+ due to the signals)
Signal SNR <CR> Std(CR) bias[ms]
Std(DT)[ms]
Eff %
SGQ5f235 10 8.24(4.2)
1.0(0.4)
0.1*24.21%
0.8 10099.4
7 5.99(4.2)
1.0(0.3)
0.006*15.22%
1.2 96.84
A2B4G1 10 6.47.47(4.2)
0.90.9(0.4)
0.8 1.0
0.960.28
100100
7 4.865.36(4.2)
0.640.8(0.3)
0.85 1.0
1.50.9
78.897.05
GAUSS1ms 10 8.9(4.2)
1.0(0.5)
-0.007*18%
0.11 100
7 6.3(4.2)
0.9(0.3)
-0.015*11.3%
0.17 99.4
*: percentage of CLUSTERS detected at the exact sample (DT=0.0)
The red values are obtained adding the lower frequency channel at 40 Hz
Filters performances comparison
Efficiency vs False Alarm Rate SNR=7 (Comparison with Power filter (Red))
sgQ15f820
A1B2G1
GAU1ms
A2B4G1
sgQ5f235
sgQ5f820
Signals injected with SNR=10
give efficiency=1 with FAR=10-4
Efficiency vs False Alarm Rate SNR=7
Time error comparison std[ms]
Std[ms] QT PF KW PC EGC MF ALF AFDFA1B2G1 0.2 0.05 0.5 0.04 0.03 0.05 0.3 0.05A2B4G1 1.4 0.7 2.6 0.2 0.4 0.5 2.2 0.28GAUSS1 0.8 1.2 1.4 0.1 0.2 0.8 1.5 0.11SG235Q5 0.9 1.4 0.9 1.3 0.5 1.3 1.1 0.8SG820Q5 0.2 0.2 0.3 - 0.2 0.3 0.3 0.26SG820Q15 0.7 0.6 1.1 - 0.5 1.3 1.1 0.55
Time error comparison bias [ms]
bias[ms] QT PF KW PC EGC MF ALF AFDF
A1B2G1 -0.1 0.2 0.3 -0.05 -0.06 0.03 0.1 -0.008A2B4G1 -1.7 1.9 4.9 -1.3 -1.4 2.1 0.7 1.0GAUSS1 -0.05 1.7 3.4 -0.01 0.01 2.3 1.7 -0.007SG235Q5 -0.07 0.6 1.3 0.02 0.02 0.7 0.7 0.1SG820Q5 -0.01 0.2 0.2 - 0.01 0.1 0.1 -0.008SG820Q15 -0.04 0.2 0.3 - 0.03 0.3 0.2 -0.031
WSR7 Preliminary Results seg.27GPS time start=852852866
GPS time stop =852858889
Hardware Injections: (SNR=7.5,15,25)
Injected signals N
SGf1000Q5/Q15 34/34
SGf1300Q5/Q15 34/34
SGf1600Q5/Q15 34/33 = 271 inj.
GAUSSIAN 34/34
A2B4G1 34/34
Pre HP filter with freq. cutoff at 80 Hz
Power Spectra Estimation:
tau=1800 s
T=3.2768 s
CR: ϑ=4.0
Wiener filter (WF) +Band-Pass filters with Gaussian shape:The frequency range 0-2000 Hz is linearly divided into 10 bands (step = 150 Hz, Sigma=100 Hz) .--> 11 different filters
GAUSSIAN/A2B4G1: all signals detected
150 550 800 1150 1450 0-2000 Hz 150 550 800 1150 1450 0-2000 Hz
150 550 800 1150 1450 0-2000 Hz 150 550 800 1150 1450 0-2000 Hz
SGf1000Q15/Q5: all signals detected
150 550 800 1150 1450 0-2000 Hz
150 550 800 1150 1450 0-2000 Hz 150 550 800 1150 1450 0-2000 Hz
150 550 800 1150 1450 0-2000 Hz
SGf1300Q15/Q5: all signals detected
150 550 800 1150 1450 0-2000 Hz
150 550 800 1150 1450 0-2000 Hz
150 550 800 1150 1450 0-2000 Hz
150 550 800 1150 1450 0-2000 Hz
SGf1600Q15/Q5: all events detected
150 550 800 1150 1450 0-2000 Hz
150 550 800 1150 1450 0-2000 Hz
150 550 800 1150 1450 0-2000 Hz
150 550 800 1150 1450 0-2000 Hz
Signal <CR/SNR>
Std(CR/SNR)
Bias[ms]<Ts-Tc>
Std(DT)[ms]
Eff %
SGf1000Q15 0.99 0.11 -0.67 0.38 100
SGf1000Q5 0.97 0.11 -0.54 0.22 100
SGf1300Q15 1.15 0.11 -0.89 0.41 100
SGf1300Q5 0.9 0.10 -0.7 0.39 100
SGf1600Q15 1.22 0.10 -0.5 0.38 100
SGf1600Q5 0.9 0.12 -0.54 0.38 100
A2B4G1 0.7 0.1 -2.3 0.8 100
GAUSSIAN 0.92 0.088 -0.45 0.1 100
CR: clusters in time coincidence with the injected signals
all clusters-271 clusters in time coincidence with the injected signals
<CR>=15.22
Std(CR)=7.27
<CR>=4.44
Std(CR)=0.61
4 BIG events not due to the injected signals
(first injection time=852852651.45370)
Without 4 big events
Signal Dt=1/20000 [s]
<CR/SNR> Std(CR/SNR) Bias[ms]<Ts-Tc>
Std(DT)[ms]
Eff %
SGf1000Q15 0.996 0.116 -0.56 0.31 100
SGf1000Q5 0.97 0.11 -0.57 0.15 100
SGf1300Q15 1.15 0.11 -0.97 0.32 100
SGf1300Q5 0.9 0.10 -0.62 0.15 100
SGf1600Q15 1.22 0.10 -0.55 0.15 100
SGf1600Q5 0.9 0.12 -0.54 0.17 100
A2B4G1 0.7 0.1 -2.3 0.8 100
GAUSSIAN 0.92 0.09 -0.45 0.0065 100