GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.)

Data conditioning and veto Data conditioning and veto for TAMA burst analysis for TAMA burst analysis

Masaki Ando and Koji Ishidoshiro (Department of Physics, University of Tokyo)

and the TAMA Collaboration

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 2

Introduction (1)Introduction (1) - Veto analysis - - Veto analysis -

Veto analysis by monitor signals

GW detector is sensitive to external disturbances as well as real GWs Reject fake events using monitor signals recorded together with the main output of the detector Main output

Monitor signals

Data conditioning (Whitening, freq.-band selection, Line removal)

Burst-event extraction (Burst filter: excess-power filter)

Coincidence analysis

GW candidates Fake events

No Yes

Systematic survey of all monitor channelsFake reduction even without 　　　　　 understandings of 　　 noise mechanisms

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 3

Introduction (2)Introduction (2) - TAMA DT9 data - - TAMA DT9 data -

TAMA data acquisition and analysis systemUsed data in this work: TAMA DT9 (Dec. 2003 – Jan. 2004)200 hours of data (HDAQ 8ch, MDAQ 64ch)

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 4

Data ConditioningData Conditioning

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 5

Data conditioning (1)Data conditioning (1) - Quality of data from detectors - - Quality of data from detectors -

Data conditioning

‘ideal’ data (predictable behavior known statistics)

Stationary noiseWhite noise

Real data from detectorsNon stationary behavior Drift of noise level Sudden excitationsFrequency dependence Sensitivity degradation in high and low frequencies 　　 Line noises

0 200 400 600 800 1000 120010

–21

10–20

10–19N

ois

e L

ev

el

[1/H

z1/2 ]

Time [hour]

DT9

DT8DT6

10310–21

10–20

10–19

10–18

No

ise

le

ve

l [1

/Hz1/

2 ]Frequency [Hz]

200 500 2x103

Typical noise levels

Selected frequencies

DT6

DT8

Calibration peak

modepeaks

Violin

modepeaks

Violin

DT9

Data conditioning

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 6

Data conditioning (2)Data conditioning (2) - TAMA conditioning filter - - TAMA conditioning filter -

Data conditioning for TAMA burst analyses

Normalization of the data by averaged nose level Remove time and frequency dependenceLine removalSelect frequency band to be analyzed

In addition … Calibration : Convert v (t) to h (t)

Resampling : Data compression

Requirements Small loss in signal power Keep GW waveform

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 7

Data conditioning (3)Data conditioning (3) - Data flow - - Data flow -

Data conditioning by FFT-IFFT

Frequency selection, Whitening/Calibration, Resampling

Fourier transform 72sec data

Frequency-band selection Overwrite ‘0’ for unnecessary frequencies Outside of the observation band AC line freq. , Violin-mode freq. 　　 Calibration peak

Whitening Normalize by avg. spec.

Calibration Convert to h(f)

Raw time- series data 20kHz samp.

Inverse Fourier transform Only below 5kHz

Whitened time- series data 5kHz samp.

Calibrated time- series data 5kHz samp.

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 8

Data conditioning (4)Data conditioning (4)- Results of conditioning -- Results of conditioning -

Whitened data

50 50.5 51 51.5–15

–10

–5

0

5

10

15

Am

pli

tud

e [

arb

. u

nit

]

Time [sec]

10310–1

100

101

No

rnal

ized

sp

ectr

um

Frequency [Hz]

50.73 50.735 50.74 50.745 50.75

–15

–10

–5

0

5

10

15

Am

pli

tud

e [

arb

. u

nit

]

Time [sec]

Without resamp.5kHz resamp.

Zoom

Spectrum Whitened

Time-series data Waveform is maintained　　　　　　 with resampling

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 9

Data conditioning (5)Data conditioning (5) - Calibrated spectrum - - Calibrated spectrum -

Calibrated data

103

10–21

10–20

10–19

10–18

Str

ain

[1

/Hz1/

2 ]

Frequency [Hz]

Blue: FFT of raw data calib.Red: Conditioned data

Power spectrum of conditioned time-series data and calibrated spectrum of original data

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 10

Event extractionEvent extraction

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 11

Event extraction (1)Event extraction (1) - Excess power filter - - Excess power filter -

Burst event triggerExcess-power filter 　　 Evaluate power in given time-frequency windows Extract non-stationary events

Data conditioning　 time-series data in a given frequency band

Calculate power in a given time window

Power in a given time-frequency window SNR

Threshold, Clustering Burst events

Free selection of time-frequency window

Previous works… Spectrogram Averaged power

Time-frequency resolution were limited by FFT length

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 12

Veto analysisVeto analysis

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 13

Veto analysis (1)Veto analysis (1) - Concept - - Concept -

Coincidence analysisExternal disturbances tend to appear in some monitor signals Reject fake events

Burst events

Coincidence Reject

GW candidate

Threshold

Threshold

Main

ou

tpu

t (S

NR

)M

on

itor

(S

NR

)

Time * Simulated data

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 14

Veto analysis (2)Veto analysis (2) - Parameter optimization - - Parameter optimization -

Parameter optimization

Analysis parameters Time-frequency window ThresholdSignals to be used for veto

Optimized to have… High fake-rejection efficiency Low accidental coincidence

Time-frequency window: Highest coincidence rate selected from 50 (18) time-frequency combinations

Threshold for monitor signals: Accidental ~ 0.1%

Accidental coincidence: estimated by 1-min.time-shifted data

Coin

cid

en

ce r

ati

oAccidentalcoincidence

Coincidence

Playground data (~10%) are used for this optimization 0.1%

SNR Threshold (monitor signal)

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 15

Veto analysis (3)Veto analysis (3) - Signal selection - - Signal selection -

Signal selection

Even with small accidental rejection prob. (~0.1%) for each, many (~100) monitor signals may lose large amount of data

Select signals to be used for veto

Classify by the coincidence rate… <0.5 % Do not use for veto 0.5 - 2 % Use for veto > 2% Intensive veto with lower threshold: accidental ~ 0.5%

Intensive fake-rejection with some monitor signals with strong correlations with the main signal

Re-optimize the threshold

HDAQ: 3 signals Rec. Pow., L+ FB, Dark Pow.

MDAQ: 9 signals l- Err., l- FB, l+ Err., l+ FB, Bright Pow, Rec. Pow., EW Trans. Pow., SEIS center Z, Magnetic Field Y

Selected signals

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 16

Veto analysis (4)Veto analysis (4) - Intensity monitor - - Intensity monitor -

Example of coincident events Intensity monitor in the power recycling cavity (time series data after data conditioning)

Main

ou

tpu

tM

on

itor

sig

nal

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 17

Veto analysis (5)Veto analysis (5) - Seismic motion - - Seismic motion -

Example of coincident events Seismic monitor – center room vertical motion (time series data after data conditioning)

Main

ou

tpu

tM

on

itor

sig

nal

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 18

Veto analysis (6)Veto analysis (6) - Magnetic field - - Magnetic field -

Example of coincident events Magnetic field – center room perpend. direction (time series data after data conditioning)

Main

ou

tpu

tM

on

itor

sig

nal

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 19

Veto analysis (7)Veto analysis (7) - Veto results - - Veto results -

Survival rate

Accidental : 4.2%

MDAQ 9 signals

Accidental : 0.8%

HDAQ 3 signals

Accidental : 4.4%

Total 12 signals

Su

rviv

al ra

te

SNR Threshold

Veto results with 178 hours of data

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 20

SummarySummary

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 21

Summary Systematic survey of monitor signals

Data conditioning filter Excess power burst filter Coincidence analysis between main and monitor signals Optimization of analysis parameters

Analysis with TAMA DT9 data 200 hours data (10% are used as playground data) Correlations were found in 12 monitor signals (Intensity monitor, Seismic motion, Magnetic field)

Reject 92% (SNR>10), 98% (SNR>100) fakes with 4.4% accidental rejection probability

Current tasksHardware signal-injection test Confirm the safety of veto (already done for some of the monitor signals)

SummarySummary

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 22

EndEnd

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 23

10–6 10–5 10–4 10–3

10–22

10–21

10–20

10–19

10–18

No

ise

–le

ve

l d

rift

[

arb

. u

nit

]

Frequency [Hz]

Daily change (1.15x10–5 Hz)

DT9

DT8

Burst-wave analysis (3)Burst-wave analysis (3) - Data conditioning - - Data conditioning -

Data conditioning

Line removal AC line, Violin mode peak, Calibration peak

102 103

10–21

10–20

10–19

10–18

10–17

Str

ain

no

ise

[1/

Hz

1/2]

Frequency [Hz]

Without Line Removal

With Line RemovalMethod FFT 72sec data Reject line freq. components Inverse FFT

Normalization Track the drift of noise level　　　 Each spectrum is normalized by averaged noise spectrum

Use 30min-averaged spectrum

30min(5.6x10-4 Hz)

GWDAW-10 (December 14, 2005, University of Texas at Brownsville, U.S.A.) 24

Time series data (after conditioning)