New data analysis for AURIGA Lucio Baggio Italy, INFN and - - PowerPoint PPT Presentation

New data analysis for AURIGA

Lucio Baggio Italy, INFN and University of Trento

AURIGA AURIGA

The (new) AURIGA data analysis

Since 2001 the AURIGA data analysis for burst search have been rewritten from scratch (G. Vedovato), in parallel with major upgrades taking place on the detector The main goals and specifications to achieve were:

• Be flexible and modular, with easy adaptation to new algorithms
• Adopt the VIRGO/LIGO frame format for data storage and exchange

new data acquisition system

• pen source project, C++

widespread use of supported and well known libraries:

ROOT (http://root.cern.ch) VEGA (http://wwwlapp.in2p3.fr/virgo/vega) FrameLibs (http://wwwlapp.in2p3.fr/virgo/FrameL) FFTW (http://www.fftw.org) LAL (http://www.lsc-group.phys.uwm.edu/lal) MKFilter

(http://www-users.cs.york.ac.uk/~fisher/mkfilter)

• And, develop new algorithms, indeed! (highlight: Karhunen-Loeve decomposition)
• Recycle software (and be recyclable)

Overview (see poster)

The data analysis of raw or simulated data for burst search divides in a series of tasks

DQ MTC EVT FW DAQS DAQ FME

1. Estimate parameters of the analytic part of the noise model (Full Model Estimate, FME) 2. Remove noise correlation (Full Whitening, FW) 3. Perform a matched template filtering and event search (EVT) 4. Define epoch vetoes based on Gaussianity monitors (Data Quality, DQ) 5. Compute distribution of errors in event parameters estimators (Monte Carlo, MTC)

Event search (1)

Within this task the whitened data are optimally filtered in the frequency domain for a specified template signal. Then, the time series is passed to the event search algorithm

event search

• ptimal filter

& coarse interpolation

EVT

max-hold fine interpolation raw data noise model template bank

see poster

Event search (2)

• The time series is downsampled to a convenient sampling rate
• The absolute value of the downsampled time series is searched for the local maxima (max-

hold algorithm with a given dead time), and when it is above a proper threshold a candidate event trigger is issued

• For each event trigger, the exact time of arrival and amplitude are computed after fine

interpolation of the samples, along with sum of squared residuals (for χ2-test), Karhunen- Loeve components, etc time

Event statistics from Monte Carlo (MTC)

MTC

coarse interpolation

Phase 1 Phase 2

template bank

event search

template injection whitened data

The goal of this task is to estimate numerically the distributions of time of arrival and amplitude errors, for a bank of filter templates, possibly not exactly matched with the input signal. Software signal injection takes place in the time domain, by adding a chosen template (properly rescaled in amplitude and time-shifted) to the actually measured white noise of the system. Template injection and search is automatically cycled for specified time and amplitude increments, and can be repeated for indipendently specified signal and filter templates.

see poster

average power

≠ fδ·fδ = Σk σk

• 4 hk

2 + …

Karhunen-Loeve Decomposition (1)

signal h +noise (Sh)

• ptimally filtered

amplitude Wiener filter

with template

δ-filtered

templateless

KLD suboptimal energy estimate

δ-filter: F(ω) = Sh(ω)-1 ↔ R-1 (inverse autocorrelation matrix) (σf

• 2 = 1)

Karhunen-Loeve eigenfunctions {ψk}k=1,…,N Rψk = σk

2 ψk

(Σkσk

• 2 = 1)

Define: AKL

2 = fδ·Rfδ = Σk σk

• 2 hk

2 + Σk σk

• 2nk

2 + 2 Σk σk

• 2 hknk

without signal: AKL ~ Chi(N) with signal: AKL= (Σkσk

• 2hk

2)1/2 + Σk σk

• 2hknk (Σiσj
• 2hj

2)-1/2 + O2(n/AKL)

∫

π ω ω ω 2 ) ( ) ( | |

2

d S H

SNRh= Gauss(0,1) input: Σkhkψk + Σknkψk nk ~Gauss(0,σk) signal noise fδ = R-1 h = Σk σk

• 2 hkψk + Σkσk
• 2nk ψk

http://www.ligo.caltech.edu/docs/P/P010019-01.pdf

Karhunen-Loeve Decomposition (2)

SNRδ probability density amplitude SNRh linear filter with mismatched template Karhunen-Loeve decomposition

• Pros:

The signal-to-noise ratio through KLD equals the maximum one achievable with template knowledge

• Cons:

increased tail of fake events definition of event baricenter?

Summary

• Brand new code, rewritten from scratch in C++, running on standalone PCs
• Integrated ARMA noise simulator, generating stationary or time varying correlated

gaussian noise, possibly polluted with power line harmonics, periodic signals and bursts.

• Adaptive parametric noise model estimate
• Support for non-parametric frequency-dependent calibration function
• Support for template bank search.
• Embedded Monte Carlo and tools for measuring efficency.

To do:

• Re-implementation of data conditioning, study for optimization of the (still) empirical

vetoing rules. Tuning on forthcoming sensitivity and stability of the detector.

• Make the analysis more robust with respect to heavy data corruption by spectral

lines and transient disturbances.

• Training on templateless search, tuning of time interval size for K-L decomposition

comparison with time-frequency methods

• Intensify collaboration with other research groups, in order to share algorithms

see also poster

Pararametric noise model estimator

raw data SDFT

FME

Phase 1 Phase 2

Periodograms quality check Iterative fit and data conditioning

FFT1 FFT2 FFT3 FFTn

Time series smoothing

Outlier removal

• 13

• 12

• 11

• 10

• 9

• 8

Phase 1

Periodograms quality check Iterative fit and data conditioning FFT1 FFT2 FFT3 FFTn

Hz 800 900 1000 1100 1200 1300 1400 1500 1600 1700 Volt^2/Hz 10

• 13

10

• 12

10

• 11

10

• 10

10

• 9

10

• 8

FME Phase 1 avr : 64x26.8 Wed Apr 2 19:18:48 2003