Towards a Data and Detector Characterization Robot

1
Towards a Data and Detector Characterization
Robot For Gravitational Wave Detectors Soumya
D. Mohanty, Soma Mukherjee AEI GEO600
collaboration

• Motivation
• As far as their design is concerned,
  Gravitational wave detectors are supposed to
  behave stably and the data produced by them is
  supposed to be stationary.
• Thus, any change in the state of a detector when
  it is operating is an abnormality.
• Here are some examples of what is meant by
  "change"
• Change in the Power Spectral Density of a
  channel.
• Change in the transfer function connecting one
  channel to another.
• Transients.
• Change in the time delay distribution between
  transients of a certain type occurring in
  coincidence in two different channels.
• And more
• Requirements for Interferometric GW detectors
• Large data rate from these detectors (all
  channels) would make it necessary to first
  process the data through an automated alarm
  system (i.e.,a trigger generator) and then follow
  up each trigger with more labour intensive
  studies using short data stretches.
• This automated system will have false alarms and
  false dismissals since one is dealing most of the
  time with noise in all channels (as opposed to
  predictable signals). An uncontrolled false alarm
  rate can overwhelm follow up studies.
• Thus, not only do we need an automated alert
  system, we also want it to be reliable, i.e., it
  must have at least a known false alarm rate (or a
  robust upper limit on false alarms).
• Challenges
• Given that this system must deal with PEM data
  also, which can be highly non-stationary/non-Gauss
  ian or generally difficult to model, achieving a
  known reliability becomes a non-trivial task.
• In addition, the reliability will not be a simple
  function of the performance of individual tools
  alone. As an example, line tracking tools can be
  influenced by the presence of transients and the
  same is true for the converse.
• Thus, the design of the system must account for
  such interdependencies from the start and also be
  robust, where required, against the different
  statistical character of data from different
  channels.
• The flow of data through this algorithm will be
  highly non-linear --- not a straightforward flow
  through a chain of independent algorithms. The
  latter is best described by the term pipeline
  but not the former. To make this distinction, we
  use the term robot instead.

DCR Products
First steps to a DCR A PSD and Transfer
function change point alarm system
Input All data channels
DETECTOR CHARACTERIZATION ROBOT
GW Aux. channel for regression
Aux. channels
GW channel
Start and Stop times of transients
Drop transients
D C R
All channels
Break into segments
Remove lines (e.g. Power supply, mechanical
resonances etc.) This needs to be done using a
model independent method which leaves transients
unaffected.
Drop transients
Start Stop times of GW segments
Find corresponding segments
Drop transients
Start Stop times of GW segments
Start Stop times of Aux. segments
Estimate Statistics for more sensitve line
removal methods such as Kalman filter, etc.
Break into PSDwise homogenous segments
Search for transients This needs to be done with
a method having a known false alarm rate
Estimate Xfer functions
Store statistics
Start Stop times of transients Time freq. info
etc.
Start Stop times of transients
Start Stop times of segments
Start Stop times of transients
Main Output Alarms Change points in data
Store Xfer functions
1
4
3
2
D A T A B A S E
Subsidiary output Summary information (e.g.,
trends).
Subsidiary output updated statistics for data
conditioning algorithms such as Kalman Filtering
GW channel
Start Stop times of transients in GW Aux.
channels as Vetoes
Rate of change of PSD/Trends in PSD.

• Implementation of the DCR
• Develop a Digital Signal Processing library in
  C.
• Make it portable by using data structures from
  Standard Template Library (STL).
• Library will aid rapid prototyping and testing of
  DCR designs.

Start Stop times of segments
Statistics
DATA ANALYSIS PIPELINE
Monitor for Requesting the recalibration of the
detector
Xfer functions
A Preview of some results MBLT (Median Based Line
Tracker) A model Independent, time domain method
for line removal
Information for Detector Diagnostics Classes of
transients, Rates of transients, Cross-Channel
Correlations etc. etc.
Break into PSDwise homogenous segments Remove
Power lines/Violin modes/other lines
List of Candidate Events with Confidence level
for each event
6
5
Caltech 40m 1994 run data PSD of IFO_DMRO
channel before (top) and after (bottom) line
removal using MBLT
Caltech 40m 1994 run data PSD of IFO_Seis_1
channel before (top) and after (bottom) line
removal using MBLT
Spectrogram of Seismic channel before (top) and
after (bottom) line removal. Notice that
transients were not affected substantially.
Using Time Series Modeling to track
non-stationarity Given a stochastic time series,
one can model it (up to second order moments) as
being the result of white noise filtered via a
zero-pole filter. The zeros, poles and Gain of
the filter can be estimated using the data
provided. Once the filter parameters have been
obtained, an estimate of the power spectral
density (in the ensemble sense) can be obtained
from the modulus square of the filter transfer
function itself. This way of obtaining the
spectral estimate is called parametric as opposed
to the non-parametric methods (such as Welchs
averaged periodogram method) that are more
familiar. The advantage of making a parametric
fit to the data is that if the data does indeed
follow the general model outline above, the
signal to noise ratio of the estimated Power
Spectral Density is much better. However, in the
case of interferometric data, the presence of
line features prevents the full exploitation of
the power of time series analysis. We can remedy
that by using MBLT.
Spectrogram of Seismic channel (post line
removal) Raw (top) and AR(100) model (bottom).
Note the better (at least visual) SNR of the
bottom plot
Each plot shows the ratio of PSD (IFO_DMRO
channel) obtained from Welchs method to that
obtained from an AR(200) model (top) without
line removal, (bottom) with line removal.
Caltech 40m 1994 run data Auto-correlation
(un-normalised) of IFO_DMRO channel before (top)
and after (bottom) line removal using MBLT. A
slowly decaying autocorrelation requires much
higher order models.