Minutes of Detector Characterization Teleconference
(June 2, 2000)
Present:
AEI-Potsdam: Sintes
Caltech: Lazzarini,
Shawhan, Zweizig
Dublin: Ottewill
Florida: Klimenko
Hanford: Raab,
Sigg
Michigan: Riles
MIT: Daw
Oregon: Mauceli, Strom
Penn State: Finn, Mohanty,
Siddiqui
Syracuse: Penn
Data Monitor Tool Status (John Zweizig):
-
John is working on the next release of the DMT (new features described
in last teleconference).
-
Interested users can get a preview of the new release on the Hanford sand
machine in the directory ~dmt/new and can link to it by setting the environment
variable DMTVERSION=new (instead of the normal 'pro' value).
-
The new release represents an extensive change from the January release
(1.1B) and debugging is not quite complete. At the moment, John is improving
the message passing interface and the trigger manager/client communication.
The display manager will not be ready for this release, but a hook for
it is in place.
-
Code contributed by Adrian Ottewill (line tracking) and Teviet Creighton
(IIR filter generation) have not yet been incorporated. The code from Teviet
is an extracted subset of the LAL library with a wrapper. John will complete
the merging of the code into the DMT.
-
John will be at Hanford June 4-9 and generally available at 509-372-8300
(x195).
Status reports on performance characterization priority 1 tasks:
-
Line noise:
-
B. Allen, A. Ottewill (Adrian reporting)
The line tracking code was sent to John in April (see above and minutes
of last teleconference). Adrian has mainly worked recently on inter-channel
correlations (see below).
-
S. Klimenko
Only small changes have been made to line tracking code since the last
teleconference. Sergey has mainly worked on data compression code recently
(see below). He asked how best to allow users to specify lines to be tracked.
It was suggested that he write a config file reader. He will do so and
give his complete code to John soon.
-
A. Sintes
Code for removing 60 Hz lines and harmonics has been converted from
Matlab to LAL C code. Again, it was suggested that a config file reader
be provided (Peter Shawhan requested that config files be placed in a default
directory to allow easy restart of monitoring processes. John is considering
writing a standardized config file interface.)
John has agreed to write a wrapper for the LAL code. Some discussion
ensued on whether it was wise for the DMT to support direct linking to
LAL libraries, as opposed to incorporating source code subsets. John would
like to gain more experience in LAL incorporation before a decision is
made. KR expressed concern that DMT releases not be tied to LAL releases.
John mentioned he had head that the LAL code uses unix signals for error
handling, which would be undesirable in the DMT environment. Sam Finn volunteered
to check this out.
-
Characterizing seismic noise:
-
E. Daw
A seismic monitor has been running continuously at Livingston
since April 19. It monitors (every 16 seconds) the power in two frequency
bands: 0-1 Hz and 1-5 Hz and records to a disk file. Before sending this
code to John, Ed plans to incorporate IIR filters instead of the FIR filters
used now, in order to obtain sharper corners on the filters. He will look
into using Teviet Creighton's code of which he was unaware before this
meeting. He asked whether he should focus on the seismic monitor task or
the general-purpose band-limited rms moniotr task (below). He was encouraged
to give higher priority to the latter, and to gain more experience with
real-data seismic monitoring before releasing a "standard" seismic monitor.
He stated that it would be important to store seismic data permanently
and wondered how best to do so, which provoked a more general discussion
on storing periodically recorded detector characterization information.
Daniel Sigg pointed out that it will soon be possible to generate framed
second- and minute-trend data outside of the DAQ system and suggested that
trend files would be a natural place for regularly recorded performance
characterization information, as opposed to making regular entries in the
meta-database. Peter, who strongly agreed, said that the meta-database
was more natural for hourly updates and, of course, transient triggers.
He pointed out that although trend data is stored in the CACR archive,
there exist no tools yet for accessing it. Such tools do exist now for
trend data stored on site. No objections were raised to Daniel's proposal.
-
Inter-channel correlations:
-
B. Allen, A. Ottewill (Adrian reporting)
Adrian had planned to release his code by today, but a last-minute
check of the code on 1994 40-meter data turned up a segmentation violation
deep in the DMT infrastructure code, which he hasn't tracked down. He and
John will talk offline to sort this out. It may be related to the fact
that the old data's frame time interval wasn't a power of two times the
time sampling bin. Once this technical problem is sorted out, Adrian will
give the code to John.
-
Bilinear cross-couplings
-
S. Penn
Although he hasn't had much time to work on this task in the last month,
Steve still plans to have at least preliminary code ready in a month, based
on the analysis techniques described by Albert Lazzarini at the March LSC
meeting.
-
Operational state:
-
R. Gustafson, K. Riles (KR reporting)
KR had planned to give the complete code to John by today, but didn't
quite make it. The essential code is written, however, and runnable on
the Hanford sand machine (look in ~keithr/osc to find a code snapshot,
sample program and makefile). Remaining work to be done before giving to
John: support of config/log files, cleanup of input/output, automatic booking
of daq channels used in state definitions, and fleshing out of documentation.
This should be ready for John next week.
Status reports on performance characterization priority 2 tasks:
-
Bandlimited RMS
-
Time / Frequency plots
-
Non-Gaussian noise
-
S. Finn, S. Siddiqui
Sharif has installed the DMT at Penn State and is working with Soumya
on conversion of existing power spectral density transient detection code
to the DMT. The same infrastructure can be used to convert code for quantifying
non-Gaussian noise. Sam offered to provide that information in a DMT monitor.
Status reports on transient analysis priority 1 tasks:
-
Power spectral transients:
-
S. Mohanty, S. Siddiqui
As discussed at the last teleconference, this task has been broken
into two stages: 1) providing a general-purpose tool for monitoring &
displaying power deviations from the average in frequency bands vs time
and 2) using pattern recognition to detect and record statistically significant
transients. Sharif has run into a technical problem in creating displayable
images, but was given several suggestions on persons to contact. Once this
is sorted out, the first package will be given to John. It will include
a graphical display to serve as a general-purpose time/frequency plot (see
above).
Fred Raab asked John how long it would take for the code to be incorporated
into the DMT release. A summer student is arriving Monday at Hanford to
start testing out DMT code. It was suggested that Soumya get the existing
code running soon on his account at Hanford so that the student need not
wait for the next DMT release. Sam offered to make this a top priority
of the Penn State group and hoped to have the code available for trial
early next week.
-
Servo instability:
R. Gustafson, K. Riles (KR reporting)
Nothing new to report since the last teleconference. This code should
be ready in early to mid July, perhaps before the next teleconference.
-
Event catalog:
-
J. Sylvestre, R. Weiss (KR reported on behalf of Sylvestre who was
out of the country)
There have been no changes to the core program since the last teleconference,
but some testing has been done with white noise and chirp signals. A summer
student, Olivier D'Arcy, is developing a GUI in root for the monitor which
includes a data viewer and control panel. The viewer could, in principle,
be adapted for use with other DMT monitors. A filter bank to detect overhead
airplanes (see
Julien's analysis of the engineering run data) will be generated
and included in the first release on June 15. Julien will be at Hanford
June 19-23 to install the program and show interested physicists how to
use it.
Status reports on transient analysis priority 2 tasks:
-
Flickering optical modes:
-
Transient detection using adaptive denoising methods
-
E. Chassande-Mottin (no report)
This task has a milestone of October 1.
-
Impulse recognition:
-
Although no one has volunteered for this task, it has much in common with
the matched filtering technique used in the event catalog (see Sylvestre
report above). Determining the appropriate filters to use to recognize
impulses needs work. A volunteer is welcome!
-
Magnetic field transients
-
E. Mauceli
A basic version of the code is now running at Hanford and has
detected the periodic current draw of dust monitors. Since Evan is also
planning to write some priority 3 transient monitors (quakes, wind gusts,
etc), he has decided to revamp substantially the structure of his code
to be more object-oriented, giving more flexibility for defining monitoring
parameters. He hopes to have this complete in two weeks and certainly by
the next teleconference.
Status report on Data Set Reduction: D. Strom
-
David began by giving the floor to Sergey who reported on recent work in
data compression.
-
Sergey has investigated both lossless and lossy compression. His own data
compressor is based on wavelet decomposition, but he has also investigated
other available packages, including gzip, arj and eri for lossless compression
and mpeg3 (designed for audio channels) for lossy compression.
-
As a test, he has tried compressing 10 16-KHz channels from the April engineering
run. He finds a lossless compression factor of 2.05 using simple differentiation
followed by gzip and a factor of 2.40 using a technique he calles Random
Signal Generation (RSC - that exploits the nearly Gaussian behavior of
the signals) followed by wavelet transforms. The second technique is also
5-8 times faster than the first.
-
He has also tested and quantified the energy loss in using mpeg3 and a
dynamic range reduction technique (integer trucation after a scale factor
division, but applied to the wavelet transforms). He finds comparable performance.
For example, both can give roughly five times compression, with about 1%
energy loss.
-
To proceed further, he needs guidance on defining a benchmark for compression
performance, e.g. minimum acceptable energy loss for given compression
factor.
-
A discussion ensued on how to go about defining benchmarks that aren't
arbitrary. In the end, David agreed to propose a small set of specific
benchmark tests (e.g., line removal and regression performance for original
vs compressed data), but it was also decided that the group as a
whole should help out in this effort. In particular, toward the end of
the summer, the data set reduction subgroup will provide a sample of compressed
real data (most likely from the April engineering run). Persons working
on DMT algorithms will then be asked to run them on both the original and
compressed samples and to report back on any differences found.
-
David also reported on other data reduction work. He expects a DMT data
decimation routine to be ready by August 1, compatible with having a standardized
sample of reduced data by summer's end.
Time
Dependence of length control calibration in April engineering run (6/14/00 revision) :
P. Shawhan
-
Peter has tried reproducing offline the calibrations done online every
four hours during the April engineering run, in order to investigate observed
time dependence of the calibration constants.
-
The calibrations were based on several-minute dithers of the input mass
position at 30, 300 and 1000 Hz. He derives a new calibration every
minute from the ratio of length control signal power spectrum bin to the
dither amplitude. (This analysis was done on the Hanford fortress machine
using the Diagnostic Test Tool (DTT).) A correction was made for baseline
noise in the spectrum, using adjacent frequency bins.
-
His goal was to determine whether the long-term calibration changes observed
were due to real long-term drifts or due to intrinsic spread in the calibration,
hence the 1-minute intervals for repeating the calibrations.
-
More fluctuation (few %) was observed on short time scales than expected.
Although he hasn't quantified the statistical error on each measurement,
Peter believes it is much smaller than the spread observed, suggesting
a systematic fluctuation. (see June 14 revision of note above for
supporting evidence)
-
During two calibration periods observed, the arm lost lock and regained
it. In one case, the calibrations showed pronounced drifts just before
the lock loss. In the other case there was no apparent warning.
-
One calibration was done during a period immediately following lock reacquisition
when alignment servo gains are automatically adjusted. The derived calibrations
show large deviations from nominal values, indicating large sensitivity
to the machine's alignment state.
-
As a cross check, the shadow sensor signals from the end mass were checked
for stability. Except for the one calibration during auto-alignment, the
end masses were indeed stable.
Detector Characterization Group's participation in LDAS mock data challenges:
K. Riles
-
A teleconference of eight persons was held a week ago to hash out issues
and put together a rough plan for integrating detector characterization
into upcoming LDAS Mock Data Challenges. (MDC's).
-
It was agreed not to try this integration in the first MDC scheduled
for the end of July to test the data conditioning API. Instead it was proposed
to plan on one or more followup "mini-MDC's", since at least one will be
needed to test the frame-reading API, anyway.
-
Ideally, there would be one such followup in early November to test 1)
frame reading, 2) time interval vetoes, 3) line removal and 4) regression
correction. It should be noted, however, that no one has volunteered to
lead this followup MDC. (KR and Kent Blackburn hope to persuade Sam Finn
to take this on, since he is leading the first MDC.)
-
An important technical note is the agreement reached to derive and use
line removal and regression coefficients in the data conditioning API,
not to derive them in the DMT for downstream use in the API. Since we want
essentially the same code to be running in the DMT, however, for diagnostic
feedback, it was decided and John agreed that the DMT will fully support
the data structures used in the data conditioning API. This will allow
code to be written for both environments with minimal variation. It is
assumed, though, that monitor parameters would be quite different for the
two environments.
-
Adrian, Sergey, and Evan have all volunteered to help in this MDC. The
next teleconference of participants will be June 23.
-
Peter had pointed out at last week's meeting that we are evolving toward
an arrangement where it will be difficult to reproduce detector characterization
algorithms as part of astrophysics analysis, since those analyses must
follow the C-only LAL specification, while both the DMT and data conditioning
code are written in C++.
-
Sam raised the possibility that a hook could be inserted into LAL routines
that allows for a C++ call. There is already a hook for any necessary additional
data conditioning. He will raise this issue with Kent Blackburn.
-
It was pointed out that if detector characterization or data conditioning
had to be repeated / improved offline, there should be no fundamental reason
it couldn't be done on a stand-alone node of an analysis Beowulf system
before being shipped to LAL code for examination.
A.O.B.
-
John mentioned that he has rerun his trigger generation code (see minutes
of last teleconference) on the April engineering run data and that it would
be possible to run other monitor triggers on that data in the near future,
as a test. After some discussion, it was decided to plan on running such
a joint test in mid to late July, after more transient analysis code becomes
available. The existing entries in the meta-database will be wiped to be
superseded by the new set.
-
Next teleconference: Friday July 7 at 11:00 a.m. EDT (8:00 PDT).