PDnMon

Introduction

The multiple redundent photodiodes used on the asymmetric port allow a consitency check that might detect:

Interuption of a single beam by dust, mis-steering.
Failure of the one photodiode
Failure of the photodiode filtering, demodulation or readout electronics.
Signal saturation.

Strategy:

Measure consitency between signals from different photodiodes by generating a chi-square quantity.

Filter out uncorrelated low-frequency noise

20HZ high-pass

Calculate a best estimate 'x_mean' of the "true" signal

simple average of the signal of all independent photodiodes over (4s) stride.
rejected weighted average, efificiency modulation correction, etc - just not enoughinformation qabout the "real signal" or relative errors on the individual signals.

Calculate relative gain and offset with a linear fit of each independent signal to x_mean

find the best fit {a_n, b_n} to minimize
Chi-Square = Sum(i=0:Nsample; (Xn[i] - a_n * x_mean[i] - b_n[i])²)
resulting parameters are trended, plotted.
a_n ~1 (± few percent). b_n ~ 0.

Calculate Chi-square terms for each sample

term[i] = w * Sum(n=0:Npd; (Xn[i] - a_n * x_mean[i] - b_n[i])²)
weight (w) calculated to give ChiSqure = NDF when exponential averaged over 100 strides. NDF ~ Npd - 1.
Plots Here.

Look for something unusual

Average values printed out e.g. here
Average Chi-square over stride (4s = 1024 samples) calculated/trended
Number of terms (sample chi-squares) > threshold (usually 16) is counted/trended
Alarm set if scale < 0.9 or bias > 0.5
Triggers in future?

sample plots

The following figure shows the PDnMon output when the AS_PD1 beam was mis-steered off the photodiode. The apparent increse in agin of the remaining 3 photodiodes (between hours 5.6 - 9) is a result of the true signal estimate decreasing by 25% from the loss of PD1. When the PD1 beam was restored, all the measured gains returned to ~1. The average chi-square and number of "glitches" i.e. the number of samples with large chi-square clearly went nuts when the AS1 signal disappeared and the calmed down significantly when it was restored.

figure from LLO elog

Offline Segment Generation

I have begun to look into the generation of DQ segments from PDnMon results. The only data available at present are the trends of PD gains, PD biases, Glitch rates (i.e. number of samples with Chi² > 12) and average Chi² over a 4s stride. After some digging I have determined that the trends are valid starting ~GPS 861150000.

Missing PD Signals

One point that is evident from the data is that there are a two cases where the L1-AS1 photodiode signals were off. They are tabulated below:

IFO-PD	GPS Times	Dates (UTC)	S5 Science Segments	Comments
L1-AS1	863903542-863913094	5/22/2007 21:12 - 5/22/2007 23:50	5335-5338	AS1 beam mis-steered (elog)
L1-AS1	866958072-866982418	6/27/2007 05:40 - 6/27/2007 12:26	5634-5635	Not noted. Possibly not reset after DC Readout tests (elog).

This can be seen as a peak at zero in the histogram of all L1 Scale-factors (maximum per minute), shown below:

The first two columns of the following table contain histograms for each IFO and Phase of the glitch rate (i.e. the number of samples with a point Chi² > 16 in a 4-second stride) and Chi² (averaged over a 4-second stride) maximized over each minute. The colored histogram is for minutes containing a kleineWelle trigger with significance > 80, and the white histogram is for all minutes. The third column contains a Rate vs. Chi² plot, with kleineWelle triggers indicated by the red dots and minutes without kW triggers indicated by a black dot. Data for these plots were pre-selected on a minute-by-minute basis as follows:

Must start after GPS 861150000
Must be completely contained in a science segment
Must not overlap with an AS_TRIGGER, SEVERE_LSC_OVERFLOW, or ASI_CORR_OVERFLOW DQ segment
Must not overlap with the Missing PD signal segments above.

IFO-Phase	Glitch Rate Histogram	Max Chi-Square	Rate v. Chi-Square
H1-ASI
H1-ASQ
H2-ASI
H2-ASQ
L1-ASI
L1-ASQ

Preliminary segments

Although I still believe that the best way to generate segments from PDnMon is to generate triggers in the program itself and to select appropriate triggers for DQ segments after the fact. I believe that this will be more successful at separating the glitches from the background distribution and can be used to veto segments of less than 1 minute. A trigger scheme like this will take a lot more work and require reprocessing of the entire S5 data sample.

Nevertheless, after careful selection of the existing data, I have been able to define some reasonable segments from the trends produced by the on-line processes (summarized in the table above). The following table summarizes the segments and the IFO names link to the segment files:

ifo	Segment definition	Segment Length (s)	kleineWelle Triggers (s)
H1	Chi-I > 15 \|\| Rate-I > 100 \|\| Rate-I/(Chi-I^3) > 0.5 \|\| Chi-Q > 50	13380	10920 (81.6%)
H2	Chi-I > 30 \|\| Chi-Q >30	2880	1620 (56.3%)
L1	Chi-I > 12 \|\| Chi-Q > 30	33240	24480 (73.6%)

Sample event

L1:LSC-AS_I[1-4] plots at 874442650. Raw (top) and averages suppressed (bottom).