IREU Final Report Characterization of the Omicron Trigger ... · Characterization of the Omicron Trigger Generator and Transient Analysis of aLIGO Data Hunter Gabbard Virgo Detector

IREU Final Report • August 2014

Characterization of the OmicronTrigger Generator and Transient

Analysis of aLIGO Data

Hunter Gabbard∗

Virgo Detector Characterization Group, Laboratoire de l’accelerateur lineaireCentre scientifique d’Orsay, Batiment 200, Orsay, France

Department of Physics and Astronomy, University of MississippiUniversity, Mississippi 38677-1848 USA

Department of Physics, University of FloridaGainesville, FL 32611-8440 USA

[email protected]

August 12, 2014

Abstract

Omicron is a burst-type trigger generator. We performed coincidence tests between Omicron generatedtriggers of ER5 LIGO data and various types of injection waveforms (Sine-Gaussian, White-Noise-Burst,and String Cusp) using a Coincidence Finder program that we developed. Through these tests wedetermined the efficiency at which the Omicron trigger generator is able to detect specific transient eventswith varrying sets of parameters. We tested and debugged a new version of Omicron and utlized Omicronto perform a close analysis of lock time data at the gravitational wave detector in Livingston, Louisiana.From this analysis we were able to classify noise events and determine several of their sources.

I. Introduction

Einsteins Theory of General Relativity pre-dicts the existence of gravitational wavesthat are small perturbations in the fabric

of space-time given in the Minkowski Metricgµν[1]

gµν = ηµν + hµν, (1)

Where ηµν is the Minkowski metric for flatspace-time and hµν is a small purturbation inspace-time. Possible sources of gravitationalwaves could result from supernovae explosions,asymmetries of spinning neutron stars, binaryneutron star systems, binary black hole sys-tems, and several other sources. Some energyis released from these systems in the form ofgravitational waves which propogate outward

from the system at the speed of light. Thesegravitational waves (GW) produce a strain onfree-falling test masses that is miniscule inamplitude. For example, If we consider apair of 10 M� black holes revolving aroundeach other at 200 Mpc, we see that the strainamplitude that they produce is equivalent toh ∼ 1 × 10−24, where h is the amplitude ofthe GW[1]. This is an extremely small changein distance as we are trying to measure suchchanges on the order of 1/1000th the diameterof a proton. To measure these minute changesin length we utlize a larger and enhanced ver-sion of the Michelson-Morley Interferometer.

The Laser Interferometer Gravitational-Wave Observatory Scientific Collaboration con-sists of two large interferometric gravitational

∗Advisor: Dr. Florent Robinet

1

mailto:[email protected]


wave detectors, each with an arm length of∼4km. One is stationed in Hanford, Wash-ington and the other in Livingston, Louisiana.These two interferometers are a part of a largerworldwide detector network with sites in Han-nover (GEO 600), Pisa (Virgo), and Tokyo (KA-GRA).

A beam of coherent light (∼ 1064nm) is pro-duced and sent through an input mode cleaner(IM), which is an in-vaccuum suspended trian-gular optical cavity. The IM ’cleans’ the laserbeam by minimizing directional and geomet-ric fluctuations, as well as providing frequencystabilization[2]. The coherent light then travelsthrough a beam spliter and both beams entera Fabry-Perot cavity. This cavity amplifies thepower of the laser as well as the distance trav-eled. These two beams are then reflected backand forth. The two beams then meet again atthe beam splitter where, if a gravitational wavedoes not pass through the system, they com-bine destructively to produce a dark fringe.

When a gravitational wave passes througha system the wave plus-polarizes and cross-polarizes a set of particles in a plane.

Figure 1: Example of plus polarization

When this happens there is a relative differ-ence in arm length between the x and y arms,as the end test masses act as these free particles.Thus the two coherent light beams travelingthrough the interferometer have different pathlengths due to stretching and contracting ofspacetime along the x and y arms. We measurethe path difference in the laser on the photodetector when the two seperate beams interfereat the beam splitter.

We measure the strain amplitude time-series h(t) that is induced on the interferometerfrom gravitational waves in gravitational wave(GW) channels

h(t) = F+h+(t) + F×h×(t), (2)

Where F+ and F× are the attenna paternfunctions, which depend on the sky positionof the GW source, while h+, h× representthe plus and cross polarization of a gravita-tional wave respectively[1]. Sometimes, exter-nal forces cause unwanted disturbances withinGW channels that can mimic a gravitationalwave signal, we call these transients or glitches.At the LIGO-Virgo Collaboration (LVC) there isa group solely dedicated to the detection andclassification of these transients, the detectorcharacterization group (detchar). In the up-coming sections I will discuss the intent andpurpose of the detchar group, as well as myresearch in characterization of the Omicrontrigger generator and LIGO subsystems.

II. Detector Characterization

The purpose of the detchar group is to identifynon-astrophysical instrumental and/or envi-ronmental disturbances in the interferometergravitational-wave observatories. The detchargroup sometimes performs a process called"noise hunting" where we isolate each noisesource and determine a coupling to other chan-nels if there is any.

Both LIGO and Virgo utilize hundredsof different sensors, called auxiliary chan-nels, that measure these external disturbances.These channels include, but are not limitedto, accelerometers, microphones, seismomitors,and voltage monitors. We use auxiliary chan-nels to determine if a candidate event in a GWchannel is really a gravitational wave, or justa glitch coupled with an external disturbanceseen in the auxiliary channels.

The detchar group identifies events withsimilar properties (e.g. same frequency band).We see if an event is correlated with some en-vironmental or physical disturbance and checkthe event times with external scheduled events.

2


Then investigations are made to figure out ifan event in one channel is coincident withan event in other auxiliary channels by usingvarious statistical algorithms such as the use-percentage-veto algorithm (UPV). If we canidentify the source of the noise we attempt toreduce/mitigate the noise source. If we arenot able to understand the source of the noisewe must veto the corrupted data by using dataquality flags[3]. There are several pipelines thatare used to identify "glitches" (Kleine Welle,PCAT, dmt_Omega, Omicron). Some pipelinesare considered better than others in specific ar-eas. For example, Kleine Welle performs wellat high frequency bands, but poor in low fre-quency bands. So, it is essential to understandwhich pipeline is the most useful to detchar in-vestigations. In the upcoming section I will ex-plain my work in understanding the strengthsand weaknesses of the Omicron trigger genera-tor, as well as the role it will play in the futurefor the LIGO/Virgo Collaboration.

III. Characterization of the

Omicron Trigger Generator

I. How Does Omicron Work?

Omicron is a burst-type trigger generator thatreads raw data from given LIGO subsystemchannels specified by the user at a given sam-pling frequency or working frequency (mustbe a power of 2).

Figure 2: This figure is taken from the raw data timeseries in the DARM channel, this is the maininterferometer output signal for GW detection.

GPS Time (s)1088.15494 1088.15495 1088.15496 1088.15497 1088.15498 1088.15499 1088.155

610×

Am

plit

ud

e

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

-610×

Amplitude vs. GPS Time

The pipeline loads data by chunks and thenbreaks these chunks into segments to be ana-lyzed. These segments have a overlap definedby the user (usually 2s) so as to avoid edgeeffects. The power spectrum density for eachchunk of data is then computed. The data vec-tor for each segment is fourier-transformed andnormalized using the PSD. Data is projectedonto a paramter space which is tiled in threedemensions, time, frequency, and Q planes. Qrepresents the quality factor (how gaussian isthe waveform). The amount of Q planes aredefined by a mis-match threshold that is setby the user. The mis-match threshold is theloss in energy between tiles, which can be nogreater than 20 percent; this also determinesthe resolution of the tile. Frequency is then log-arithmicaly distributed in rows which are thenlinearly distributed over time. The amplitudesignal-to-noise ratio (SNR) is then computedfor each tile. Omicron produces triggers whichcan be defined as a tile with an energy above agiven threshold[4].

II. Coincidence Finder Programs

Seeing as one of the main purposes of thedetchar group is to classify glitch families, itis advantageous to determine how well Omi-cron identifies and characterizes discrete typesof simulated burst wave forms. To do this Ideveloped several codes to find coincidencebetween a set of discrete injections (sine-gauss,white noise burst, string cusp). Injections areproduced by the LIGO toolset lalapps_binj andtriggers are produced by Omicron. We utilizedan eight hour long stretch of science data fromthe Hanford Interferometer.

The Coicindence Finder program iswritten in a C++ format and utilizesROOT/GWOLLUM libraries. CoincidenceFinder reads injection files produced bylalapps_binj. Lalapps_binj is a program thatcreates a set of randomly distrubted injectionparamters for a specific type of injection. Coin-cidence Finder stores the relavant parametersfrom the injection file in vectors. CoincidenceFinder performs checks to verify that injection

3


parameters (amplitude (hrss) range, frequency,time) are within Omicron search range. Trig-gers are read and clustered in a .root formatand trigger variables are stored in TTree con-tainers. Time coincidence testing is then per-formed between Omicron triggers and injec-tions based upon a pre-defined coincidencewindow. Coincidence Finder will define a coin-cident event if a trigger and an injection overlapin time. If there are several matches within thewindow, we take the first injection in time, thuspreventing any bias in time.

To quantify the performance of Omicronseveral figures of merit are computed in Coin-cidence finder.

εhrss =NdetNtot

, (3)

Where εhrss represents the detection effi-ciency as a function of amplitdue, Ndet are theinjection events found to be coincident by Co-incident Finder, and Ntot are the total amountof injection events.

ε f req =Ndet f req

Ntot f req, (4)

Where ε f req represents the detection effi-ciency as a function of injection frequency,Ndet f req is detected injection event at a givenfrequency, and Ntot f req is the total amount ofinjection events with given frequency values.

∆peak = (ttrig − tinj), (5)

Where ∆peak is the peak time or change intime between Omicron trigger and injectiontime, ttrig is the trigger time, and tinj is theinjection time.

Λasy =((log10( ftrig)− log10( finj))

(log10( ftrig) + log10( finj)), (6)

Where Λasy is the symmetry of the fre-quency reconstruction, ftrig is the trigger fre-quency, and finj is the injection frequency. It isimportant to explain, in regards to the upcom-ing investigations, a phenomenon called ran-dom coincidence. This phenomenon results from

a statistical artifact in the Coincidence Finderprogram, whereby random triggers found byOmicron at low SNR are found to be in coinci-dence with injections (within overlap window)made by lalapps_binj.

III. Sine-Gaussian Injections

We first test how robust Omicron is in re-trieving all-sky sine-Gaussian injections. Sine-Gauss injections are dependent upon a Q valuewhich determines the width of the waveform,a frequency value, and an hrss value (ampli-tude).

Figure 3: Efficiency curve for sine-Gaussian injectionswith a Q of 75 as a function of the log10(hrss).

log10(Injection hrss)-22 -21 -20 -19 -18 -17 -16 -15

De

tect

ion

Effic

ien

cy

0

0.2

0.4

0.6

0.8

1

Detection efficiency vs hrss

We see in figure 3 that Omicron is able todetect sine-Gaussian injections starting at anhrss of 10−22, quite a promising result. How-ever, as we can see in figures 4 and 5, as theQ value decreases it appears the detection eff-ciency of Omicron decreases as well.


log10(Injection hrss)-22 -21 -20 -19 -18 -17

De

tect

ion

Effic

ien

cy

0

0.2

0.4

0.6

0.8

1


4



log10(Injection hrss)-21.5 -21 -20.5 -20 -19.5 -19 -18.5 -18 -17.5

De

tect

ion

Effic

ien

cy

0

0.2

0.4

0.6

0.8

1


Over the course of our investigations wefound that this discrepency in the Q factor wascorrelated to an issue with the detection eff-ciency as a function of the injected frequency.Because Omicron does not determine a triggerbased on its inherant frequency, there shouldbe no relationship between the detection effi-ciency of Omicron and the injection frequency.However, as we will see in figure 6 below, thisis not case.

Figure 6: Sine-Gaussian effciency curve as a fucntion ofinjection frequency for a Q of 5.

Injected Frequency1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2

De

tect

ion

Effic

ien

cy o

f F

req

ue

ncy

0

0.2

0.4

0.6

0.8

1

Detection efficiency of frequency vs Injected Frequency

We concluded that this trend is due to a con-version done in the injection program we wereutilizing. The injection program "lalapps_binj"takes the input hrss range the user gives in thecommand prompt, and converts this to a newhrss. This is done to make injections lie alongthe 50 percent efficiency curve better[5]

hrsstrue =√

2× π × hrssinj × ( f /Q). (7)

Where the hrssinj represents the convertedhrss, hrsstrue the inputed hrss, f the injectionfrequency, and Q the quality factor, which isrelated to the width of the waveform by therelation below

λw =2π

Q2 . (8)

Where λw is the width of the waveform,and Q is the quality factor. So we see thatthe total overal amplitude of the injections de-creases with a small f and a large Q, and viceversa. The quality factor relationship is illus-trated well in figure 7 below where we see thatthe overal amplitude shifts as Q changes.

Figure 7: Number of injections as a function of ampli-tude. Q of 75 is in black, Q of 30 in blue, andQ of 5 in red.

log10(amplitude)-23 -22 -21 -20 -19 -18 -17 -16 -15

Nu

mb

er

of

Co

inci

de

nt

Eve

nts

0

10

20

30

40

50

60

log10(Amplitude) for All Wave Forms

Figure 8: Plot of Q=75, Q=30, and Q=5 coincident trig-gers. This illustrates the linear relationshipbetween amplitude and frequency.

log10(frequency) (Hz)1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4

log1

0(am

plitu

de)

-23

-22

-21

-20

-19

-18

-17

-16

-15

0

5

10

15

20

25

log10(amplitude):log10(frequency)

In figure 8 we see that there is also a linearrelationship with amplitude as a function of

5


frequency in the injection parameters. We cor-rected both of the issues seen in figures 7 and 8by first adding a step in Coincidence Finder toconvert back to true hrss. Then, to unbias ourinjection sample we weighted injections so thatthose with a high amplitude in low frequencybins were favored more than injections with ahigh amplitude in high frequency bins.

Winj = hrssinj, (9)

Where Winj represents the weight given toa particular set of injections, and hrssinj repre-sents the injected hrss.

ε =

Ndet∑

i=0hrssinj

Ninj

∑i=0

hrssinj

, (10)

Where ε represents the efficiency,Ndet∑

i=0hrssinj

is the sum of the detected injections at a given

amplitude, andNinj

∑i=0

hrssinj the sum of total in-

jections at a given amplitude.Thus, we can confidently say that Omicron

detection effciency is not related to the Q value;or frequency value, the trend we saw is insteadan artifiact of the injection program lalappsbinj and Coincidence Finder.

As mentioned previously in Eq(5), we cal-culated the peak time for each injection type.

Figure 9: Plot showing peak time for Sine-Gaussian Qof 30.

Time (s)-0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08

Nu

mb

er

of

Co

inci

de

nt

Eve

nts

0

20

40

60

80

100

Peak Time

The peak time is calculated by subtracting

the trigger time from the injection time to findout how close in time Omicron retrieves injec-tions. As we can see in Figure 9 the distributionis centered around 0. However, according toCoincidence Finder, Omicron detects injectionsbetween +20ms and −20ms. Ultimately we de-termined this is not an issue associated withOmicron, but instead one related to Coinci-dence Finder. When an injection is made it israndomly given a sky location. The injectiontime is then referenced by the time at the cen-ter of the earth. We calculated that the timeit takes for a gravitational wave to propagatefrom the center of the earth to the surface was∼ 20ms. Thus, depending upon the sky lo-cation the time at which Omicron detects theinjection at the surface of the earth will varybetween + − 20ms. In the future we shouldadd arguments to Coincidence Finder to com-pensate for this discrepency.

IV. Bound Limited White Noise BurstInjections

White noise burst injections are dependentupon an equivalent isotropic radiated energy(solar mass per parsec squared), duration ofinjection (seconds), and the bandwidth of theinjection (Herz). We first measured the detec-tion efficiency as a function of hrss.

Figure 10: Efficiency vs. hrss curve.

log10(injected hrss)-24 -23 -22 -21 -20 -19 -18 -17

de

tect

ion

effic

ien

cy

0

0.2

0.4

0.6

0.8

1


The poor efficiency of the white noiseburst (wnb) injections can be attributed to theunits that are associated with them (equiva-lent isotropic radiated energy). Because we did

6


not know what value the injection generatorprogram would produce for a given isotropicradiated energy, we had to guess and checkwith our amplitude ranges. At low value injec-tion amplitudes (∼ 10−24) there is a 100 percentdetection efficiency. This is due to a statisticalartifact where there are only one or two in-jections made at that injection amplitude andCoincidence Finder happens to find randomcoincidence.

Figure 11: Peak Time of White Noise Burst injection.

Time (s)-0.15 -0.1 -0.05 0 0.05 0.1 0.15

Nu

mb

er

of

Co

inci

de

nt

eve

nts

0

10

20

30

40

50

60

70

80

90

Peak Time

In figure 11 we see that again there is anexcellent reconstruction of the Omicron trig-ger time around the injection time. However,we notice that there is an increase in the over-all background noise ∼ 10. This anomaly isrelated to the duration of white noise burst in-jections. Since white noise bursts can have aduration of up to 2 seconds we initially decidedwe had to increase the size of the injection win-dow. However, because the window for pos-sible coincidence was much larger, there wasa large increase in random coincidence foundby Coincidence Finder. So, we changed therequirements for coincidence to determine thatthere was a match if an injection event andtrigger event overlap in time.

Figure 12: Reconstruction of frequency symmetry foundby Coincidence Finder

Frequency (Hz)-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1

Nu

mb

er

of

Co

inci

de

nt

Eve

nts

0

100

200

300

400

500

600

700

Symmetry of Frequency

Figure 13: Symmetry of Frequency as a function of in-jection frequency.

log10(injection frequency) (Hz)1.8 2 2.2 2.4 2.6 2.8 3 3.2

Sym

me

try

of

Fre

qu

en

cy

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Frequency symmetry vs. frequency inj

While investigating the frequency recon-struction of the Omicron trigger generator wecame across a recurring problem in all injec-tion types. As we can see in figure 12, notall injections are detected at their "true" injec-tion frequency. Instead, we see that there area large amount of injections detected at higherfrequency than they actually are while thereare also a handful detected at a lower frequencythan the "true" injection frequency. This trendshould not exist since Omicron does not deter-mine the significance of a trigger based uponits central frequency value.

In figure 13 we note that as the central injec-tion frequency increases the asymmetry in coin-cident triggers decreases. At a certain injectionfrequency, the asymmetry flips and becomesnegative. We believe that this is not the faultof Omicron, so we took a closer look at thelock data that we were using. It turns out that

7


in the H1:LDAS-STRAIN channel during thistime chunk there is an excess of noise around1KHz. Due to this set of transient events, injec-tions with a central frequency value lower thanthat of this transient family would be foundto be randomly coincident with said transientseries at a higher value and vice versa. This isclearly illustrated in figure 14 where we see thefaint structures of this "glitch" family.

Figure 14: Frequency-time decomposition of 100 secondtime window plotting frequency as a functionof time with a verticle scale in SNR.

V. String Cusp

The third and final injection types we investi-gated were string cusps. These injections aredependent upon a cut off frequency (Hz) andan amplitude. The string cusp wave equationis given below,

h( f ) = A ∗ f−5/3 ∗ θ( f − finj), (11)

Where A is the amplitude of the waveform,f is the frequency, θ represents whether thewaveform is positive or negative, and finj is thecutoff frequency.

Figure 15: Detection efficiency as a function of injectionamplitude.

Injection Amplitude-22 -21 -20 -19 -18

De

tect

ion

Effic

ien

cy

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Detection Efficiency vs. Amplitude

In the detection efficiency graph of figure15 we can see that Omicron starts to detect in-jections at an amplitude of ∼ 10−20. As youwill notice, the efficiency curve does not ap-pear to continue falling towards zero, but in-stead flatens out at ∼ 10−21. This non-linearbehavior represents the zero percent line ofthe efficiency curve and is essentially randomcoincidence matched by Coincidence Finder.

Figure 16: Peak Time of string cusp injections

Time (s)-0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08

Nu

mb

er

of

Co

inci

de

nt

Eve

nts

0

10

20

30

40

50

60

70

Peak Time

In figure 16 above, we see that the peaktime is centered around zero. We can assumethat the elevated background activity aroundthe peak point is due to stochastic noise coinci-dence.

VI. Optimized Omicron

There is an updated version of Omicron thatwill be utilized in both the Virgo and LIGOdetectors on a daily/hourly basis coming out

8


in August 2014. Before the release of the up-dated Omicron we performed tests utlizingCoincidence Finder and GWOLLUM functions.The new version of Omicron was optmized insuch a way that no changes were made to thealgorithms used to calculate the significanceof triggers. For example, one of the changesmade was that some steps in Omicron thatwere previously done during the trigger gen-eration process, were moved to the beginningof the program so as to decrease computationtime. Thus, we should see no change in thetriggers produced by the optmized version ofOmicron.

Initially we found through CoincidenceFinder that new Omicron detects injections ata lower efficiency than its predecessor. As wecan see in figure 16 below, the overall signal tonoise ratio of transients decreases, thus produc-ing triggers with lower resolution at a giventime and frequency band.

Figure 17: The figure below shows triggers producedby the old version of Omicron in black su-perimposed on triggers produced by the newversion of Omicron in red.

In order to determine the root cause of thedecrease in overall SNR we go step-by-stepthrough the Omicron process. In figure 17 be-low we plot the data time series of the timechunk after high pass and filtering with thenew version of Omicron in blue and the oldversion in black doted lines.

Figure 18: Data time series after high pass and filtering.

GPS Time (s)1088.154968421088.1549684221088.1549684251088.1549684281088.154968431088.1549684331088.1549684351088.1549684381088.154968441088.1549684431088.154968445

610×

Am

plit

ud

e

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

-610×

rawdata_L1:LSC-DARM_IN1_DQ_1088154938_1088155002

Figure 17 is a zoomed in view of the ampli-tude of the data as a function of time. As wecan see the new verision of Omicron in blue isthe exact same as the old in the dotted blackline. Next, we can take a quick look at thepower spectrum density.

Figure 19: Zoomed in view of power as a function offrequency computed over the full data timeseries.

log10(frequency) (Hz)

log

10

(Po

we

r)

(Am

p^2

/Hz)

-2910

-2810

-2710

-2610

-2510

-2410

-2310

-2210

-2110

-2010

-1910

-1810

-1710

-1610

-1510

PSD_L1:LSC-DARM_IN1_DQ_1088154938_1088155002

Again, we clearly see the "bug" is not re-lated to this process. After further investi-gations we determined that the cause of this"bug" was related to the conditioning of thedata. During conditioning the data time se-ries, after it has been high passed and filtered,is normalized by the power spectrum densitycomputed over the full data chunk. Duringthis process there was an errant square rootfunction.

9


IV. Transient Analysis of

Commissioning LIGO Data

Recently, the interferomter in Livingston,Louisiana has achieved several 2hr plus locks,where the commisioning team held the twoarms of the interferomter in the correct posi-tions for the laser light to resonate in the cavityproperly[6]. We performed a detailed analy-sis of these time stretches to understand andmitigate "glitch" families within the differentialarm lock channel.

I. Investigation Procedure

First, we run the Omicron trigger generatorover the full stretch of locked data in a GWchannel. We then utilize the Omicron functionGetOmicronPlots which produces several differ-ent histograms (frequency distribution, Time-frequency map, SNR distribution, etc) that wecan use to have a good overall understandingof the lock stretch. In particular, we look at theglitchgram, which is a frequency vs. time distri-bution of triggers with a verticle scale in SNR.We then identify discrete sets of "glitch" fami-lies by a set of triggers in a particular frequencyband.

A few trigger times are identified withineach of these glitch families, we then use thesetrigger times to make Omiscans. An Omiscanis a Omicron function that produces a set offrequency-time decomposition plots for a setof channels given by the user, and an SNRthreshold also defined by the user. Channelsthat appear to have similar transient events tothe main GW channel are compiled in a list.UPV is run over this channel list to identifyany glitch-to-glitch couplings. These resultsare then discussed with individuals on the In-terferomter site to verify the source of glitchesand transients within the main GW channel.

II. June 30th, 2014 Lock

During this time period we looked specificallyat the DC readout time in the differential armlock channel (DARM) ∼ 45 minutes. This chan-nel measures the difference between the x and

y arm lengths (Lx-Ly). This is the main in-terferomter output for GW detection. Afterrunning Omicron we initially see in the glitch-gram, which plots frequency as a function oftime with a verticle scale in SNR, that there aretwo distinct "glitch" famlilies at (∼ 2− 3kHz)and (∼ 250Hz).

Figure 20: Glitchgram plotting frequency as a functionof time.

The series of "glithces" at ∼ 2− 3kHz hasa duration of ∼ 1s and a large bandwidth of∼ 3 − 4kHz with an SNR of ∼ 9 − 13. Weinitially believed this may be correlated withthe "glitch" family at 250Hz, as they both oftenoccur at the same time.

Figure 21: Example of 2-3kHz "glitch" family usingOmiscan. Frequency is plotted vs. time witha verticle SNR scale.

The 250Hz "glitch" family has a durationof ∼ 1s and a short bandwidth with an SNRof ∼ 9− 13. This type of event is usually pre-ceded by three loud glitches at lower frequency

10


bands < 60Hz, suggesting that this transientevent may be correlated with the suspensionchannels.

Figure 22: Example of 250Hz "glitch" family usingOmiscan. Frequency is plotted vs. time witha verticle SNR scale.

To determine if there is any possible co-herence with the differential arm lock channelamong other channels, we run an algorithmcalled use-percentage-veto (UPV). UPV is analgorithm to study glitch-to-glitch couplingsbetween channels. When a coupling is identi-fied it is possible to produce data quality vetoes.The UPV algorithm considers Omicron triggersfrom two input channels: an auxiliary channeland a main channel. These triggers are timeclustered. A time coincidence is performed be-tween these two cluster samples. Two clustersare coincident if they overlap in time. Thenseveral figures of merit are computed.

We consider that the coupling between twochannels is real if the use-percentage is greaterthan a given threshold (usually 0.5). A largeuse percentage value indicates that a triggerin the auxiliary channel is coupled (coincident)with a high probability to a trigger in the mainchannel[7].

After running the algorithm we find thatthere are several candidates for coincidence.However, many have a use percentage thatis far too high to be considered "safe" chan-nels. This is due to the assumption that chan-nels with an unusually high use percentage(50− 100 percent) generally are the "same" asthe main channel, thus they are considered "un-safe" channels. We found that the suspension

channels in the end test masses of both the Xand Y arms showed strong correlation withthe DARM channel, and are considered some-what "safe" depending upon the actuation ofeach channel (veto efficiency = 40 percent to 3percent).

After presenting my results and discussionswith the detchar team in Livingston, we deter-mined the sources of these two "glitch" families.The series of events at ∼ 250Hz was the resultof laser intensity noise in the pre-stablized laser(PSL) optical table and periscope. We foundthat there were some alignment issues. Thus,we understand why the PSL channels did notperform well in UPV due to the fact that thereis no glitch-to-glitch coupling between the two.The "glitch" family at ∼ 2− 3khz is possiblyrelated to noise modulated by angular fluctua-tions in the differential ETM control module.

III. July 15th, 2014 Lock

During this time segment we want to see iftransients present in the previous lock timewere either mitigated, or reduced. We can havea good initial idea of this by comparing theglitch grams of the two lock times.

Figure 23: Glitchgram of July 15th lock time plottingfrequency as a function of time.

The "glitch" family at ∼ 2− 3kHz is com-pletely mitigated while the family at 250Hzhas been reduced significantly. Altough, we dosee that between 100 Hz and 200 Hz there aremany more lines of transients that have been

11


produced. We also notice a strong line around1.5kHz. If we look closer at the discrete lineswe notice that the transient events at 103Hzand 128Hz are essentially same. This is alsotrue for the pair at 158Hz and 207Hz.

Figure 24: Omiscan of 103hz glitch family.




We believe that these two pairs are essen-tially a part of the same "glitch" family. Afterrunning UPV over the targeted channels we de-termined that these broadband low frequency"glitches" in the DARM channels were the re-sult of motion in the PSL tables, as noted bycoherence found in the PSL accelerometer chan-nel by STAMP-PEM runs. These might also becoupled with suspension channels as given bythe effciciency (vetoe efficiency = 10 percent).UPV did not select acceleromter channels wellbecause the coupling is probably non-linear.The source of the family at high frequenciesmay be from a mix of mechanical and/or digi-tal artifacts.

V. Final Remarks

This project has been fairly successful in deter-mining the robust nature of Omicron whendealing with different injection "types". Inregards to sine-gaussian injections we deter-mined that, unlike Kleine Welle, Omicron isable to detect transients with an equal effi-ciency at both low and high frequency. Inaddition, we found that injections with an am-plitude as low as a strain amplitude of 10−22-10−21 were still able to be detected. Despitethe reference time of the injection, Omicronis able to reconstruct the peak time and fre-quency with a low error. In White noise burstsignals we initially found an asymmetry in thefrequency reconstruction, and determined thecause not to be the fault of Omciron, but in-stead associated with background noise in the

12


channel.Through these studies we can conclude that

Omicron classifies "glitch" families with a greatprecision and has been useful in the charac-terization of aLIGO sub-systems on post-ER5data. The Coincidence Finder programs willbe used in future versions of Omicron to makesanity checks. The results from this project areimportant as they validate the usefulness ofOmicron as both a trigger generator and bursttype search pipeline that will be used in futureobservational runs at LIGO and Virgo.

VI. Acknowledgments

I would like to thank the National ScienceFoundation for funding this IREU program.Thank you to the University of Florida, Dr.Bernard Whitting, Dr. Guido Muller, and Dr.Andonis Mitidis for allowing me this wonder-ful opportunity to experience a new and excit-ing research environment. A special thanks toDr. Florent Robinet for his constant help andsupport during this project (I sometime won-der how he ever got any work done). Thankyou to all the members of the LAL group at Or-say, who were extremely warm and welcoming.Lastly, thank you to my advisor at my homeinstitution Dr. Marco Cavaglia, who partly in-spired me to pursue research in gravitationalphysics.

References

[1] Saulson, Peter R. Interferometric Gravita-tional Wave Detectors. Singapore: World

Scientific Publising, 1994. Print.

[2] "LIGO Channel List." CIS. Caltech, n.d.Web.<https3A2F2Fcis.ligo.org2Fchannel2F>.

[3] LSC Members. "The Characteriza-tion of Virgo Data and Its Impact onGravitational-wave Searches." ArXiv(2012): 1-50. ArXiv. Web. June-July 2014.<http://arxiv.org/>.

[4] Robinet, Florent. "GWOL-LUM and Friends." GWOL-LUM. LAL/Virgo, n.d. Web.<http3A2F2Fvirgo.in2p3.fr2FGWOLLUM2Findex.html3FFriends2Fomicron.html>.

[5] Brown, Duncan. "LALApps âAT LSC Al-gorithm Library Applications." Universityof WisconsinâASMilwaukee (2005): 1-182.LSC Algorithm Library Applications.Web. June-July 2014. <https://www.lsc-group.phys.uwm.edu/daswg/projects/lalapps/lalapps-5.0.pdf>.

[6] "LIGO Locks Its First Detec-tor!" LIGO Locks Its First Detec-tor! N.p., n.d. Web. 29 July 2014.<http://www.phys.lsu.edu/mog/mog17/node11.html>.

[7] Isogai, Tomoki, and The Ligo ScientificCollaboration A Collaboration. "Used Per-centage Veto for LIGO and Virgo Bi-nary Inspiral Searches." Journal of Physics:Conference Series 243 (2010): 012005.Web.

13

Documents

IREU Final Report Characterization of the Omicron Trigger ... · Characterization of the Omicron Trigger Generator and Transient Analysis of aLIGO Data Hunter Gabbard Virgo Detector