@PhDThesis{Ferreira:2023:ApTéMa,
author = "Ferreira, T{\'a}bata Aira",
title = "Aplica{\c{c}}{\~a}o de t{\'e}cnicas de Machine Learning no
estudo de transientes dos detectores Advanced LIGO",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2023",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2022-12-09",
keywords = "glitches, LIGO, aprendizado de m{\'a}quina, an{\'a}lise de
redes, ondas gravitacionais, machine learning, network science,
gravitational waves.",
abstract = "As detec{\c{c}}{\~o}es diretas de ondas gravitacionais n{\~a}o
apenas trouxeram mais uma evid{\^e}ncia da Teoria da Relatividade
Geral de Einstein, mas inauguraram uma nova astronomia. Os
observat{\'o}rios LIGO foram os pioneiros na detec{\c{c}}{\~a}o
desse tipo de sinal e dezenas de eventos j{\'a} foram
catalogados. O n{\'u}mero progressivo de detec{\c{c}}{\~o}es
fomenta a {\'a}rea e gera expectativas em diferentes
observat{\'o}rios e cientistas ao redor do mundo. Entretanto,
n{\~a}o s{\~a}o apenas os sinais de eventos
astrof{\'{\i}}sicos que aparecem nos dados destes detectores,
mas diferentes ru{\'{\i}}dos transientes oriundos de diversos
fatores ambientais, instrumentais ou antropog{\^e}nicos. Estudar
esses invasores locais, usualmente denominados glitches, {\'e}
sempre um desafio para a colabora{\c{c}}{\~a}o
cient{\'{\i}}fica, pois alguns deles t{\^e}m alta taxa de
ocorr{\^e}ncia, podem mimetizar ondas gravitacionais, poluir os
dados e diminuir a signific{\^a}ncia estat{\'{\i}}stica de um
sinal astrof{\'{\i}}sico real. Infelizmente, alguns desses
transientes n{\~a}o t{\^e}m causas identificadas ou definidas, e
a tentativa de buscar tais ind{\'{\i}}cios incentivou este
trabalho. Esta tese apresenta uma forma alternativa para
caracterizar e encontrar classes de glitches no canal
gravitacional, a partir dos denominados glitchgramas. Duas
t{\'e}cnicas computacionais foram utilizadas para avaliar a
efici{\^e}ncia dessa caracteriza{\c{c}}{\~a}o proposta. A
primeira aplicou ferramentas de An{\'a}lise de Redes e a segunda,
de Aprendizado de M{\'a}quina; ambos resultados foram comparados
com as classifica{\c{c}}{\~o}es pr{\'e}vias do Gravity Spy,
ferramenta utilizada pela colabora{\c{c}}{\~a}o para classificar
transientes. A an{\'a}lise de redes obteve resultados excelentes
para determinadas classes, mas nem tanto para outras e, portanto,
limita{\c{c}}{\~o}es no uso dessa t{\'e}cnica a partir de
glitchgramas foram encontradas. No geral, o m{\'e}todo teve 75,
03% de concord{\^a}ncia com Gravity Spy e, com o cosseno de
similaridade, apresentado na t{\'e}cnica, foi poss{\'{\i}}vel
atribuir classes a glitches desconhecidos. O segundo m{\'e}todo
foi efetivo na busca de todas as classes investigadas. Com a
aplica{\c{c}}{\~a}o de uma ferramenta do aprendizado de
m{\'a}quina supervisionado, uma valida{\c{c}}{\~a}o cruzada foi
realizada e o m{\'e}todo concordou em 94, 70% com o Gravity Spy.
Valor que poderia ter sido maior, pois o m{\'e}todo apontou erros
de classifica{\c{c}}{\~a}o do atual modo de an{\'a}lise do
LIGO. O aprendizado de m{\'a}quina ainda mostrou-se independente,
p{\^o}de ser aplicado em an{\'a}lises di{\'a}rias de
funcionamento do LIGO, em busca da presen{\c{c}}a de classes em
canais auxiliares, abriu campos para diferentes
aplica{\c{c}}{\~o}es e permitiu concluir que os glitchgramas
caracterizam bem os glitches. Para exemplificar isso, esta tese
tamb{\'e}m apresenta um estudo das duas classes mais presentes
durante a terceira corrida observacional do LIGO: a Scattered
Light e a Fast Scattering. Para esta {\'u}ltima, uma
investiga{\c{c}}{\~a}o sobre sua rela{\c{c}}{\~a}o com
movimentos micross{\'{\i}}smicos e antropol{\'o}gicos foi
realizada. ABSTRACT: The gravitational wave detections not only
provided further evidence for Einsteins Theory of General
Relativity but also inaugurated a new astronomy. LIGO
observatories were pioneers in detecting such types of signals,
and dozens of events have already been cataloged. The progressive
number of detections has been promoting the area and generating
expectations in scientists worldwide. However, it is not only the
signals of astrophysical events that appear in the data of these
detectors, different transient noise sources arise from various
environmental, instrumental, or anthropogenic factors. Studying
these local invaders, generally called glitches, is always a
challenge faced by the scientific collaboration, as some of them
have a high occurrence rate, may mimic gravitational waves,
pollute the data and decrease the overall statistical significance
of a real astrophysical signal. Unfortunately, some of these
transients do not have identified or present well-defined reasons,
and the attempt to look for such indications encouraged this work.
This dissertation presents an alternative way to characterize and
find classes of glitches in the gravitational channel, based on
the so-called glitchgrams. Two computational techniques were used
to evaluate the efficiency of this proposed characterization. The
first applied Network Analysis tools and the second Machine
Learning tools; both results were compared with previous Gravity
Spy classifications, the tool used by the collaboration to
categorize transients. Network Science obtained excellent results
for some classes but not so much for others; therefore,
limitations in using this technique from glitchgrams were found.
Overall, the method had 75.03% of agreement with Gravity Spy, and,
with the cosine of similarity, presented in the method, it was
possible to assign classes to unknown glitches. The second method
was effective in searching all investigated categories. With a
supervised machine learning tool, cross-validation was performed,
and the technique agreed at 94.70% with Gravity Spy. The value
could have been higher, as the method pointed out classification
errors in the current LIGO analysis mode. Machine learning still
proved to be independent, could be applied in daily analyzes of
LIGOs operation, in search of the presence of classes in auxiliary
channels, opened fields for different applications, and allowed
concluding that glitchgrams characterize well glitches. In order
to exemplify this, this dissertation also studies the two most
common classes during the third LIGO observational run: the
Scattered Light and the Fast Scattering. For the latter, an
investigation into its relationship with microseismic and
anthropological motions was performed.",
committee = "Aguiar, Odylio Denys de (presidente) and Costa, C{\'e}sar Augusto
(orientador) and Tinto, Massimo and Jablonski, Francisco Jos{\'e}
and Lenzi, C{\'e}sar Henrique and Tosta e Melo, Iara",
englishtitle = "Application of Machine Learning techniques in the study of
transients from Advanced LIGO detectors",
language = "pt",
pages = "120",
ibi = "8JMKD3MGP3W34T/48AFMK5",
url = "http://urlib.net/ibi/8JMKD3MGP3W34T/48AFMK5",
targetfile = "publicacao.pdf",
urlaccessdate = "27 abr. 2024"
}