Emissioonijoontega tähtede tuvastamine Gaia kosmoseteleskoobi andmetest
Files
Date
2009
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Tartu Ülikool
Abstract
Aastal 2011 stardib Euroopa Kosmoseagentuuri (ESA) kosmoseteleskoop Gaia. Teleskoop vaatleb viie aasta jooksul hinnanguliselt miljard objekti, kogudes seeläbi kokku ühe petabaidi jagu andmeid.
Käesoleva bakalaureusetöö eesmärgiks oli uurida, kuidas saab kasutada masinõppe algoritme Gaia kosmoseteleskoobi madalalahutusega BP/RP fotomeetri andmete töötlemisel. Kuna reaalseid andmeid veel ei eksisteeri, siis kasutati töös olemasolevatest kõrge lahutusega spektritest koostatud Gaia vaatluste simulatsioone.
Bakalaureusetöö tulemused saab jagada kaheks osaks. Kõigepealt uurisime tugivektormasinate kasutamist kahe erineva klassifikatsiooniülesande lahendamisel. Esimese ülesandena proovisime eraldada H-alfa emissiooni sisaldavaid spektreid H-alfa emissioonita spektritest. Saadud tulemused olid hoolimata piiratud andmetest head. Teise klassifitseerimisülesandena katsusime eraldada Be tähti Wolf-Rayet’ tähtedest.
Teise ülesandena üritasime me hinnata tugivektorregressiooni algoritmiga H-alfa spektrijoone ekvivalentlaiust. Selleks sobiva andmestiku saamiseks kirjutasime me lihtsa emissioonijoontega tähespektri mudeli, mis ignoreerib tegelikkuses spektrit tekitavaid füüsikalisi protsesse. Mudel kasutab peakomponentide analüüsi, et genereerida vaatlustega sarnaseid tähe kontiinuume. Spektrijooned lähendasime Gaussi funktsioonidega. Esialgset tulemused tugivektorregressiooni kasutamisel olid positiivsed.
Tulevikus on plaanis katsetada erinevaid meetodeid klassifitseerimisalgoritmide effektiivsuse parandamiseks, lisada emissioonijoontega tähespektri mudelile astrofüüsikalistest kaalutlustest tuletatud spektrijooni ning uurida tugivektorregressiooni kasutamist teiste spektrijoonte effektiivlaiuse tuvastamisel.
The Gaia probe, set to launch in 2011, will measure an estimated billion astronomical objects during its five-year mission, producing roughly one petabyte of data. The goal of this thesis is to study the possibilities of using machine learning algorithms for analysing data produced by the low-resolution BP/RP photometer of the space probe. We use simulations of the Gaia observations in our experiments. The results of the thesis are divided into two parts. First, we look into using support vector machines for solving two different classification tasks. For the first classification task, we try to separate spectra with H-alpha emission from regular stars, obtaining reasonable results. We proceed to further classify Be stars and Wolf-Rayet’ stars. Second, we try to estimate the equivalent width of the H-alpha spectral line using support vector regression. We introduce a simple, non-physical model for the spectrum of an emission line star. The model applies principal component analysis in order to generate the spectral continuum. Spectral lines are approximated using Gaussian functions. Preliminary results in using support vector regression are encouraging. Future work includes improving the effectiveness of the classification algorithms, adding additional spectral lines to the spectral model as well as applying support vector regression to other spectral lines.
The Gaia probe, set to launch in 2011, will measure an estimated billion astronomical objects during its five-year mission, producing roughly one petabyte of data. The goal of this thesis is to study the possibilities of using machine learning algorithms for analysing data produced by the low-resolution BP/RP photometer of the space probe. We use simulations of the Gaia observations in our experiments. The results of the thesis are divided into two parts. First, we look into using support vector machines for solving two different classification tasks. For the first classification task, we try to separate spectra with H-alpha emission from regular stars, obtaining reasonable results. We proceed to further classify Be stars and Wolf-Rayet’ stars. Second, we try to estimate the equivalent width of the H-alpha spectral line using support vector regression. We introduce a simple, non-physical model for the spectrum of an emission line star. The model applies principal component analysis in order to generate the spectral continuum. Spectral lines are approximated using Gaussian functions. Preliminary results in using support vector regression are encouraging. Future work includes improving the effectiveness of the classification algorithms, adding additional spectral lines to the spectral model as well as applying support vector regression to other spectral lines.