PPGINF PÓS-GRADUAÇÃO EM ENGENHARIA DA INFORMAÇÃO FUNDAÇÃO UNIVERSIDADE FEDERAL DO ABC Phone: Not available http://propg.ufabc.edu.br/ppginfo

Banca de DEFESA: CAROLINE PIRES ALAVEZ MORAES

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
DISCENTE : CAROLINE PIRES ALAVEZ MORAES
DATA : 03/02/2020
HORA: 14:00
LOCAL: sala 301, 3º andar, Bloco B, Campus SA da Fundação Universidade Federal do ABC, localizada na Avenida dos Estados, 5001, Santa Terezinha, Santo André, SP
TÍTULO:

New Approaches to Blind Source Separation in Post-Nonlinear Mixtures


PÁGINAS: 84
GRANDE ÁREA: Engenharias
ÁREA: Engenharia Elétrica
SUBÁREA: Telecomunicações
ESPECIALIDADE: Sistemas deTelecomunicações
RESUMO:

In the signal processing area, the Blind Source Separation (BSS - Blind Source
Separation) problem occupies an important position in view of its versatility and possible
practical applications. Even though the theoretical framework is well consolidated in
the linear context, the generic non-linear approach still lacks methodologies capable
of ensuring the separation of the sources, which makes this research field very actual
and challenging. In the context of nonlinear mixtures, the prior knowledge of some
additional information such as temporal structure and a priori knowledge of certain
characteristics of the sources can help in the development of new separation methods
that are more robust. This work proposes two approachs using the post-nonlinear model
PNL (Post-Nonlinear), one based on the minimization of mutual information and
the second one using second order statistics. In the first approach, it
is necessary to estimate the distribution of the sources, which can be
done using kernel functions. Usually, the Gaussian kernel function is used.
However, other kernel functions with interesting properties can be applied, such
as the Epanechnikov kernel. We apply both functions to estimate the pdf, showing that the method developped with the Epanechnikov kernel performs better than with the Gaussian kernel. In addition,
the prior knowledge of some additional information such as temporal structure and a
priori knowledge of certain features of the sources can help in the development of new
separation methods that are more robust. Most separation techniques involve higher
order statistics and algorithms that use neural networks or metaheuristic. In this work, we also developed a simple separation algorithm called
A-SOBIPNL, which is based on the gradient descent, and only uses second-order statistics
to explore the temporal structure of the source signals. For this, we combine two
classical algorithms, AMUSE and SOBI, to apply on the linear and nonlinear stage,
respectively, obtaining a good performance of the algorithm.


MEMBROS DA BANCA:
Presidente - Interno ao Programa - 1544392 - ALINE DE OLIVEIRA NEVES PANAZIO
Membro Titular - Examinador(a) Interno ao Programa - 1761107 - RICARDO SUYAMA
Membro Titular - Examinador(a) Externo à Instituição - LEONARDO TOMAZELI DUARTE
Membro Suplente - Examinador(a) Interno ao Programa - 1761105 - MURILO BELLEZONI LOIOLA
Membro Suplente - Examinador(a) Externo ao Programa - 1932365 - FABRICIO OLIVETTI DE FRANCA
Notícia cadastrada em: 04/01/2020 17:10
SIGAA | UFABC - Núcleo de Tecnologia da Informação - ||||| | Copyright © 2006-2021 - UFRN - sigaa-2.ufabc.int.br.sigaa-2