Banca de QUALIFICAÇÃO: ALEXANDRE MICCHELETI LUCENA

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : ALEXANDRE MICCHELETI LUCENA
DATE: 16/05/2025
TIME: 10:00
LOCAL: Remoto (https://conferenciaweb.rnp.br/sala/ricardo-33)
TITLE:

On the Blind Source Separation of Nonlinear Mixtures

PAGES: 43
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

Blind source separation can be considered one of the main problems within the realm of
signal processing theory and has vast applications in various fields. While linear mixture models
have been extensively studied, many real-world applications involve nonlinear mixing, posing
additional challenges. Regular Independent Component Analysis (ICA) methods, which rely
on statistical independence, are insufficient for general nonlinear cases. However, the work by
Ehsandoust et al. (2017) proposed a novel approach for smooth nonlinear mixtures, leveraging
local linear approximations to enable ICA-based separation. Interestingly, the initially proposed
method — based on deriving the observed signals, performing an adaptive linear blind source
separation, smooth the coefficients of the separation matrices and integrate the solution — can
be reduced, in some cases, by simply performing an adaptive linear blind source separation and
smooth the coefficients of the separation matrices.
This study investigates the effectiveness of this method under varying conditions on the
blind source separation of nonlinear mixtures. We analyze a different model for the source
signals considering autoregressive (AR) sources with distinct frequency characteristics to assess
the impact on separation performance. Additionally, we modify the general algorithm by
incorporating a General Regression Neural Network (GRNN) as a replacement for the nonlinear
regression step and evaluate its performance. Result shows that this adaption of the algorithm
achieves the best performance in most of tested scenarios. Simulations showed that the usage
of the GRNN as a smoothing algorithm improved the source separation performed even when
applied without the local approximation in some cases. Moreover, a different type of smooth
source based on AR(1) signals is considered, showing that the algorithm was able to perform
separation, hence opening space to considering random processes in this kind of modeling.

COMMITTEE MEMBERS:
Presidente - Interno ao Programa - 1761107 - RICARDO SUYAMA
Membro Titular - Examinador(a) Interno ao Programa - 1544392 - ALINE DE OLIVEIRA NEVES PANAZIO
Membro Titular - Examinador(a) Interno ao Programa - 1946319 - DIOGO COUTINHO SORIANO
Membro Suplente - Examinador(a) Interno ao Programa - 1761105 - MURILO BELLEZONI LOIOLA