Banca de DEFESA: GILVAN FARIAS DA SILVA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
DISCENTE : GILVAN FARIAS DA SILVA
DATA : 30/10/2020
HORA: 14:00
LOCAL: https://conferenciaweb.rnp.br/webconf/laboratorio-de-sistemas-digitais-ufba
TÍTULO:

CORROSION DETECTION USING PULSED EDDY CURRENTS AND NEURAL CLASSIFIERS WITH TRAINING CONSTRAINTS

PALAVRAS-CHAVES:

Non-destructive testing, Machine learning, Artificial neural netwoks

PÁGINAS: 113
GRANDE ÁREA: Engenharias
ÁREA: Engenharia Elétrica
RESUMO:

Corrosion of metal pipes is a common problem in the industry. If not detec-
ted in time, it can cause pipe disruption, resulting in loss of production and
serious accidents. At insulated industrial pipes the corrosion can remain hid-
den under thermal insulation, or inside the pipe. To detect corrosion visually,
it is necessary to temporarily remove the thermal insulation. However, this
removal may be impractical with the industrial plant in operation. There-
fore, in this work, a type of non-destructive test called pulsed eddy currents
(PEC) was used to detect this hidden corrosion. The interpretation of the
test depends on the experience and technical skill of the operator. At aplica-
tions like this, each misclassification has different consequences. Classifying a
corroded pipe as not having corrosion inhibits the execution of proper main-
tenance, increasing the chances of the pipe breaking. In this work, it was
proposed a neural classifier to assist the operator in decision making. This
classifier was trained with particle swarm optimization with restrictions, in
order to consider the different classification errors in different ways. It was
considered errors that increase the risk of accidents are more undesirable than
errors that cause only financial losses. To evaluate the proposed method, it
was used PEC signals acquired from pipes of a petrochemical industry. The
classifier trained by the proposed method did not present classification errors
that compromise the safety of the industrial plant. Contrary to what was
observed with classifiers trained by the different variations of the backpropa-
gation method analyzed in this work.

MEMBROS DA BANCA:
Presidente - 1856678 - PAULO CESAR MACHADO DE ABREU FARIAS
Interno - 2506534 - EDUARDO FURTADO DE SIMAS FILHO
Externo à Instituição - Nadia Nedjah
Externo à Instituição - Werner Spolidoro Freund