Desenvolvimento de redes neurais auto-associativas para classificação de ruído acústico emitido por isoladores cerâmicos.

Abstract

This work aim ed to develop an algor ithm of an auto-associative Neural network to recognize patterns of ultrasound emitted by an electric insu lator, as well as to com pare the perform ance of the deve loped algorithm with an existing Neural Network, dedicated to the same task. The preliminary results promoted the development of an advanced algorithm that is capable of improving the m atch rate at th e expense of higher computational effort during the traini ng phase. The results achieved with this algorithm are prom ising, especially consideri ng that various factors that influence the performance of an AutoAssociativ e Neural N etwork were not analyzed, like the algorithm used to train the Network. Consider ing that the computational effort involved is mainly in the train phase this algorithm might be useful in areas where the m atch rate obtained with the trad itional m ethods are n ot enough. All th e algorithm s were developed and tested use the MatLab Software.