Filtragem de ruído em sinais de eletrocardiograma e áudio utilizando autoencoders e mecanismos de monitoramento

Abstract

This study proposes an approach based on autoencoders with residual blocks and monitoring mechanisms for filtering electrocardiogram signals contaminated by noise, both in isolation and in combination, as well as spectrograms corrupted by common urban environmental noise generated from audio files. Two distinct architectures were evaluated: one using a Squeeze-and-Excitation block and another incorporating a monitoring mechanism called the Attention Block. The models were tested on electrocardiogram signals and spectrograms contaminated with noise to assess their performance in signal reconstruction and noise removal tasks. The Attention Block demonstrated superior performance in both tasks. This advantage is attributed to its ability to integrate channel-wise attention mechanisms and local noise detection, which proved relevant when multiple types of i