Modelos geradores para detecções de anomalias em atividades sonoras

Abstract

Several data domains allow the use of anomaly detection, including audio. An important feature of these systems is to identify when something is different from ordinary. For this purpose, several studies using machine learning were performed. The state of art in anomaly detection in images uses architectures based on GAN (Generative Adversarial Network), however, few studies demonstrate the use of these or other generating architectures in the domain of sounds. To overcome this problem, this work aims to develop a method for identifying anomalies in sound activities using data captured through microphones. The anomaly identification process is carried out through a generator model from a deep network architecture. Tests using real databases show that some changes in the architectures used for images can achieve promising results. This approach has been validated using the DCASE 2021 dataset, which includes over 180 hours of audio from industrial machinery. We evaluated the classification of anomalies, reporting an weighted average of 88,16% AUC and 78,05% pAUC, results superior to those presented by baselines