Comparative Study of Different Activation Functions for Anomalous Sound Detection

Abstract

Anomaly detection is of great importance in our modern life as it can be very useful in many ways such as lowering costs, avoiding potential injuries, and even saving lives. The method of anomalous sound detection using the Self-Supervised Learning (SSL) approach is effective and has a relatively low training time with the use of a Convolutional Neural Network (CNN). The use of techniques such as the pre-training and hard sample with the SSL approach led to producing very high results scoring results higher than 0.9 for the Area under the curve (AUC) score especially for the non-stationary sounds. An AutoEncoder (AE) based system developed by the Detection and Classification of Acoustic Scenes and Events (DCASE2020) competition’s organizers were used for comparative purposes to compare the results of the SSL method with the results of the baseline system. The results of the baseline system and the results of the SSL approach with different configurations and the SSL method had shown higher results in most cases.