ANOMALY DETECTION WITH SELF-SUPERVISED AUDIO EMBEDDINGS

Abstract

The majority of approaches to machine condition monitoring viaanomalous sound detection are based on supervised learning.  Themetadata of the datasets is used as data labels for training super-vised models.  However, data labeling is expensive and often im-possible  for  industries  with  significant  amount  of  equipment.   Inthis  case  self-supervised  methods  could  solve  the  problem  sincethey do not require labeled data. In this work we applied the recentself-supervised approach to compute embeddings of audio signalsnamed BYOL-A and classical machine learning method Local Out-lier Factor (LOF) to compute outlier scores for anomalous sounds.The main focus of this work is to not use any labels from the meta-data of the datasets and explore a self-supervised learning approach.