Machine-type independent anomaly sound detection by using conditional autoencoder

Abstract

Recently, anomaly sound detection methods based on reconstruction error of autoencoder have been widely used, but conventional methods require separate models for each machine type. Thus, they train models only on single machine-type data and do not use whole training data containing various machine-type data. To address these two problems, we propose to use a single conditional autoencoder in order to deal with all machine types of sound data simultaneously. Experiments on the task2 of the DCASE 2022 challenge show the effectiveness of our proposed method.