Deep Convolutional Variational Autoencoder for Anomalous Sound Detection

Anomalous sound detection (ASD) is one of the most important fields in industrial facility maintenance. For this task, semi-supervised approaches are preferred thanks to their simplicity and no training data labels required. These methods train an autoencoder (AE) with only normal sound data and detect anomalies based on anomaly scores of actual samples. In this paper, we propose applying the convolutional variational autoencoder (CVAE) to ASD task. Through experiments using machine sound data, the CVAE is proven to be effective in detecting abnormal sound and outperform existing methods.