Activity-Informed Industrial Audio Anomaly Detection Via Source Separation

Jaechang Kim, Yunjoo Lee, Hyun Mi Cho, Dong Woo Kim, Chi Hoon Song, Jungseul Ok

Abstract: We discuss a practical scenario of anomaly detection for industrial sound data where the sound of a target machine is corrupted by not only noise from plant environments but also interference from neighboring machines. This is particularly challenging since the interfering sounds are virtually indistinguishable from the target machine without additional information. To overcome these challenges, we fully exploit the information of machine activity or control that is easy to obtain in the industrial environment, and propose a framework of source separation (SS) followed by anomaly detection (AD), so called SSAD. We note that the proposed SSAD utilizes the activity information for not only AD but also SS. In our experiment based on industrial dataset, we demonstrate that the proposed method using only mixture signal and activity information achieves comparable accuracy with an oracle baseline using clean source signals.