Activity guided industrial anomalous sound detection combined with source separation

Abstract

We suppose a practical scenario of anomalous sound detection in industry where the recorded sound of a target machine includes background noise from factories and interference from nearby machines. This is especially challenging since the neighboring machines often generate sounds which are hardly distinguishable from the target machine without additional information. To overcome these challenges, we fully utilize the information of machine activity or control that is comparatively easy to obtain in the industries and propose a framework of source separation (SS) followed by anomaly detection (AD), coined as SSAD. We note that the proposed SSAD exploits the activity information for not only anomaly detection but also for source separation. In our experiments based on the industrial dataset, results demonstrate that the proposed framework using mixture signal and source activity information shows comparable performance in terms of AUC with oracle baseline using clean source signals.