Two stage system for anomalous sound detection in industrial environments

Abstract

Machine breakdowns and maintenance breaks cause costly downtime in factories and power plants. Recognizing a breaking machine before the actual breakdown can reduce the downtime and size of the damage. The existing condition monitoring systems are usually based on measuring the vibrations in the machines. In industrial environments the acoustic properties are relatively homogeneous during normal operation, and machine failures cause change in those properties. Therefore a change in acoustic conditions reflects an anomalous event that can be detected through analysis of audio signals at the scene. However, sometimes normal operation like talking or door slamming can cause a significant change in the acoustic conditions, and those should be ignored.

This thesis presents a two-stage acoustic anomaly detection system. The motivation behind using two stages is to offer the operator a possibility to silence certain anomaly types. This makes it possible to ignore normal events that are anomalies from acoustic point of view but do not indicate a need for alarm.