Condition monitoring based on anomalous sound detection via autoencoders

Condition Monitoring is the process of monitoring a parameter of a particular machine, for the purpose of identifying developing anomalies. In this thesis, in cooperation with 221e S.r.l. during an internship, an autoencoder-based Condition Monitoring system is proposed, with the aim of detecting anomalies in machines using sound signals. Sound indicators in Condition Monitoring offer multiple advantages over more traditional metrics like temperature, vibration, or voltage. Anomalies can be detected before major malfunctions occur, the machine can be monitored without physical contact and a large set of different anomalies can be detected. The system was developed and evaluated on three real user scenarios provided by the company. Different conditions and settings, with customized data acquisitions and training of the model, were considered. In the end, an embedding of the monitoring solution into the microcontroller multi-sensor board STWIN is considered and validated on the device.