Industrial Machine Monitoring: Real-Time Anomalous Sound Event Detection on Low-Powered Devices

Sammanfattning: Traditionally fault detection in industrial machinery has been performed manually by experienced machine operators listening to the machines. However, it is desirable to automate this process to increase efficiency and improve the working environment of the operators. The main challenge in this thesis is to create a system that can accurately detect when an anomalous sound event occurs, at the same time the system may not report too many false alarms. Additionally, the system must perform the detection fast enough for the detected anomalies to still be relevant. This thesis therefore explores lightweight machine learning approaches to anomalous sound event detection, such as Gaussian Mixture Models (GMM) and One-Class Support Vector Machines (OCSVM). The experiments evaluate how low-level descriptors from the time-, spectral- and cepstral-domain perform as features modeling the characteristics of a sound segment. Another set of experiments evaluates if it is possible to detect anomalies fast enough to achieve real-time anomaly detection. Lastly, a real-time anomaly detection application based on the findings is presented together with the results for a few test runs. The results indicate that it is possible to detect anomalies of sufficient magnitude in relation to the expected signal. Furthermore, it is found that it is possible to detect anomalies at a speed fast enough to enable real-time anomaly detection on limited hardware such as a Raspberry Pi 4.