Detecting Abnormal Machine Sounds Using An Ensemble Approach with Data Augmentation Techniques

In this paper, we present our methodology in the DCASE 2021 Challenge, specifically Task 2: Unsupervised Detection of Anomalous Sounds in Machine Condition Monitoring under Domain Shift Conditions. The primary goal was to ascertain whether sounds emitted by a machine are normal or anomalous, utilizing a test dataset devoid of any abnormal sounds in its training counterpart. To tackle this challenge, we utilized deep learning models, namely MobileNetV2 and ResNetV2-50, which were enhanced with data augmentation techniques like mix-up to improve anomaly detection in machine sounds. The results of our evaluation are presents the harmonic mean AUC and pAUC across various machines including ToyCar, ToyTrain, Fan, Gearbox, Pump, and Slider. Notable results are an AUC of 54.77 and pAUC of 58.83 for ToyCar, with other machines showing varied performance metrics, such as a 44.10 AUC for Gearbox and a 59.02 AUC for Slider.