Poster Abstract: Adapting Pretrained Features for Efficient Unsupervised Acoustic Anomaly Detection

Abstract

Faults in industrial equipment lead to significant costs due to downtime and unplanned maintenance interventions. Acoustic Internet of Things (IoT) sensors combined with machine learning offers the possibility of early fault detection to mitigate these costs. However, prior approaches to Acoustic Anomaly Detection (AAD) are poorly suited to operating within the resource constraints of IoT systems as they require the acquisition of large volumes of data to train models from scratch for different machine types or operating environments. To overcome the limitations, we introduce a system utilizing pretrained low-dimensional features and Gaussian mixture models. Preliminary results on real-world datasets show that our proposed approach outperforms state-of-the-art solutions based on the area under the curve (AUC) score with an average of 4 different machine types. Furthermore, our approach requires far less training data, making it more suitable to operate within the power and network constraints of IoT devices.