OutlierNets: Highly Compact Deep Autoencoder Network Architectures for On-Device Acoustic Anomaly Detection

Human operators often diagnose industrial machinery via anomalous sounds.\nAutomated acoustic anomaly detection can lead to reliable maintenance of\nmachinery. However, deep learning-driven anomaly detection methods often\nrequire an extensive amount of computational resources which prohibits their\ndeployment in factories. Here we explore a machine-driven design exploration\nstrategy to create OutlierNets, a family of highly compact deep convolutional\nautoencoder network architectures featuring as few as 686 parameters, model\nsizes as small as 2.7 KB, and as low as 2.8 million FLOPs, with a detection\naccuracy matching or exceeding published architectures with as many as 4\nmillion parameters. Furthermore, CPU-accelerated latency experiments show that\nthe OutlierNet architectures can achieve as much as 21x lower latency than\npublished networks.\n