Low power on-line machine monitoring at the edge
Bert Boons, Marian Verhelst, Peter Karsmakers · 2021
On-line machine monitoring enables more efficient maintenance, by continuous analysis of machine parameters (e.g., vibration, sound, temperature, ...) to enable early identification of malfunction. This task is best implemented as an anomaly detection scheme, as no or limited data is available regarding malfunctioning conditions. This scheme should moreover be integrated on the sensory edge device, to avoid the burden of centralizing the analysis of the raw sensor data on the compute and network infrastructure. In this work we therefore present a workflow for training a quantized deep learning anomaly detection model (Deep SVDD), suitable for edge devices with limited resources (e.g. memory, compute, power). We show that using minimal architectural interventions we achieve a reduction of 50% in terms computational resources when compared to autoencoder based approaches. Our final models run at 4 bits without loss of detection performance in terms AUC.