ABSTRACT
Using convolutional neural networks (CNN) to discover sensor data patterns help predict upcoming failures in industrial machines. Traditionally, this pipeline is deployed on the cloud as deep neural networks have high computational requirements. Alternatively, an on-device deployment to make decisions locally for this pipeline could lower the energy requirements by not sending the sensor data back and forth to the cloud. However, this is challenging from an edge deployment perspective due to the limited computational resources and energy budget. To approach this issue, we propose a hierarchical architecture that leverages multiple smaller networks dividing the bigger problem into smaller sub-problems in a divide and conquer approach. With this architecture, we achieve a 9x reduction in energy consumption going from 0.045) per inference for a non-hierarchical CNN classifier to 0.0051 At the same time, our approach delivers low latency with comparable accuracy to the baseline, while running completely at the edge.
