Monitoring a process/phenomenon of specific interest at the network edge is prevalent in Cyber-Physical Systems (CPS), remote healthcare, smart buildings, intelligent transport, etc., that are essential building blocks of smart cities. Todays monitoring systems extensively use Internet-of-Things (IoT) sensors. In the era of the Edge Intelligence, there is a major research thrust for deploying small Machine Learning (ML) models on the IoT sensors making them capable of doing local inference on the collected data. The small ML models consume lower energy at the cost of lower inference accuracy compared to large ML models, namely, Deep Neural Netowrks (DNNs) that run on edge servers. In this context, there are several unanswered questions on the Total System Energy (TSE) consumption in the monitoring systems. A natural question is: where should the inference be performed for a data sample so that TSE is reduced? Another impending question is: when should the sensors smaple in order to further reduce the TSE? The latter question is inspired by the fact that in today’s system, the data collected sensors has high redundancy.

The GreenEdge project answers the above questions by exploring the TSE savings that can be achieved in a monitoring application using intelligent sampling and scheduling the inference between the edge server and the IoT sensors. GreenEdge will achieve this while respecting the applications’ Quality of Service (QoS) requirements. This will be conducted in three stages: (1) performing measurements of energy consumption, processing times, and communication times on the IoT sensor and the edge server, (2) establishing models and algorithmic solutions that schedule the sampling and the inference by exploiting the trade-offs between the TSE consumption, inference accuracy, IoT battery limitation, delay in detecting essential events etc., and (3) applying the new findings and validating the efficacy of the proposed algorithms in two exemplary applications with varied characteristics, namely, a cognitive assistance application and a wildfire monitoring testbed.

Project PID2021-128223OA-I00 funded by: