Resource-constrained Intelligent Edge Systems for Internet-of-Things Applications

Miniature Internet-of-Things (IoT) systems can be uniquely useful in many applications including wearables and environmental monitoring due to their ability to obtain and analyze data directly at the source. These devices use millimeter-sized batteries, resulting in a limited energy budget that constrains intelligent edge-computing, memory capacity, and wireless communication.  This thesis introduces techniques at the algorithmic, hardware and system level that tackle these challenges, and demonstrates three resource-constrained IoT systems.

The first work is a 6.7 x 7 x 5mm ultra-low-power imaging system with deep learning and image processing capabilities for intelligent edge monitoring. It leverages data and energy management techniques to maintain less than 50µW average power, and proposes image-enhancement layers to improve neural network performance in constrained embedded systems. The second work presents an OFDMA baseband localization processor fabricated in 22nm, co-designed with a low-power crystal-less narrowband RF receiver, that efficiently estimates the channel frequency response in real-time for IoT localization. This system is deployed in a cm-scale localization tag that achieves 4.3X longer distance and 6.6X lower power than the state-of-the-art. The final work demonstrates an 8.5 x 20mm tag that integrates the proposed baseband localization processor with a suite of other low-power ICs and custom antenna, to enable insect tracking.

CO-CHAIRS: Professor David Blaauw and Professor Hun-Seok Kim