Energie-efficiënte Edge Computing

Abstract

The Internet of Things (IoT) is expected to reach up to 41 billion connected devices by 2030, driving a shift from centralised cloud computing toward edge computing, in which sensing, processing, and decision-making are performed closer to the data source. This shift reduces latency, bandwidth usage, and dependence on continuous connectivity, but it also exposes IoT devices, which are often battery-powered and deployed in remote, hard-to-maintain environments, where they are subject to severe constraints in energy, computation, and communication. This thesis addresses three research questions, each of which addresses these constraints: (1) how a low-power, real-time middleware platform can coordinate heterogeneous sensors and actuators; (2) how data transmission can be minimised while preserving the accuracy of the transmitted information; and (3) how edge intelligence can be realised on sustainable, battery-free devices while still supporting complex tasks. To answer these questions, this thesis makes three main contributions. First, it presents a hardware/software platform for low-power multimedia sensing and edge intelligence over cellular networks that is built on a Linux-capable System-on-Chip (SoC) with an integrated LTE-M/NB-IoT modem and a middleware layer that manages heterogeneous sensors and actuators, handles uplink/downlink networking and synchronisation, and controls low-power sleep modes. Second, building on the platform, it proposes a pipeline combining edge data processing, low-power wireless transmission, and backend model adaptation, validated on two industrial audio-based fault detection case studies, where it achieves substantial energy savings in data transmission while preserving accuracy close to an uncompressed baseline. Third, it extends the platform into what is, to the best of our knowledge, the first battery-free edge intelligence platform capable of reliably executing standard deep learning (DL) models in the field through solar energy harvesting, supercapacitor-based energy storage, and energy-aware task scheduling. Together, these contributions show how low-power middleware design, communication-efficient data pipelines, and sustainable hardware/software co-design can enable practical, long-lived, and intelligent IoT deployments in resource-constrained and remote environments, validated through real-world industrial and environmental field studies.