On Environmental Learning in Resource-Constrained Edge Systems
Guardado en:
| Publicado en: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Autor principal: | |
| Publicado: |
ProQuest Dissertations & Theses
|
| Materias: | |
| Acceso en línea: | Citation/Abstract Full Text - PDF |
| Etiquetas: |
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Resumen: | Deploying resource-constrained edge systems in real-world environments presents challenges beyond those encountered in controlled testbeds. Real-world conditions are dynamic and unpredictable, shaped by numerous uncontrollable factors such as building structures, noise, weather, and traffic, all of which can significantly impact system performance. Strict resource constraints further complicate the situation. To address these challenges, machine learning with edge systems offers a promising solution. However, traditional approaches relying on centralized servers are impractical due to resource and communication limitations.Edge systems can leverage machine learning in two primary ways: offline training with online inference and on-device training with inference. The offline approach involves training models on powerful servers and deploying them to edge devices for inference, making it well-suited for resource-constrained environments. In contrast, the on-device approach performs both training and inference directly on the edge device, eliminating the need for centralized communication. This method is particularly useful in scenarios with data privacy concerns or limited connectivity, enabling local adaptation and resource-efficient solutions.This dissertation introduces a comprehensive framework to enable edge systems to autonomously learn and adapt to their environments by integrating advancements in hardware, software, network optimization, and sensing technologies. At the core of this research is MKII (“Mach 2”), a low-power acoustic edge device designed for resource-constrained scenarios. MKII features an ARM Cortex-M7 and M3 architecture, an ultra-low-power runtime, and long-range LoRa multi-hop networking, allowing it to operate independently of infrastructure such as Wi-Fi. It supports both edge learning approaches and has been successfully deployed in diverse environments, including New York City, airport terminals, and campus-scale monitoring systems.To address communication challenges in urban environments, this dissertation proposes innovative LoRa-based networking solutions. These include a frequency selection method that combines offline and online training, as well as a reinforcement learning-based dynamic frequency allocation approach for online training. These methods help mitigate interference and reduce energy consumption, and have been validated through real-world experiments, demonstrating improved reliability and efficiency in long-range wireless communication.Lastly, this dissertation addresses domain adaptation in multi-label sound classification. A novel unsupervised domain adaptation framework for edge system deployments is introduced, eliminating the need for labeled deployment data. The framework incorporates class-specific adaptive thresholds to improve pseudo-label accuracy and uses diversity regularization to enhance model robustness. Designed for resource-constrained devices, it performs on-device inference and triggers selective retraining based on high-confidence data, minimizing computational overhead while maintaining high classification accuracy. Extensive evaluations using the SONYC-UST dataset confirm the effectiveness of the proposed methods in real-world deployment scenarios. |
|---|---|
| ISBN: | 9798314890967 |
| Fuente: | ProQuest Dissertations & Theses Global |