Abstract:
The rapid proliferation of the Internet of Things (IoT) has transformed healthcare by enabling continuous monitoring through wearable devices, yet challenges remain in achieving reliable, low-power communication persist, particularly in mobile scenarios. Long Range Wide Area Network (LoRaWAN), a prominent Low-Power Wide-Area Network (LPWAN) technology, offers extended range and energy efficiency, but traditional Adaptive Data Rate (ADR) mechanisms exhibit inefficiencies, leading to packet loss and high energy consumption. This thesis introduces a lightweight Tiny Machine Learning (TinyML) framework that enhances resource allocation in LoRaWAN for health monitoring applications. Building upon the AI-ERA approach, a previously developed AI-based resource allocation method, the framework employs an ensemble model integrating Support Vector Classifier (SVC), XGBoost, and Deep Neural Network (DNN), with Out-of-Fold (OOF) predictions (predictions for each data point made by a model that was not trained on that point) and a meta-classifier (Logistic Regression), achieving 86% accuracy in optimal Spreading Factor (SF) prediction. Using NS-3 simulations, the framework demonstrates up to 32% improvement in Packet Success Ratio (PSR) for static scenarios and 28% for mobile ones compared to standard ADR, alongside reduced energy consumption and reducing convergence time (to optimal SF allocation) to 6 hours. Deployable on constrained devices, it enables real-time adaptability in wearables while addressing privacy, latency, and power constraints in health IoT. This work advances scalable health monitoring systems through proactive, edge-based intelligence, with potential extensions to Federated Learning for enhanced privacy. Keywords: TinyML, LoRaWAN, Ensemble Learning, Resource Allocation, Health Monitoring, IoT.
