Intelligent Proactive Edge Health Monitoring for Resource-Constrained Devices

Abstract

The rapid expansion of edge computing and the Internet of Things (IoT) has transformed the deployment of intelligent applications by shifting computation from centralized cloud infrastructures to resource-constrained edge devices. Although this paradigm enables low-latency processing, reduced bandwidth consumption, and improved data privacy, it also introduces significant challenges related to limited computational resources, thermal stress, and device reliability. Traditional health monitoring techniques remain largely reactive, detecting failures only after performance degradation has already occurred. Consequently, there is an increasing need for intelligent, proactive approaches capable of anticipating system degradation before failures affect device operation. This thesis addresses this challenge by proposing an intelligent proactive health monitoring framework for resource-constrained edge devices. The proposed framework combines continuous hardware monitoring, machine learning-based forecasting, and Large Language Model (LLM)-driven orchestration to predict future device health and dynamically adapt monitoring strategies according to the current operational context. The proposed architecture continuously collects multiple hardware and system-level metrics, including CPU utilization, processor temperature, memory usage, storage activity, network performance, power conditions, and thermal throttling indicators. These heterogeneous measurements are transformed into semantic health states that provide an interpretable representation of the device's operational condition. To predict future health states, several forecasting models—including Random Forest, XGBoost, LightGBM, and Long Short-Term Memory (LSTM) networks—are trained and evaluated using representative workload scenarios. Their predictive performance, computational overhead, and suitability for resource-constrained environments are systematically compared. Building upon these forecasting capabilities, the framework introduces an LLM-based orchestration layer that dynamically selects the most appropriate application-level forecasting model according to the predicted health state, available computational resources, and application requirements. Unlike conventional static approaches that rely on a single forecasting model, the proposed adaptive orchestration mechanism balances prediction accuracy with computational efficiency while reducing unnecessary resource consumption. The framework is implemented and experimentally evaluated on Raspberry Pi 5 edge devices under diverse workload conditions representative of real-world edge computing environments. The experimental evaluation demonstrates the effectiveness of semantic health representation for interpreting device conditions, compares the forecasting accuracy and efficiency of the candidate machine learning models, and validates the benefits of LLM-driven adaptive model selection. The results indicate that intelligent orchestration enables more reliable and resource-efficient health monitoring by selecting forecasting models according to changing operating conditions while maintaining high prediction performance. Overall, this thesis contributes to the fields of edge intelligence, predictive maintenance, and sustainable computing by presenting a context-aware health monitoring framework that integrates predictive analytics with intelligent decision-making. The proposed approach demonstrates how combining machine learning forecasting with LLM-based orchestration can improve the reliability, efficiency, and longevity of resource-constrained edge devices, providing a foundation for future adaptive health management systems in edge computing environments.