Publication

Abstract : Piezoelectric (PE) energy-harvesting IoT devices enable autonomous sensing but face major challenges in large-scale cloud environments due to unpredictable harvested energy, fluctuating workloads, and limited resources. This study proposes a hybrid learning architecture that integrates convolutional neural networks (CNNs) with a Double Deep Q-Learning (DDQL) agent to achieve efficient task scheduling under these constraints. The CNN module extracts high-level spatial–temporal patterns from telemetry and vibration-based piezoelectric sensor traces to estimate task complexity and energy requirements. IoT nodes powered by piezoelectric harvesters generate sensing, preprocessing, or data-upload tasks, requiring a scheduler capable of optimizing latency, success rate, and energy usage. The cloud-edge scheduler decides task offloading, resource allocation, and execution order. CNN-derived features and system state are fed into a DDQL agent, which selects actions such as local execution, offloading, or delaying, while mitigating over-estimation bias found in standard DQN. The model incorporates battery behavior through a stochastic transduction and Markov Decision Process (MDP) framework. Experiments on a MATLAB/Simulink and Python-based digital twin show that the CNN-DDQL system reduces latency by 21–34%, energy consumption by 15–28%, and improves deadline-meeting rates by 17–26%. Results confirm the framework’s robustness in low-energy and unstable network conditions, making it highly suitable for self-powered IoT systems.