Semantic-Value-Aware Federated Learning for Energy-Harvesting AIoT Systems

초록

Energy-harvesting AIoT (EH-AIoT) systems enable sustainable edge intelligence under intermittent and time-varying energy supply. Federated learning (FL) has been widely adopted in such environments; however, existing approaches primarily focus on energy-aware control or communication reliability, implicitly assuming that all local updates are equally useful. In practice, many updates provide only a limited contribution to global learning, leading to inefficient energy utilization. In this paper, we propose a semantic-value-aware federated learning framework for EH-AIoT systems, where update transmission is determined based on both energy feasibility and its learning contribution. We define semantic update value as an approximation of the expected global learning improvement and estimate it using task urgency, gradient novelty, forgetting risk, and distribution shift. Based on this formulation, we develop a unified control framework that integrates energy stability, learning performance, and update significance using a Lyapunov-based objective. A lightweight policy is designed to adaptively control local training intensity, rehearsal ratio, and transmission decisions. Experimental results under non-IID data distributions and realistic solar energy traces show that the proposed method reduces unnecessary transmissions while maintaining or improving accuracy. It also achieves lower blackout rates and improved energy efficiency compared to existing approaches. These results demonstrate that semantic-value-aware update control provides an effective solution for resource-efficient federated learning in EH-AIoT systems.

키워드