Edge-Intelligent Wireless Sensor Networks: A Federated Learning Framework for Energy-Aware Distributed Inference

Abstract

The internet of things (IoT) is also using sensor networks that are increasingly becoming wireless as well as incorporating edge computing to achieve distributed inference instead of processing the data through a centralised cloud computing service. While, the traditional federated learning (FL) models are not specifically optimised to operate within energy constraints characteristic of battery-operated sensor nodes, which results in untimely depletion of nodes and network saturation. This paper: this paper suggests a federated learning model that is energy- aware on edge-intelligent WSNs that optimises computation and communication overhead on joint distribution training. A mathematical model of energy consumption is formulated that models node-based computation and transmission costs and re-formulates the global learning optimization with hard constraints on the available energy. The adaptive mechanism of participation is proposed, where only those nodes that have enough residual energy and good channel conditions are eligible in contributing to every training round. Moreover, update schemes that are communication efficient are incorporated to minimise further the transmission overhead. Large-scale simulations show that the suggested framework might save as much as 28 % of the total energy usage as compared to typical FL with equivalent inference competence. The methodology also increases the network life by more than 35 % and convergence speed in non-homogeneous node environments. The above findings suggest that sustainable edge intelligence in costly wireless sensor networks needs an energy-conscious federated optimization.