International Journal of Science, Engineering and TechnologyAuthors: Dr. M. Mahil, S. Sindhu
Abstract: Industrial Internet of Things (IIoT) environments consist of a large number of heterogeneous edge devices with varying computational capabilities, energy levels, and communication conditions. Traditional Federated Learning (FL) approaches typically employ random or static client selection strategies, which often lead to excessive energy consumption, increased communication overhead, and slow model convergence. To address these challenges, this paper proposes a D4PG-based Energy-Aware Client Selection framework for Federated Learning in Industrial IoT systems. The proposed approach models client participation as a sequential decision-making problem and utilizes Deep Distributed Distributional Deterministic Policy Gradient (D4PG) reinforcement learning to dynamically select clients based on battery level, CPU utilization, bandwidth availability, latency, and expected training contribution. A multi-objective reward function is designed to jointly optimize learning accuracy, energy efficiency, and communication cost. The framework is evaluated in a heterogeneous non-IID federated environment consisting of 20–50 edge clients. Experimental results demonstrate that the proposed D4PG-based strategy improves convergence speed, reduces communication overhead, and lowers energy consumption compared with conventional client select Federated Learning ion methods. The findings highlight the effectiveness of distributional reinforcement learning for intelligent resource-aware federated optimization in Industrial IoT deployments.
