MEC-Enabled Hierarchical Federated Learning for Resource-Aware Device Selection in IIoT

Hierarchical federated learning (HFL) combined with the Mobile Edge Computing (MEC) paradigm has attracted extensive research interest in the Industrial Internet of Things (IIoT) due to its ability to deploy computational resources near edge devices and effectively reduce communication overhead. However, in real-world applications, the dynamic participation of edge devices and their diverse training objectives can lead to instability in model convergence, affecting overall system performance. To address this challenge, this paper proposes a device selection strategy based on task completion probability to determine participating devices dynamically in each training round. Furthermore, to balance system resource consumption and model performance, we formulate an optimization objective to minimize the loss function under resource constraints. By leveraging theoretical analysis, we reformulate the objective as a loss upper bound minimization problem related to resource allocation, which is subsequently decomposed into multiple subproblems for iterative solving. Simulation results demonstrate that the proposed method achieves superior resource efficiency and training stability. Compared to the state-of-the-art HFL method, DSRA-HFL reduces the average training delay by approximately 18% and energy consumption by 22% under dynamic conditions, while maintaining a competitive model accuracy. This validates the effectiveness of our joint optimization strategy in practical IIoT scenarios.