Near-Field Wireless-Power-Transfer-Assisted Federated Learning IoT Networks

In this article, we investigate a near-field wireless power transfer (WPT) assisted federated learning (FL) Internet of Things (IoT) network, where multiple near-field IoT users jointly train a global model with the help of one hybrid access point (HAP). We consider practical communication scenarios where the users with limited battery are unable to support the local training and model upload within a given latency, and hence the near-field WPT is applied into the FL in a time-division multiplex (TDM) mode to harvest energy to the users from the HAP. By taking into account the joint constraints from both the latency and energy consumption, we optimize the training performance of the near-field WPT-FL IoT network by maximizing the number of users successfully participating in the FL training, through jointly exploiting the system temporal, power and bandwidth resources. Specifically, when the instantaneous channel state information (I-CSI) of wireless links in the system is known, we optimize the near-field WPT duration ratio, RF transmit power and wireless bandwidth among users to maximize the instantaneous number of users successfully participating in the FL training, through jointly using the deep Q-learning network (DQN) and greedy method. When only statistical CSI (S-CSI) is known, we optimize the system resources to maximize the expected number of users successfully participating in the FL training, through jointly using the DQN and golden ratio method. Simulation results are provided to demonstrate that our I-CSI-and S-CSI-based resource allocation (RA) schemes are superior to other competing schemes in both test accuracy and convergence rate.