Optimization algorithm of UAV-assisted federated learning resources based on wireless energy harvesting

Federated learning (FL) enhances data privacy and security by enabling multiple terminals to collaboratively train a global model without transferring data to a central location. To tackle the challenges of limited terrestrial communication resources and device energy in FL, a UAV-assisted federated learning and energy collection resource optimization algorithm is proposed. This algorithm harvests the working energy of FL terminals from radio frequency signals transmitted by the UAV via wireless energy collection. Considering energy causality and limited resources, we formulate a problem aimed at minimizing the completion time of FL training tasks. As this problem is NP-hard, we initially employ a greedy algorithm to optimize the UAV's position, then decompose it into three sub-problems: computing resources, power control, and bandwidth allocation. We derive and establish the optimization objective function for computing resources as convex, obtaining the expression for optimal computing resources. The Brent algorithm, based on golden section interpolation, iteratively solves the power distribution sub-problem, while the Lagrange-quasi-Newton algorithm optimizes bandwidth allocation for each user. Simulation results demonstrate that the proposed algorithm effectively reduces training completion time while maintaining training quality.