Multi-cell Air-ground Federated Learning with New Samples under Wireless Packet Error

Unmanned aerial vehicles (UAVs) have emerged as pivotal platforms for edge intelligence (EI) and federated learning (FL) further enhances UAV-enabled EI by enabling decentralized model training across distributed clients. Existing FL frameworks face critical limitations, including restricted network coverage in single-server systems, unreliable air-ground communication channels, and inefficiencies in handling new samples. This paper proposes the multi-cell air-ground FL (MAGFL) with new samples under packet error, addressing these gaps. UAVs act as wireless clients, selecting samples without pretraining to minimize computational overhead. Terrestrial servers optimize UAV-server associations to mitigate packet error rate (PER) impacts. A joint optimization problem is formulated to simultaneously refine sample selection and association strategies. The framework analytically quantifies wireless errors’ effects on FL convergence, derives a closed-form expression for the expected convergence rate, and reformulates the problem as a mixedinteger nonlinear program (MINLP). The proposed algorithm solves this MINLP with guaranteed convergence and extensive simulations validate MAGFL’s superiority in dynamic datasets, demonstrating robustness to PER.