Federated Learning-Based SB-SPS Scheduling for 5G NR-V2X Under Heterogeneous Vehicular Environments

Abstract

This study proposes a federated-learning-assisted sensing-based semi-persistent scheduling (SB-SPS) framework for 5G NR-V2X sidelink communication in dense and heterogeneous highway environments. Although the conventional SB-SPS enables distributed resource allocation without centralized coordination, its static reservation and reselection behavior can lead to persistent collisions, poor packet delivery, and reduced throughput under dynamic vehicular traffic conditions. To address this limitation, the proposed method integrates a lightweight collision-risk predictor into the SB-SPS reselection loop and trains the predictor collaboratively across vehicles using federated learning, without sharing raw local data. The framework is evaluated in two SUMO-based highway scenarios, highway_high and highway_mixed_hetero, and compared against the legacy SB-SPS baseline as well as FedAvg and FedProx variants with $\mu = 0.6, 1.0$ , and 1.4. The results show that the federated methods improve the reliability and throughput over the conventional SB-SPS in both scenarios. In highway_high, FedProx( $\mu =0.6$ ) improves packet delivery ratio from 0.79% to 70.21% and throughput from 0.003 to 0.253 pkt/s, while maintaining a mean delay of approximately 10 ms. In highway_mixed_hetero, the same method improves the packet delivery ratio from 24.29% to 72.92% and throughput from 0.078 to 0.233 pkt/s, again with a mean delay of approximately 10 ms. Across the evaluated settings, the federated schedulers consistently reduce collision levels relative to the baseline SB-SPS, whereas FedProx provides the strongest overall trade-off under heterogeneous traffic conditions. These findings indicate that federated learning is a promising approach for collision-aware and privacy-preserving enhancement of distributed NR-V2X sidelink scheduling.