Resource-Adaptive Semantic Transmission and Client Scheduling for OFDM-Based V2X Communications

Area of Science:

• Vehicular communication systems
• Machine learning for intelligent transportation systems
• Wireless network resource management

Background:

• Orthogonal frequency-division multiplexing (OFDM) based vehicle-to-everything (V2X) uplink scheduling faces challenges with variable resource-block allocation impacting semantic encoder output.
• Conventional encoders produce fixed token counts, leading to data truncation and corrupted federated learning gradients, particularly for non-line-of-sight vehicles.
• Urban V2X deployments exhibit spatial patterns where vehicles at intersections have poor conditions but crucial data, yet are excluded by throughput-driven client selection.

Purpose of the Study:

• To develop a unified framework addressing token budget violations and biased client selection in OFDM-based V2X federated learning.
• To enhance the efficiency and accuracy of federated learning models for V2X safety applications by optimizing data transmission and client representation.

Main Methods:

• Implemented a Sensing-Guided Adaptive Modulation (SGAM) module for per-slot token budget derivation and Gumbel-TopK pruning with hard capacity clipping.
• Introduced a Channel-Decoupled Federated Learning (CDFL) module for independent client partitioning by channel quality and data complexity.
• Utilized facility location optimization for diverse client representative selection and inverse propensity weighting for partition-size imbalance correction.

Main Results:

• Achieved a Macro-F1 score of 0.710, an 8.7-point improvement over the baseline, on the NuScenes dataset with 20 non-IID vehicular clients.
• Demonstrated zero budget violations throughout training and a 75% reduction in training variance.
• Significantly improved worst-class F1 score, more than doubling it compared to FedAvg, indicating better performance for underrepresented clients.

Conclusions:

• The proposed framework effectively resolves token budget constraints and improves client selection fairness in OFDM-based V2X federated learning.
• The SGAM and CDFL modules enhance model accuracy, reduce training variance, and ensure critical safety data from challenging environments is utilized.
• This approach offers a robust solution for real-world V2X federated learning, particularly in complex urban scenarios.