A novel distributed architecture incorporating deep learning and biased selection for vehicular communication mmWaves beamforming

Vehicle to Infrastructure (V2I) connectivity has historically relied on Dedicated Short Range Communication (DSRC) and more recently Cellular Vehicle to Everything (C-V2X). However, DSRC adoption has slowed due to high deployment costs, whereas C-V2X, limited to the 5.9 GHz sub 6 GHz band, provides modest data rates mainly suitable for safety critical messages. Emerging V2I services, such as high resolution sensor sharing and cooperative perception, demand multi gigabit throughput to transfer large volumes of data (4–10 GB) between vehicles and Mobile Edge Computing (MEC) servers, requirements exceeding the capacity of sub-6 GHz technologies. This study explores a novel distributed architecture utilizing a federated learning paradigm for optimizing mmWave beamforming processes in V2I communication systems. By leveraging multiple non-RF modality sensors (GPS and LiDAR) and deep learning models, this approach aims to enhance the global model's adaptability and reduce the sub-6 GHz channel usage. The proposed system uses client-biased selection strategies, including MaxLoss and Heuristic Multi-Arm Bandit, to train and update the global model, demonstrating significant improvements in convergence rates and overall performance. Simulation results using the Infocom FLASH dataset validate the framework's efficiency, highlighting its potential for real-world deployment in dynamic environments.