Abstract:
Base station (BS) selection is important in establishing reliable communication links in millimeter-wave (mmWave) systems. The selection procedure typically requires each BS to perform a handshake with user equipment (UE), which becomes increasingly complex as the number of BSs and UEs grows, leading to significant communication overhead. Our article investigates the selection of BS for autonomous vehicles (AV) using deep learning (DL) with light detection and ranging (LiDAR). Two federated learning (FL) approaches are implemented to reduce the communication overhead, where the BS broadcasts its parameters to all connected nodes. A comprehensive comparison of these FL variants is presented. Additionally, the article describes the generation of datasets by ray-tracing approaches and LiDAR data. To evaluate the robustness and generalizability of the proposed model, transfer learning (TL) is employed across different scenarios, such as varying the number of training samples, the city used for simulation, and the number of BSs involved. Simulation results demonstrate that our proposed approach outperforms traditional received signal strength indicator (RSSI)-based BS selection by a factor of 1.9 in accuracy while achieving a remarkable 96.38% reduction in data size with FL-based BS selection. Our proposal significantly minimizes communication overhead, validating the efficiency. © 1967-2012 IEEE.
