Joint channel extrapolation and LoS/NLoS identification for UAV to ground communications: A multi-task learning approach

With the advent of the sixth-generation (6G) mobile networks, UAV has been increasingly important, acquiring accurate channel state information (CSI) with low overhead becomes a prior to improve the performance of mobile networks. In addition, the identification of line of sight (LoS) and non-LoS also plays a vital role in for UAV-to-ground communication. Current research resolves the above problems separately without considering the potential of utilizing the common representations of the above tasks. Inspired by the multi-task learning (MTL) in the field of deep learning, we proposed CEI-MTL, a multi-task learning framework based on Transformer to perform time-antenna-domain channel extrapolation and LoS/NLoS identification simultaneously. To effectively improve the performance of the the proposed multi-task learning framework, we carefully design the weight of combining the loss function of channel extrapolation. Simulation results demonstrates the effectiveness of adopted weight value. Extensive simulation show that the proposed multi-task learning framework can effectively channel extrapolation and LoS/NLoS identification simultaneously. Compared with the state-of-art single-task framework, the proposed framework can achieve comparable channel extrapolation or LoS/NLoS identification performance with increasing the inference speed. This indicates that the proposed model has great potential in practical mobile networks.