Learning-Aided Beam Training for Extremely Large-Scale MIMO With Spatial Non-Stationarity

We propose a beam training scheme to achieve low overhead and precise beam alignment in extremely large-scale multiple-input multiple-output (XL-MIMO) systems with near-field and spatial non-stationarity (SnS). First, the deep learning-based beam training for near-field SnS (DBNS) method is developed. Considering that the sub-channels differ for each subarray due to SnS, a distinct codeword index is obtained for each subarray. Inspired by the weakened near-field and SnS characteristics of subarrays with fewer antennas than the full array, the beam training overhead is reduced by leveraging the received signals from the subarray-wise far-field codebook. Second, a visible region-adapted alternating beam refinement (VABR) algorithm is introduced. Using the outputs of the DBNS method as inputs, the VABR algorithm enables the acquisition of off-grid estimates for both angle and distance, while accounting for SnS within each subarray. Numerical results demonstrate that the proposed beam training methods effectively address SnS, achieving high beam gain and prediction accuracy while significantly reducing the beam training overhead in near-field environments.