Environment-Specific Beam Training for Extremely Large-Scale MIMO Systems via Contrastive Learning
Xiangyu Zhang, Haiyang Zhang, Chunguo Li, Yongming Huang, Lüxi Yang
Abstract
Extremely large-scale massive multiple-input multiple-output (XL-MIMO) systems are poised to be a crucial technology in the 6G era. However, beam training (BT) in XL-MIMO encounters high training overhead due to near-field channel effects. To address this issue, we propose a two-stage BT scheme coupled with a two-level codebook that capitalizes on the spatial non-stationarity property. Specifically, the spatial non-stationarity property is identified by the vision region (VR), that a sub-array only sees a portion of the communication region. Based on this property, our BT scheme utilizes the high-level codebook to determine the users’ VR and the corresponding sub-array in the first BT stage. In the second BT stage, we employ the low-level codebook to align the beam inside the VR. To accommodate this BT process, our codebook is designed as an environment-specific codebook that includes spatial non-stationarity information, which depends on the scatterer and obstacle of the propagation environment. For designing this codebook, we introduce an efficient contrastive learning framework, which achieves higher beam gain with fewer codewords by learning a more appropriate spatial partition. Simulation results demonstrate that the proposed method attains high beam alignment accuracy and spectral efficiency while maintaining a compact dictionary and reducing BT overhead.