SERN-BT Algorithm for Beam Training in Extremely Large-Scale MIMO System
Abstract
Considering the problem that traditional near-field beam training requires joint angle-distance optimization, which increases the pilot overhead as the antenna size grows. To reduce the pilot overhead while maintaining achievable rate performance, an efficient beam training (BT) algorithm is proposed based on Residual Network with Squeeze-and-Excitation (SERN) in this paper, named as SERN-BT algorithm. Inspired by the hierarchical codebook, the SERN network utilizes received signals from far-field wide beams to infer the optimal near-field codeword. The strategy results in a substantial decrease in pilot overhead when contrasted with traditional narrow beam sweeping. Furthermore, the SERN framework incorporates a dual-branch structure to decouple the nonlinear mapping of angle and distance. Within the SE-ResNet blocks, the squeeze-and-excitation mechanism is innovatively integrated to adaptively recalibrate channel weights. The mechanism enables the network to suppress noise interference and selectively emphasize the sparse spatial features inherent in wide-beam signals. Finally, the well-trained SERN model is capable of rapidly inferring the near-optimal near-field codeword. Simulation results demonstrate that the presented algorithm can dramatically reduce the pilot overhead while sustaining achievable rate in the near-field domain.