Deep Learning Beamspace Channel Estimation for mmWave Massive MIMO with Switch-Based Selection Network

In this paper, we introduce a deep learning-based beamspace channel estimation approach that better exploits the inherent sparsity of the mmWave MIMO channel. On one hand, we replace conventional complicated phase shifter networks with switch-based selection networks, whose sparse connectivity is more adapted to the sparsity of mmWave channels. On the other hand, we propose an attention-Unet model for accurate beamspace channel estimation. The architecture comprises an encoder-decoder structure with attention mechanism. By selectively focusing on the dominant part, the attention mechanism can further capture the sparsity of the beamspace channel. Simulation results demonstrate that the proposed approach outperforms the existing phase shifter-based techniques under both the widely used Saleh-Valenzuela channel model and the open-source DeepMIMO dataset based on ray-tracing.