Active Sensing for Beam Management: A Deep-Learning Approach

Abstract

Millimeter wave (mmWave) systems rely on predefined codebooks for both initial access and data transmission. To compensate the high pathloss of mmWave signal, base station(BS) and user equipment(UE) to be equipped with large antenna arrays which make those codebooks consist of a large number of candidate narrow beams. Both the BS and UE needs to search for a optimal beam from their codebooks that provides maximum received power, such procedure may cause huge beam training overhead. Besides, codebook based beam management limits the maximum beamforming gain as it is bounded by the spatial granularity of the codewords. To overcome these limitations, in the paper, we design a deep learning (DL) based beam training method with partial codebook sweeping. Unlike the existing works using machine learning (ML) or DL to predict the best beam ID from the codebook, the DL model directly outputs the beamforming weights of the analog phase shifters which maximize certain metric, e.g. received signal to noise ratio (SNR). The neural network (NN) is trained offline using simulated environments according to the 3GPP channel models and is then deployed online to predict the optimal beamforming vector with partial beams sensing. Simulation results show that our proposed model outperforms the standard DFT-based codebook with significantly reduced beam training overhead, and enhance the beamforming gain which reflects on the achievable rates.