Deep transfer learning-based adaptive beamforming for realistic communication systems

Abstract

To support high data rate/large capacity in next-generation communication, beamforming techniques on a massive multiple-input multiple-output (MIMO) system are actively proposed. For designing beamformer, however, a downlink channel state information (CSI) is required. Estimation of the downlink CSI in the massive MIMO system is a challenging task due to an increase in overhead according to the channel size. Therefore, most studies assume a time division duplexing (TDD) system in which the uplink and downlink channels are the same. However, in realistic communication system, non-reciprocity of uplink and downlink channels occurs even in the TDD system due to the influence of hardware such as an radio frequency (RF) chain. Recently, deep learning (DL)-based channel estimation and beamforming techniques have been proposed, thanks to the advantage of being applicable regardless of the type of system and channel. However, the DL-based approach requires the large number of training data for performance above a certain level. In this paper, we aim to train the deep neural network (DNN) for designing the beamformer, and introduce deep transfer learning (DTL) using information from pre-trained DNN to reduce the number of required training data. When DTL-based training method is applied, it is possible to train the DNN to properly adapt to the surrounding channel environment with a small number of channel data. Also, through the simulation, the performance of the DTL-based beamforming in various conditions and systems is confirmed. And the factors affecting the performance of the DTL-based beamforming are analyzed. Furthermore, we propose `step-by-step' DTL as the extended DTL-based method that can flexibly respond to the uncertainties of realistic communication environment. The simulation results show that DTL-based DNN training method not only improves the performance by considering the non-reciprocity between uplink and downlink channels regardless of the system and channel type, but also significantly reduces the number of training data for adapting to realistic hardware influence of a specific base station (BS).