Machine Learning-Assisted Channel Estimation in Massive MIMO Receiver

We propose a new Machine Learning (ML) approach to channel estimation (CE) in Massive Multiple Input Multiple Output (MIMO) receivers. The algorithm employs a recurrent neural network (RNN) for iterative channel tap search and nonlinear de-noising of their amplitudes in the time domain. Our method outperforms the sparse minimum mean square error (MMSE) CE that is based on time-domain windowing and a further linear denoising of channel taps within the window. Simulation results are presented for user speed of 5km/h in non-line-of-sight scenarios of the 5G QuaDRiGa 2.0 channel. The results are provided in both antenna and beamspace domains of the 64 antennas receiver.