In this paper, we present a new beam selection approach for the beamspace channel estimation (CE) in 64 antennas Massive Multiple-Input Multiple-Output (MIMO) receiver. Usually, the beamspace CE is implemented via digital transformation of antenna signal to a priori selected sub-space of discrete Fourier transform (DFT) directed towards propagation channel taps. This results in less complexity of CE and MIMO detector units. We propose a new data-based sub-space selection method, which outperforms the DFT-based beam selection thanks to employing prior knowledge of channel tap distribution in the spatial domain. Simulation results are presented for the non-line-of-sight models of the 5G QuaDRiGa 2.0 channel.