Deep Learning-based Channel Estimation for Massive MIMO-OTFS Communication Systems

This paper introduces a deep learning (DL)-enabled framework for channel estimation of high mobility massive multiple-input multiple-output orthogonal time frequency space (MIMO-OTFS) wireless cellular networks. By modulating data in the delay-Doppler domain, OFTS is able to transform a frequency-selective time-varying fading channel into a quasi-time-invariant channel. Employing OTFS can overcome challenges that traditional modulation schemes encounter in high-mobility applications, and consequently can significantly improve the quality of wireless transmission. To realize these benefits, OTFS systems require accurate channel estimation, which becomes challenging as the number of antennas grows. Channel estimation for OTFS is formulated as a sparse signal recovery problem, and is solved by a novel deep network design consisting of the following three convolutional neural networks (CNN): (i) based on spatial features of doubly-selective fading channels, PositionNet is designed to find the indices of non-zero elements (support) in the sparse channel matrix, (ii) PositionNetr then refines the solution, and (iii) AmplitudeNet obtains the values of the non-zero elements. This approach provides improvement in bit error rate (BER) and normalized mean squared error (NMSE) as well as significant reduction of 80% in computation. Simulation comparisons demonstrate the merits of the proposed approach.