Residual Learning based Channel Estimation for OTFS system

Abstract

Orthogonal time frequency space (OTFS) systems can effectively balance the Doppler shift by transforming the channel with a drastic change in the time-frequency (TF) domain into a stable channel in the delay-Doppler (DD) domain. In order to take full advantage of the OTFS system, accurate channel estimation results are critical in OTFS systems. In this paper, a model-driven deep learning (DL)-based channel estimation technique is proposed for OTFS in the DD domain. The presented channel estimation scheme has two parts. The first part takes advantage of the traditional orthogonal matching pursuit (OMP) algorithm to generate preliminary channel estimation results. The second part uses a deep residual learning network (ResNet) to further process the rough estimation results to get an accurate OTFS channel estimation. Simulation results demonstrate that the performance of the proposed model-driven ResNet-based scheme is significantly better than the traditional OMP algorithm, and there is about 6dB performance gain when the size of an OTFS frame is 128×16 and the normalized mean squared error (NMSE) is 0.00173. It also proves that the proposed ResNet-based channel estimation scheme can be applied to different scenarios and achieve good robustness.