Deep Learning with a Self-Adaptive Threshold for OTFS Channel Estimation

The recently developed orthogonal time frequency space (OTFS) technology has proved its capability to cope with the fast time-varying channels in high-mobility scenarios. In particular, the channel model in the delay-Doppler (DD) domain has a sparse representation, and its associated channel estimation can be realized by adopting one embedded pilot scheme. However, it may face performance degradation in scenarios with unknown and burst noise. In this paper, we develop a deep learning (DL)-based method to deal with complicated noise. In particular, we consider the sparsity of the OTFS channel and propose a deep residual shrinkage network (DRSN) to implicitly learn the residual noise for recovering the channel information. In addition, to further improve the channel estimation accuracy, we adopt a self-adaptive threshold to eliminate the irrelevant features to ensure channel sparsity. Simulation results verify the effectiveness of our proposed DRSN-based approach in complicated noise scenarios.