Model-Driven Channel Estimation Network for Orthogonal Time-Frequency Space Systems

Orthogonal time-frequency space (OTFS) modulation has garnered significant attention in wireless communications, particularly in high-mobility scenarios, due to its exceptional resilience to Doppler shifts. While the delay-Doppler domain channel estimation is mostly considered for reliable communication, it is often hindered by fractional Doppler shifts in complex channel conditions. To overcome this challenge, this paper proposes a model-driven channel estimation network (CENet) for accurate channel estimation in the delay-time (DT) domain, with the initial least-squares (LS) solutions as prior information for training. Specifically, a LS-driven CENet with parallel structure is proposed to efficiently handle the high-dimensional DT domain channel matrix, which enables the row-wise training and estimation with fast convergence. Simulation results demonstrate that our proposed CENet significantly outperforms the conventional LS and deep neural network-based methods.