Channel Estimation Based on Model-Driven Residual Networks for Zero-Padded OTFS

Abstract

Orthogonal time frequency space (OTFS) has been a potential option for high-mobility communication scenarios due to its strong Doppler resilience. In our previous work, considering the practical rectangular pulse shaping and fractional Doppler shifts, a novel pilot pattern for Zero-Padded OTFS channel estimation was proposed, which can significantly reduce the pilot power of the embedded pilot scheme based on compressed sensing. To further improve the accuracy of the channel estimation, this paper presents a simplified model-driven residual network (ResNet) to refine the initial results. The proposed ResNet has a smaller input size and thus, lower training complexity by leveraging the beneficial features of the delay-Doppler (DD) domain channel. Simulation results demonstrate that the proposed simplified ResNet outperforms the preliminary estimator in terms of normalized mean square error (NMSE) performance. And the former can obtain more than 5 dB and 4.5 dB gain over the latter in the embedded pilot case and our proposed pilot pattern, respectively. The results also show that the channel estimation of the proposed ResNet can improve signal detection performance.