Deep Residual Attention Network for OTFS Channel Estimation

Abstract

Orthogonal time frequency space (OTFS) has shown its potential in dealing with fast time-varying channels under high-mobility scenarios. The embedded pilot frame reduces pilot overhead to achieve higher spectral efficiency but introduces interference between data and pilot. To reduce interference between data and pilot, a deep learning based residual attention network (DRAN) is proposed in this paper. The DRAN consists of four modules, i.e., the residual module, the channel attention module (CAM), the position attention module (PAM) and the attention feature fusion (AFF) module. The residual module transforms the received signal into a series of domains. The CAM re-weighted information from these domains to extract the pilot information, which could be regarded as a filter in the delay-Doppler (DD) domain. And the PAM is further used to reduce the fluctuation in the filter. Finally, the AFF is used to fuse the results from CAM and PAM. The proposed framework extracts the pilot from the signal and reduces the interference from data and noise. Simulation results show that the proposed DRAN achieves higher accuracy compared with the previous OTFS channel estimation methods.