Lightweight Deep Learning-Based Receiver Design for Coded OTFS Modulation in Vehicular Networks

Vehicular networks have strong requirements for reliable communication in high-mobility environments. Orthogonal time frequency space (OTFS) modulation addresses the challenges of vehicular double-selective channels by mapping information symbols into a two-dimensional delay-Doppler (DD) domain, allowing for enhanced utilization of channel diversity across time and frequency. Traditionally, signal detection in OTFS receivers usually requires accurate channel state information (CSI) which is difficult to obtain. In this paper, we design an efficient intelligent receiver for coded OTFS in vehicular networks based on deep learning (DL), utilizing its power ability in two-dimensional data processing. Specifically, without perfect CSI, we use neural networks to directly extract features from signal data in the DD domain for OTFS signal detection. Moreover, in order to solve the problem of the limited reception field of convolution modules, we design a shift padding (SP) method to enhance data at the edge of DD domain data frames. Extensive simulation results validate the feasibility of the SP method and the effectiveness of the proposed intelligent receiver of coded OTFS, whose performance is considerably ahead of classic signal detection methods in low signal-to-noise ratios (SNR).