Deep Learning-Based Downlink OTFS-SCMA

Sparse code multiple access (SCMA) is a scheme with high potential for application to orthogonal time frequency space (OTFS) systems. However, errors arising during OTFS detection propagate to the SCMA decoder, resulting in performance loss. In this letter, a deep learning (DL)-based scheme is proposed for downlink OTFS-SCMA system. The proposed scheme includes an autoencoder (AE)-based SCMA codebook mapping and a multi-layer perceptron (MLP)-based SCMA decoder. Considering that SCMA codewords are independently identically distributed one-dimensional vectors, the designed MLP can efficiently decode all SCMA codewords of an OTFS frame simultaneously, which effectively reduces the latency. By jointly training and optimizing with end-to-end error as loss function, the proposed scheme is robust against error propagation. Simulation results indicate that the proposed DL-based scheme achieves a gain of up to 5 dB compared to traditional schemes. Moreover, the proposed scheme shows significant robustness to channel state information (CSI) compared to the cross-domain schemes while significantly reducing complexity.