Deep learning-based Collision-aware Multi-user Detection for Grant-free Sparse Code Multiple Access Systems

In grant-free sparse code multiple access (SCMA) systems, SCMA codebooks (CBs) are used for efficient grant-free random access. However, CB collisions can occur when multiple active users select the same CB, degrading the performance of multi-user detection (MUD) at the base station (BS). The existing methods modify the factor graph on the message-passing algorithm (MPA) for each CB collision scenario, resulting in high computational complexity. In this paper, we aim to confirm that even in the presence of CB collisions, MUD performance can be ensured through a deep learning (DL)-based receiver and explore its limitations. We propose a single DL architecture for collision-aware MUD (CA-MUD) that can tolerate CB collisions, without resorting to the distinct MUD processes associated with individual collision scenarios. To facilitate the generation of training data for CA-MUD that comprehensively represents the grant-free SCMA scenario, we introduce a transceiver model that regulates the number of active CBs and sets the maximum tolerable CB collisions. Simulation results demonstrate that our proposed approach allows a single CA-MUD network to handle various CB collision scenarios, including 2-fold CB collision subject to a limited number of active users.