Deep Learning-Based Collision-Aware Multi-User Detection for Channel-Modulated Codebooks in Grant-Free Sparse Code Multiple Access Systems

Abstract

In grant-free sparse code multiple access (SCMA) systems, each active user transmits data using randomly selected SCMA codebook along with its associated preamble. When multiple users select the same codebook, i.e., leading to codebook collisions, the detection of channel-modulated codebooks is still possible through collision-aware multi-user detection (CA-MUD) using their associated preambles. However, traditional CA-MUDs are designed with unique configurations tailored to each of the extensive codebook activity scenarios, thereby significantly enhancing the detection complexity and limiting the practical implementation of GF-SCMA systems. In this paper, our objective is to propose a deep learning (DL)-based CA-MUD capable of efficiently handling diverse codebook activities with a single detector, even in the presence of codebook collisions. Toward this end, we propose a multi-task learning-based DL architecture for CA-MUD that can tolerate codebook collisions, without resorting to distinct CA-MUD processes for individual collision scenarios. A key innovation in our approach is an input pre-processing method for efficient CA-MUD training that generates a channel-modulated codebook vector at the receiving end, enhancing the learning process. Simulation results demonstrate that our proposed approach enables a single CA-MUD network to manage various codebook activity scenarios, including 2-fold codebook collision, within a limited number of active users, while ensuring robustness against channel estimation errors.