Deep Learning Multi-User Detection for PD-SCMA

The performance of hybrid multi-radio access technologies depends on the sufficiency of the multi-user detection (MUD) at the receiver. For optimal performance of the hybrid power-domain sparse code multiple access (PD-SCMA), robust detection strategies are necessary to alleviate MUD complexity and reduces computational time. Deep learning (DL) based MUD techniques are the most promising as they can detect all symbols of an overloaded PD-SCMA without requiring additional operations of channel estimation and interference cancellation. This work proposes a deep neural network (DNN) aided MUD scheme (DNN-MUD) for an uplink PD-SCMA system supporting near users (NUs) and far users (FUs) multiplexed in power-, and code-domain, respectively. The proposed DNN-MUD features a unified framework that jointly performs successive interference cancellation (SIC) and message passing algorithm (MPA)/expectation propagation algorithm (EPA) operations to overcome interference propagation of SIC and computational complexity of MPA/EPA. The DNN training is enhanced by batch normalization to reduce the internal covariant shifts, thus enhancing the efficiency of detection. Performance results show that the average symbol error rate (SER), complexity and computational time of the proposed DNN-MUD significantly outperforms the conventional joint SIC-MPA/EPA schemes.