Deep learning-based sequential models for multi-user detection with M-PSK for downlink NOMA wireless communication systems

Abstract

Non-orthogonal multiple access (NOMA) techniques have the potential to achieve large connectivity requirements for future-generation wireless communication. NOMA detection techniques require conventional successive interference cancellation (SIC) techniques for uplink and downlink transmissions on the receiver side to decode the transmitted signals. Multipath fading significantly impacts the SIC process and correct signal detection due to propagation delay and fading channel. Deep learning (DL) techniques can overcome conventional SIC detection limitations. Signal detection for a multi-user NOMA wireless communication system that relies on orthogonal frequency-division multiplexing (OFDM) is discussed using various DL approaches in this paper. For multi-user signal detection, different deep learning-based sequential model neural networks, gated recurrent unit (GRU), long short-term memory (LSTM), and bi-directional long short-term memory (Bi-LSTM) are applied. The deep neural network is initially trained offline with multi-user NOMA signals in the OFDM system and used to recover transmitted signals directly. DL-based sequential models with different cyclic prefixes and fast Fourier transforms with various M-phase shift keying (M-PSK) modulation schemes are discussed with deep learning optimization algorithms. In simulation results, the conventional SIC technique with minimum mean square error approach is compared to the effectiveness of DL-based models for signal detection of multi-user NOMA systems by their bit error rate performances. The root mean square error performance of different deep learning-based sequence models with other optimizers is also discussed. Moreover, the robustness of the Bi-LSTM is evaluated with the reliability of other DL-based sequential model applications in the multi-user downlink NOMA wireless communication systems.