Deep Learning Empowered Semi-Blind Joint Detection in Cooperative NOMA

Abstract

In this paper, we propose a multi-user symbol detection in cooperative-non-orthogonal multiple access (C-NOMA) schemes via deep learning (DL). We use a DL-based detection (DLDet) in both users instead of conventional detectors. Therefore, an iterative detector (i.e., successive interference canceler (SIC)) at the near user (UE1) and a combining plus optimum detector (i.e., maximum ratio combining (MRC) and maximum-likelihood (ML) detector) at the far user (UE2) are not required anymore. The proposed DLDet can detect symbols at the both users simultaneously. Besides, the DLDet does not require an additional channel estimation algorithm to acquire the channel state information (CSI) at the receivers. The multi-user symbol detection is performed based on the received pilot responses simultaneously, thus calling semi-blind. We train the proposed DLDet offline over Rayleigh fading channel and then, we use the offline-trained DLDet as an online detection algorithm. We compare the error performance of the DLDet with two benchmarks: conventional C-NOMA and threshold-based selective C-NOMA (TBS-C-NOMA). With the extensive simulations over Rayleigh fading channels, we reveal that the DLDet outperforms both C-NOMA and TBS-C-NOMA and achieves the full diversity order (i.e., 2). Besides, this performance gain is up to 10 dB and 3-7 dB in C-NOMA and TBS-C-NOMA, respectively, which is very promising for the energy-limited networks. Moreover, in order to improve the error performance of the C-NOMA, the TBS-C-NOMA introduces two disadvantages, which are a signaling overhead to obtain the optimum threshold value and a capacity performance decay due the silence of the relay when the threshold is not satisfied. To this end, without using a threshold, outperforming TBS-C-NOMA is essential to revoke these disadvantages. The DLDet accomplish this; hence, the power of the DLDet is unveiled in coping with the error propagation. Then, to reveal the robustness of the DLDet against different fading conditions, we use the offline-trained DLDet as an online detector over different fading channels (i.e., Nakagami- m and Rician). We show that the DLDet outperforms conventional detectors over those fading channels and again outperforms the conventional C-NOMA and TBS-C-NOMA performances, although it is trained over Rayleigh fading channel. Indeed, the DLDet has better performance than the C-NOMA and TBS-C-NOMA even though they are assumed to have perfect CSI whereas the DLDet uses only a single pilot signal. The DLDet again provides the full diversity order (i.e., 2m for Nakagami- m and 2 for Rician). This proves the robustness of the DLDet and shows that the DLDet performs well regardless of the fading conditions once it is trained for any fading channel.