Deep learning-based flexible joint channel estimation and signal detection of multi-user OFDM-NOMA

Abstract

This paper proposes a deep learning (DL)-based joint channel estimation and signal detection in multi-user orthogonal-frequency division multiplexing-non-orthogonal multiple access (OFDM-NOMA) schemes over Rayleigh fading channels. In the considered model, we assume that channel state information (CSI) is not known at the receivers (users), thus, to obtain CSI responses, we use two type pilot insertions (i.e., block type and comb type). According to the received pilot responses and data signal, the proposed DL-based detector (DLD) can detect symbols at all users without requiring any additional operations (e.g., channel estimation, interference canceler, etc.). Then, we evaluate the error performance of the proposed DLD in Monte Carlo simulations and compare the results with the benchmark (i.e., successive interference canceler-based detector (SICD) with perfect CSI at the receiver). It is revealed that the proposed DLD outperforms the SICD significantly even if the perfect channel state information at the receiver (CSIR) is available for the SICD and is not for the DLD. Moreover, the proposed DLD provides a flexible detection, where all users can detect their symbols once the offline training is completed in case either equal to or less than the number of the users used for training are allocated. Based on the extensive simulations, we also reveal that the proposed DLD outperforms the SICD with perfect CSIR when less users are allocated in the testing than the offline training. This proves the flexibility of the proposed DLD and enables grant-free access. Lastly, we also present BER performances of DLD for different fading channel conditions where the pre-trained DLD with Rayleigh fading channel is used as online joint channel estimation and detection algorithm without re-training. We demonstrate that the DLD still outperforms conventional SICD and this reveals the robustness of the DLD against to fading conditions. Thus, the DLD could be used in any fading conditions without re-training and this is quite promising for practical implementations.