Deep Learning BASED Signal Detection for High Performance NOMA Communication

The presented research outlines a deep learning strategy for the signal detection for a Non-Orthogonal Multiple Access (NOMA) framework. In NOMA, the receiver carried out the successive interference cancellation (SIC) process and then decodes multiple users in a sequential manner. The accuracy of detection significantly relies on correctly identifying previous users due to the effects of error propagation. In this scenario, a deep learning methodology is proposed, specifically employing the Gated Recurrent Unit Deep Neural Network (GRU-DNN), to detect transmitted symbols without the explicit estimation of channels. The proposed method is capable of performing both channel estimation and signal detection at the same time. Training on a simulated channel dataset, the proposed DNN is designed to approximate the relationship between the received superposed symbol and its transmitted symbols. According to the findings, the GRU-DNN algorithm demonstrates superior performance compared to conventional methods like Least Squares (LS), Minimum Mean Square Error (MMSE) and alternative DNN-based channel estimation techniques. This superiority is evident through the Symbol-Error Rate (SER) outcomes observed in the results.