Comparisons on Deep Learning Methods for NOMA Scheme Classification in Cellular Downlink

Abstract

Various non-orthogonal multiple access (NOMA) schemes have been proposed in multi-user downlink scenarios for the enhanced spectrum efficiency. While coherent detection at the receiver requires the knowledge of transmission configuration, the direct transmissions of the deployed NOMA scheme leads to heavy signaling overhead. To this end, this paper proposes an automatic classification framework for diversified NOMA schemes based on deep learning, which can effectively determine the NOMA encoding method without prior knowledge. Specifically, we resort to fully-connected deep neural network (FC-DNN) and long short-term memory (LSTM), which respectively use a feed-forward structure and a recurrent structure to deal with multiple received signal packets of NOMA. We then analyze the performance of the above two methods in both ideal and non-ideal scenarios, including phase and frequency offsets. The results indicate that LSTM can identify five different NOMA schemes with over 90% accuracy in the ideal case, and outperforms FC-DNN especially in low signal-to-noise region. However, FC-DNN achieves better performance than LSTM in nonideal cases.