DNN-based Active User Detection for an NB-IoT Compatible Grant Free NOMA System

Abstract

Grant free non-orthogonal multiple access (GF-NOMA) is a promising access method for massive machine type communication (mMTC), which has several advantages when compared to the conventional grant based access method, such as, reduced latency, smaller scheduling and signalling overheads, and improved energy efficiency. Since there is no explicit grant given to each user in GF-NOMA, detecting all the active users present, i.e., active user detection (AUD) at the base station (BS) becomes crucial. Typically, AUD is performed using correlation with all possible preamble sequences transmitted by the GF-NOMA users. Recently, deep learning (DL) based models have emerged as a viable alternatives for AUD. However most of these works assume perfect timing and frequency synchronization of users, which hardly occurs in practice. In this work, a GF-NOMA scheme and a deep neural network (DNN) model for AUD are proposed. The proposed scheme is compatible with the narrowband Internet of things (NB-IoT) and the proposed DNN-based AUD mechanism accounts for the impact of timing and frequency offsets. It is demonstrated that the performance of proposed DNN based AUD scheme is comparable (or slightly better) than the conventional method while providing a significant reduction in the computational complexity.