A Configurable Deep Learning-based Architecture for NPRACH Reception

In this work, a novel deep learning (DL) based receiver is developed for narrowband random access channel (NPRACH) detection and uplink synchronization in narrowband Internet of things (NB-IoT) systems. Particularly, a configurable DL-based design is proposed in which the functionality and signal flow of the blocks constituting the network stay unchanged, and only their input/output dimensions are altered in accordance with the preamble format and the coverage area. It is shown that the proposed architecture provides a preamble detection probability > 99% at the targeted signal-to-noise (SNR) regimes for both preamble format 0 (PRM-0) and preamble format 1 (PRM1), along with a time-of-arrival estimation error < 3.646 s, as required by the third generation partnership project (3GPP). It is also shown that the complexity of the proposed architecture is much lower than that of the conventional NPRACH receiver adopting the two dimensional fast Fourier transform (2D-FFT) algorithm, thereby making it feasible for practical implementation.