On Employing Deep Learning to Enhance the Performance of 5G NR Two Step RACH Procedure

Abstract

To meet the latency requirements of various business usecases and applications in 5G New Radio (NR), two step grant-free RACH procedure has been proposed in Third Generation Partnership Project (3GPP) release 16 for granting access to subscribers. However, due to the limited number of preambles, there is a non-zero probability that two mobile User Equipments (UEs) selecting same preamble signatures leading to collisions. Consequently, the base stations (gNBs) in 5G Radio Access Network (RAN) are unable to send a response to the UEs. Furthermore, with the increase in the number of cellular UEs and Machine Type Communication (MTC) devices, the probability of such preamble collisions further increases, thereby leading to reattempts by UEs. This in turn, results in increased latency and reduced channel utilization. In order to reduce contention during preamble access, we propose to use deep learning based models to design a RACH procedure that predicts the incoming connection requests in advance and proactively allocates uplink resources to UEs. We have used Recurrent Neural Network (RNN) and Long Short Term Memory (LSTM) models to predict UEs which are going to participate in two step RACH procedure. On doing extensive simulations, it is observed that both RNN and LSTM models perform equally good in reducing the number of collisions in a dense user scenario thereby enabling massive user access to 5G network.