Deep-Neural-Network-Aided Cross-Slot User Equipment Scheduling for Grant-Free Random Access

Massive machine-type communications (mMTC) is an important scenario to support Internet of Things (IoT) services. However, the massiveness of user equipments (UEs) poses new challenges for existing grant-free random access (GF-RA) schemes, such as pilot collisions and accumulation of failed UEs. To address this problem, we consider consecutive RA slots, and propose a cross-slot UE scheduling strategy for collision resolution in GF-RA systems. Specifically, different types of UEs are scheduled to select different sets of pilots via the feedback information. In this way, pilot collisions can be alleviated by dynamic UE scheduling. Then, we construct three deep neural networks (DNNs) for different collision-resolution tasks in UE scheduling, and these DNNs are trained to improve the scheduling efficiency. Furthermore, we adopt a matched training strategy for DNN training, which integrates the loss function of different DNNs to improve the output accuracy. Finally, a complete GF-RA scheme with DNN-aided UE scheduling (DNN-UESch-GFRA) is established. Simulation results are provided to verify the effectiveness of the matched training strategy, and show that the DNN-UESch-GFRA scheme can effectively resolve random access (RA) collisions and improve RA throughput.