Deep Neural Network Based Active User Detection for Grant-Free Multiple Access

The dramatic increment of Internet of Things (IoT) devices has become a challenging problem in wireless communications. The IoT devices have the characteristic of sporadic transmission, which causes severe signaling overhead and latency problems for the grant-based multiple access systems. To solve this problem, grant-free non-orthogonal multiple access (GF-NOMA), with user equipments directly performing uplink transmission to the base station, has emerged as a promising solution. Without pre-allocating the resources, the active user detection (AUD) is required in GF-NOMA, and the massive number of devices operated under limited frequency resources makes the AUD problem challenging. Inspired by the interference cancellation framework, we propose a novel model-based deep learning (DL) architecture, called active user detection-interference cancellation neural network (AUD-ICNN), to address the AUD problem under the frequency selective fading channel, without needing the channel state information (CSI) and the user sparsity information. Specifically, the proposed AUD-ICNN exploits the signal structure to estimate the activation probability and then cancel the interference according to the estimated activation probability. Simulation results show that the proposed AUD-ICNN outperforms the conventional compressive sensing (CS) algorithms and significantly reduces the error rate compared to the state-of-the-art DL architectures. Meanwhile, the complexity is reduced by 6.9 and 13.9 times compared to the state-of-the-art DL and CS methods, respectively. Furthermore, unlike the existing DL-based architectures, the proposed AUD-ICNN uses a single neural network to deal with different user sparsity. Finally, a transfer learning method is proposed to extend the proposed AUD-ICNN to a robust sparsity estimation neural network.