In this paper, we study an application of deep learning to uplink multiuser detection (MUD) for non-orthogonal multiple access (NOMA) scheme based on Welch bound equality spread multiple access (WSMA). Several non-cooperating users, each with its own preassigned NOMA signature sequence (SS), transmit over the same resource. These SSs have low correlation among them and aid in the user separation at the receiver during MUD. Several subtasks such as equalizing, combining, slicing, signal reconstruction and interference cancellation are involved in MUD. The neural network (NN) considered in this paper replaces these well-defined receiver blocks with a single black box, i.e., the NN provides a one-shot approximation for these modules. We consider two different supervised feed-forward NN implementations, namely, a deep NN and a 2D-Convolutional NN, for MUD. Performance of these two NNs is compared with the conventional receivers. Simulation results show that by proper selection of the NN parameters, it is possible for the black box approximation to provide faster and better performance, compared to conventional MUD schemes, and it achieves almost the same symbol error rate as the ultimate one obtained by the complex maximum likelihood-based detectors.

中文翻译：

---

基于深度学习的NOMA多用户检测

在本文中，我们研究了深度学习在基于Welch绑定等价扩展多址（WSMA）的非正交多址（NOMA）方案的上行链路多用户检测（MUD）中的应用。几个非合作用户，每个用户都有自己的预分配NOMA签名序列（SS），它们通过同一资源进行传输。这些SS之间的相关性较低，并有助于在MUD期间在接收器处进行用户分离。MUD涉及多个子任务，例如均衡，组合，切片，信号重建和干扰消除。本文考虑的神经网络（NN）将这些定义明确的接收器块替换为单个黑匣子，即NN为这些模块提供了一次近似。我们考虑了两种不同的监督前馈NN实现，即深度NN和2D卷积NN，用于MUD。将这两个NN的性能与常规接收器进行比较。仿真结果表明，通过NN参数的适当选择，可能的是黑盒近似提供更快和更好的性能，比传统的MUD方案，以及它实现几乎相同的符号错误率如由复合得到的最终一个基于最大似然的检测器。