Deep Transfer Learning for Model-Driven Signal Detection in MIMO-NOMA Systems

In this article, a odel-driven signal detection method with deep transfer learning (DTL) is proposed for downlink multiple-input multiple-output non-orthogonal multiple access (MIMO-NOMA) systems. Specifically, we first introduce some learnable parameters to an unfolded iterative algorithm for MIMO detection and improve it through a preconditioned process to speed up its convergence. Then we combine this modified algorithm with the successive interference cancellation (SIC) structure in NOMA detection to propose our learned preconditioned conjugate gradient descent network with SIC (LPCG-SIC). The proposed network has fewer parameters and lower computational complexity in which only few trainable parameters need to be optimized, and there are only addition and multiplication operations in online detection. Furthermore, to overcome the drawback that most of the existing deep learning-based detectors only focus on signal detection in a given environment and cannot adapt to changes in feature spaces or distributions, we propose a DTL-based detection algorithm and three model-driven transfer strategies for our LPCG-SIC detector to improve the reusability of the trained network. Simulation results show that our proposed method can perform signal detection under multi-user interference and outperforms conventional detectors. The transfer strategies can obtain significant performance gain with lower training cost compared to no-transfer methods.