High SNR SCMA Detection via Transfer Learning From Low SNR Region

Abstract

Sparse code multiple access (SCMA) is a competitive candidate multiple access technology for future wireless communication systems. In this article, a transfer learning (TL)-based SCMA detection scheme is proposed to improve the performance of the deep neural network (DNN) detector for the downlink SCMA system. First, we propose a detection framework that preserves model parameters trained on datasets with varying signal-to-noise ratios (SNRs) and adaptively selects them during detection based on the estimated channel SNR. Then, we analyze the reason why the DNN detector, which can achieve the same BER performance as the message passing algorithm (MPA) in the low SNR region, fails to achieve MPA performance in the high SNR region. Later, a TL-based SCMA detection scheme is proposed, which consists of pretraining, fine-tuning and online detection. Simulation results show the proposed TL-based SCMA detection scheme can achieve improved BER performance compared to the deep-learning (DL)-based scheme trained from scratch. Moreover, to alleviate the gap between the source domain and the target domain, a successive TL strategy is proposed, which makes the transfer process smoother and further improves the performance by introducing intermediate states.