Improving Performance of Large-Scale MIMO Detector Via A Proposed Two-Step Deep-Learning Architecture

A large-scale multiple-input multiple-out (LS-MIMO) transmission technique with low-resolution analog-to-digital converters (ADCs) has become one of the promising techniques for 5G and future wireless networks. In this paper, we first present the mathematical derivation to fit the high-order 4-ary Amplitude Shift-Keying (ASK) superposition modulation scheme in the conventional deep-learning LS-MIMO signal detector designed previously for the binary phase-shift-keying (BPSK) modulation scheme. Importantly, we also propose a two-step deep-learning detection network that does not require training the network again when changing the modulation order, for example, from BPSK to 4-ary ASK and more, provided that the MIMO configuration is kept unchanged. The two-step model offers a low computing requirement and shorter processing time and still achieves better bit error rate (BER) performance over the one-step architecture. The essential attribute of the proposed two-step detector is that no further training is needed makes it promising for the Internet of Things, where devices with a low computing capacity are embedded. Moreover, the advantages of the two-step deep-learning detector also open up a great possibility to deploy the adaptive modulation transmission scheme to enjoy the channel gain variation of wireless fading channels.