Model-Driven Deep Learning-Based Signal Detector for CP-Free MIMO-OFDM Systems

Abstract

The absence of cyclic prefix (CP) can increase the spectral efficiency of multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems. However, CP removal will make signal detection challenging. In this paper, we develop a model-driven deep learning (DL)-based detector to resolve this problem. The prototype of the detector is the orthogonal approximate message passing (OAMP) algorithm, which has a strong ability to mitigate interference but involves matrix inversion with high complexity. We first use the conjugate gradient method to reduce the computational complexity of OAMP. Then, by unfolding the revised algorithm into a network and learning the optimal values of its parameters, the detection performance can be significantly improved. Complexity analysis indicates that the proposed scheme can reduce the running time to only a quarter of that required in OAMP while achieving remarkable performance in bit-error rate.