A Model-Driven Dimension-Elevating OAMP Detector for MIMO-OTFS Communications

It is crucial and also challenging to design signal detection methods that possess both high accuracy and efficiency in MIMO-OTFS systems. In this paper, we adopt the modeldriven approach, by unfolding deep learning into the classical orthogonal approximate message passing (OAMP) algorithm, to design a dimension-elevating OAMP (DE-OAMP) detector, which achieves improvement in accuracy while maintaining computational complexity. Specifically, the DE-OAMP unfolds the iterative process of the OAMP into the multi-layer neural network, with each layer containing a linear estimator and a non-linear estimator. In the nonlinear estimator, we break the paradigm of introducing trainable scalar parameters by employing the dimension-elevating vector parameters to improve the detection performance. Simulation results demonstrate that, based on the widely-used Extended Vehicular A (EVA) channel, compared to the OAMP-Net, the CG-OAMP-Net, the CG-OAMP, the OAMP, the MP and the LMMSE, the DE-OAMP reduces the bit error rate by $25.57 \%, 25.78 \%, 59.33 \%, 59.33 \%, 59.55 \%$, and $92.45 \%$ when SNR $=\mathbf{1 2} \mathbf{~ d B}$, respectively.