Sparse Signal Recovery Neural Network With Application to High-Mobility Massive MIMO-OTFS Communication Systems

A deep learning-based sparse signal recovery network <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\text{SSRnet}$</tex-math></inline-formula> is designed. This network is built on the proposed neural network <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\rm{PositionNet+}$</tex-math></inline-formula>, which takes the received signal as input and obtains the support of the desired sparse matrix without requiring a sensing matrix. Using PositionNet+, SSRnet is able to recover the sparse signal precisely, outperforming conventional methods, including least-squares (LS) estimation with perfectly known support, by virtue of its denoising behavior, while offering substantially reduced computation. The network is then utilized to perform channel estimation of high-mobility massive multiple-input multiple-output orthogonal time frequency space (MIMO-OTFS) wireless systems which is cast as a sparse signal recovery problem. In OTFS, data is modulated in the delay-Doppler domain to transform a fast time-varying and frequency-selective fading channel into a quasi-static and sparse channel. To maximize performance, OTFS systems require accurate channel estimation and low pilot signaling which are provided by SSRnet. Simulation and computational comparisons demonstrate that the proposed approach enhances performance in terms of bit error rate (BER) and normalized mean squared error (NMSE), reduces pilot symbol overhead, as well as lowers computational complexity.