Research On Nonlinear MIMO Signal Processing Based On Bayesian Inference

In the wireless communication network,in order to further increase the wireless data transmission rate,the physical layer mainly relies on increasing the operating frequency(such as millimeter wave,terahertz band)in exchange for large bandwidth;increasing the number of base station antennas in exchange for high multiple-input multiple-output(MIMO)multiplexing diversity gain;increasing the number of base stations in exchange for large user access.However,power consumption caused by 5G massive data has become a concern.5G deployment and applications face severe challenges in power consumption and Operating Expenses/Capital Expenditure(OPEX/CAPEX).The nonlinear(NL-)MIMO whose receiver is designed by low-power/low-cost envelope/phase detector and low-precision ADC,can effectively reduce the circuit power consumption of the base station,and alleviate the high OPEX/CAPEX issues in 5G/6G.However,the introduction of non-linear devices will cause the traditional ideal In-phase/Quadrature(I/Q)complex baseband observations to suffer from various nonlinear/linear distortions,such as loss of phase/amplitude observation information,analog to digital converter(ADC)quantization distortion.Traditional communication baseband signal processing algorithms are mostly based on classical I/Q complex baseband observations and cannot be applied in NL-MIMO.Therefore,it is necessary to develop new communication signal processing algorithms for the latter.This thesis mainly focuses on the three NL-MIMO schemes including halved phase-only(HPO-)MIMO,magnitude-only(MO-)MIMO,and phase-only(PO-)MIMO and discusses the following signal processing issues:1.Considering the inevitable time and frequency deviation in NL-MIMO system,a joint time-frequency synchronization algorithm is proposed,which exploitsthe repeatability of a pilot preamble.Starting-time and carrier frequency offset(CFO)are jointly estimated under a grid-searching manner.At first,continuous starting-time and CFO estimation are uniformly sampled from their potential value ranges.On all potential starting-time points,the received phase measurements are uniformly sliced according to the repetition structure of the preamble,and then the phase rotations due to CFO are removed based on each potential CFO value.The shifted phase measurements would show strong similarities if CFO is successfully compensated.The similarities among these sliced phase sequences are characterized by a metric named phase sequence distance(PSD).They are summed up as a final PSD,whose minimum value corresponds to the estimated starting time and CFO.Simulation results validate the effectiveness of the synchronization algorithm under different signal-to-noise ratios and lengths of preamble.HPO-MIMO with our synchronization algorithm can even perform similarly as its perfectsynchronized correspondence.2.In terms of the channel estimation(CE)problem in the NL-MIMO system,the spatio-temporal sparsity(STS)of the wireless channel is exploited to improve the CE in NL-MIMO.At first,The sparse CE problem is formulated as the generalized linear mixing one and resolved by a modified generalized vector approximate message passing(GVAMP)algorithm with expectation maximization(EM)mechanism.The prior distribution of sparse channel is modeled as Bernoulli Gaussian-mixture(BGM).Consequently,NL-MIMO channel responses and unknown parameters including BGM parameters and noise variance are updated by the GVAMP and EM procedure alternatingly.A gradient descent(GD)method is proposed for EM update of noise variance in NL-MIMO,where closed-form solution is missed due to the complex formulas of likelihood under phase/magnitude observations.Monte Carlo integration technique is exploited to numerically derive the gradient of the noise variance under phase/magnitude observations.Simulation results show that the transmission power and pilot length are significantly saved after exploiting the STS,and both the EM-GVAMP and GD estimators are validated.3.As for the problem of estimating the source number and the direction of arrival(DO A)in NL-MIMO system,a joint DO A and source number estimation method is proposed for NL-MIMO under the frameworks of sparse signal recovery and machine learning clustering.At first,the DOA estimation problem is converted as detecting supports of sparse signal recovered from a generalized linear mixing(GLM)problem under phase/magnitude measurements.The latter is then resolved by a modified generalized approximate message passing(GAMP)algorithm with a EM mechanism.The DOA estimates are acquired snapshot per snapshot and then clustered by density-based spatial clustering of applications with noise(DBSCAN),whose cluster centers and number are treated as DOAs and source number respectively.The proposed algorithm is applied to PO-MIMO system.Simulation results indicate that the proposed estimator can effectively handle the phase quantization losses and estimate source number and DOAs based on quantized phase measurements of PO-MIMO.The estimation performance is steadily improved by adopting more antennas and sampling snapshots.