Research On Massive MIMO Signal Detection Technique Based On Deep Learning

Massive multiple-input multiple-output(MIMO)can provide higher spectral efficiency and energy efficiency compared to conventional MIMO systems.However,as the number of modulation orders and antennas increases,the computational complexity of conventional symbol detection algorithms becomes unaffordable and their performance deteriorates.And deep learning(DL)techniques can provide flexibility,nonlinearity,and computational parallelism for massive MIMO detection to address these challenges.In fact,the applications of deep learning in MIMO signal detection are mainly classified into two categories,namely data-driven approaches and model-driven approaches.In this paper,these two approaches are combined with conventional MIMO detection algorithms to propose a data-driven accelerated multiuser interference cancellation network(AMIC-Net)and a model-driven broyden-fletchergoldfarb-shanno network(BFGS-Net),respectively.1.In Massive MIMO systems,considering that the traditional multiuser interference cancellation algorithm(MIC)ignores the impact brought by noise in the detection process,and its detection performance is limited by the defects of its own algorithm structure,this paper proposes a data-driven AMIC-Net.First,the extrapolation technique is used to accelerate the convergence speed of the MIC algorithm,and the extrapolation factor is considered as a learnable parameter to reduce the computational complexity;then,the above accelerated MIC algorithm is unfolded by a data-driven DL approach,and a relatively simple DNN network structure is obtained by using sparse connectivity;finally,in order to adapt to the communication scenario of high-order modulation,a novel activation function is designed in this paper,which is composed of multiple softsign functions with additional learnable parameters to implement a multi-segment mapping of the set of constellation points with different modulations.Simulation results show that the proposed AMIC-Net network can bring significant performance improvement to the MIC algorithm under various antenna settings,and has lower computational complexity and better detection performance compared with existing detectors,especially in the high-order QAM modulation scenario.2.To address the problem that the traditional BFGS algorithm requires a lot of computation in updating the search direction and determining the step size,a model-driven BFGS-Net is proposed in this paper.First,a model-driven DL approach is used to expand the BFGS iterative algorithm into a multilayer connected network,and the appropriate search direction and step size are updated by learning the optimal values of these network parameters in the training phase to effectively reduce the computational complexity of the BFGS algorithm;then,to improve the expressive power of the network,a nonlinear Soft S activation function is introduced in this paper;finally,the detection performance is further improved by extending the dimensionality of the learnable parameters.Simulation results show that the proposed BFGS-Net network not only effectively reduces the computational complexity of the BFGS algorithm,but also achieves significant performance improvement under both Rayleigh fading channels and spatially correlated channels.