| In recent years,with the development of 5G wireless communication,the MultipleInput Multiple-Output(MIMO)system can increase transmission rate,spectral efficiency,anti-interference ability,and coverage and reliability.As two key technologies in MIMO systems,MIMO signal detection and precoding are highly valued.Today,with the upsurge of artificial intelligence,deep learning has been developed rapidly.As a method of deep learning,model-driven deep learning combines the physical model with the deep learning network,uses the physical model to guide the learning process of the deep learning network,reduces the structural complexity,and improves the efficiency and accuracy of deep learning;On the one hand,model deep learning can learn model features from data automatically,and deals with high-dimensional and complex data effectively,which performs well in MIMO signal detection and precoding problems.Therefore,it is significant to take research in MIMO detection and precoding problems.Based on the model-driven deep learning method,this thesis will focus on the two popular issues of MIMO: signal detection and precoding.The main innovative work showed as:1.In terms of MIMO signal detection technology,this thesis combines the gradient model optimization method with the powerful model learning ability of deep learning.We propose a learnable step size momentum gradient projection network(LSMGPNet)signal detection method.Compared with the traditional detection algorithm,the bit error rate has been decreased by 5d B.2.In terms of MIMO millimeter-wave hybrid precoding technology,this thesis combines the manifold gradient optimization method with model-driven deep learning method.We propose a hybrid precoding method for fast complex oblique manifold network(FOACOMNet).Compared with the traditional millimeter-wave precoding algorithm,the spectrum efficiency performance has been significantly improved.3.In terms of RIS-assisted MIMO system hybrid precoding technology,this thesis combines the Riemannian manifold gradient optimization algorithm with model-driven deep learning,and proposes a model-driven Riemann constant modulus complex circular manifold network(RGDNet)method for precoding.Compared with traditional algorithms,the proposed method has a significant improvement in spectral efficiency performance. |
