Research On Uplink Massive MIMO Detection Algorithm Based On Model-driven Deep Learnin

Massive MIMO systems with multiple antennas can significantly improve the capacity and reliability of communication systems,and can provide support for the development of wireless communication standards such as B5 G and 6G.Because of the large number of antennas at the base station,the signal detection algorithm needs to process large amounts of data.At present,it is difficult to achieve an effective compromise between computational complexity and detection performance based on traditional signal detection algorithms.Data-driven deep learning-based detection algorithms can effectively improve detection performance,but their large number of required training sets,many network training parameters,and difficulty in interpreting the network limit their practical applications.The traditional belief propagation(BP)detection algorithm has excellent detection performance,but suffers from high computational complexity and difficulty in convergence.The specific research to address the above issues is as follows:Based on the traditional BP detection algorithm,this paper analyzes and derives the optimization strategy of simplifying information update and introducing damping mechanism to collate the IM-BP algorithm,which provides the theoretical basis for the subsequent use of model-driven deep learning to build the detection network architecture.Experimental simulations show that the IM-BP algorithm can effectively reduce the computational complexity and accelerate the convergence speed with a slight decrease in detection performance.In view of the limitations of data-driven detection algorithm,this paper proposes BP-IB-Net detection algorithm based on private integrated block in model-driven.The main principle is to replace the complex core calculation process of IM-BP algorithm by private integrated block network.In addition,in order to reduce the dimension of detection network input,the network input is preprocessed in this paper.The training set required by the training network is relatively small,which can effectively reduce the training time,and the network framework has good interpretability.Simulation results show that the BP-IB-Net algorithm requires fewer iterations and converges faster when achieving the same detection performance as the IM-BP algorithm.In addition,the robust analysis of the BP-IB-Net algorithm is carried out.In the presence of errors in the channel state information,the algorithm still maintains good detection performance.Besides,the BP-DU-Net detection algorithm is proposed based on the deep unfolding approach in model-driven.The core of the algorithm regards the IM-BP iterative algorithm as a layer-connected neural network and parameterizes the iterative process of the IM-BP algorithm.In the parameterization process,the formula for calculating the posterior probability likelihood ratio information is improved by introducing a scaling parameter and an offset parameter,aiming to compensate for the interference Gaussian approximation error in the IM-BP algorithm.The problem that the damping factor in the IM-BP algorithm is difficult to determine is solved by optimizing the parameters through neural networks.At the same time,the detection performance and computational complexity of the algorithm are balanced by increasing and deleting the number of iterative units of the network.Simulation results show that the BP-DU-Net algorithm can further accelerate the convergence speed of the IM-BP algorithm and improve the detection performance.