Research On Mmwave Blockage And Beam Prediction Based On Deep Learning

As the communication frequency band increases and large-scale antenna array technology becomes more prevalent,ensuring transmission quality and link stability has become increasingly challenging.While the sub-6GHz frequency band utilized in the past falls short in terms of data transmission rates and communication bandwidth,the millimeter-wave frequency band offers greater support due to its numerous unutilized frequency bands.However,millimeter waves suffer from greater propagation loss and weaker diffraction capabilities than sub-6GHz radio waves,making mobile communication links in the millimeter-wave band more susceptible to interruption by blockages.Large-scale antenna array technology utilizing beamforming can overcome this disadvantage,but the increase in the number of antennas poses challenges to beam alignment.In a mobile communication system integrating sub-6GHz and millimeterwave frequency bands,communication links in the millimeter-wave band are easily blocked,and beam alignment is challenging.One possible solution is to utilize past sub-6GHz channel information to predict blockage states and optimal beams for millimeter waves.This involves predicting future blockage states to allow the communication system to take countermeasures in advance to avoid link interruption and predicting the optimal beam to generate the best possible antenna array in time to avoid beam aging.With the rapid development of machine learning technology,this paper aims to study machine learning methods for blockage prediction and beam prediction in a mobile communication system integrating sub-6GHz and millimeter-wave frequency bands,and consider the potential challenges and corresponding solutions.1.A prediction model based on RNN is designed to achieve crossband prediction from the sub-6GHz channel matrix to millimeter-wave blockage state and optimal beam.Channel characteristics caused by diffraction in different frequency bands when blockage is about to occur at the receiving side are analyzed,which is to improve the accuracy of blockage prediction and beam prediction.Wireless Insite channel simulation software based on ray tracing technology is utilized to generate channel data,and experiments are designed to verify the effectiveness of diffraction features introduced in blockage prediction and beam prediction.Finally,experiments are designed to verify the generalization ability of the prediction model for different communication scenarios.2.In mobile communication systems,the acquisition of the sub-6GHz channel matrix required for blockage and beam prediction depends on CSI feedback at the receiving side.However,in actual communication,CSI may be delayed due to transmission delay,resulting in an aged input channel matrix for blockage and beam prediction.Additionally,CSI feedback methods such as aperiodic feedback and semi-persistent feedback may lead to a lack of input channel matrix for blockage and beam prediction.This paper studies the effect of aging and missing input channel matrix on the prediction accuracy of blockage and beam prediction.This paper proposes a method that uses Convolutional Neural Networks(CNN)to predict and update the aged and missing channel matrix,which is compared with the traditional Autoregressive(AR)model.Finally,the improved accuracy of blockage and beam prediction is verified after utilizing the predicted channel matrix to update and replace the aged and missing channel matrix.3.The previous research focused on addressing blockage and beam prediction issues within a single scenario and link.To broaden the application of the designed prediction model,this study considers blockage and beam prediction as a combined prediction problem.The analysis explores the necessity and performance improvements that arise from scenario and link generalization on blockage and beam prediction.Additionally,the study introduces fundamental concepts of transfer learning,establishes a generalization model that is based on transfer learning,and analyzes the potential benefits of applying transfer learning to blockage and beam prediction.The prediction model trained in the original scenario and link is utilized to predict new scenarios and links,respectively,with a focus on exploring the degree of decline of the prediction effect.The study also investigates the implementation of transfer learning from existing models to new scenarios and links and examines the impact of freezing layers and learning rates on the effectiveness of transfer learning.Ultimately,the research aims to obtain a parameter setting method that achieves better prediction results under low overhead conditions.