Design And Implementation Of Beam Tracking Enhancement Methods Based On Vision Learning

Owing to the abundant spectrum resources in the high frequency band,millimeter wave-based wireless communication systems are one of the most important means to improve business service capabilities in the future.At the same time,due to the high path loss characteristics of millimeter wave signal,it is necessary to use beam-forming technology to form a narrow beam with high gain for communication.In addition,in a mobile scenario,frequent beamforming processes are required in beam tracing due to the constant change of the relative position of the sending and receiving ends.However,the traditional beamforming algorithm needs to perform high complexity channel estimation,resulting in a large system overhead.Due to the short wavelength characteristics of millimeter waves,the communication link between the transceivers is getting shorter and shorter,which is closer to the line-of-sight connection,and beam prediction can be assisted by visual image-based vision learning.Therefore,the research content of this thesis is to design and implement the millimeter wave beam tracking enhancement algorithm based on vision learning on an embedded platform.The main contents of this thesis are summarized as follows.(1)Design of beam tracking enhancement algorithm based on vision learningTo solve the high system overhead problem of traditional beam tracking algorithm in high mobility scenes,this thesis designs a beam tracking enhancement algorithm based on vision learning for vehicle moving scene.Firstly,perform system and channel modeling for the beam tracking task studied in this thesis,and mathematically define the target of the task;Secondly,combining the vision features and beamforming vectors,propose a beam tracking enhancement algorithm based on vision learning,which mainly includes feature extraction module,data pre-processing module and beam prediction module,etc.,using historical sequence data to predict beamforming vectors at multiple moments in the future;In addition,in order to improve the generalization performance of the model,In this thesis,a number of related learning loss functions are designed for joint multi-task learning and training;Finally,the effectiveness of the algorithm proposed in this thesis is verified by the comparative experiment with the benchmark models and the self-ablation experiment.(2)Implementation of beam tracking algorithm on embedded platformsTo verify the feasibility of the algorithm,the algorithm is deployed on the embedded platform.Firstly,in order to facilitate the model deployment,the algorithm was transplanted,that is,the Pytorch model trained on the server side was converted into a TFLite model.At the same time,the data preprocessing module and the inference module were implemented on the embedded platform for actual algorithm operation;Secondly,in order to further improve the inference speed of the algorithm,this thesis carried out model optimization design,including model compression based on channel pruning,model weight quantification and joint compilation optimization;Finally,the designed optimizations are tested to verify the effectiveness.illustrating the engineering feasibility of the algorithm proposed in this thesis.