一种基于信道矩阵模式的毫米波波束预测方法

Millimeter wave beam prediction via channel matrix patterns

毫米波(millimeter wave,mmWave)是5G应用的一项重要技术,具备更高的上下行传输速率,更低的响应时延。mmWave系统通常基于码本(codebook)进行波束预测,但穷尽codebook匹配最佳波束成形向量耗时较多。在保证高吞吐量的同时,如何降低codebook搜索成本,已然成为mmWave系统优化方向之一。文章提出一种基于信道矩阵模式的毫米波波束预测方法,通过多载波信道响应整合并可视化,将结构化数据转换为图像样本,利用深度学习模型Efficient NetV2-S提取图像特征,进行迁移学习时,将特征提取层参数固定,仅训练最后一层神经元参数,保证预测准确率同时降低训练时间成本,再基于codebook进行波束预测。经验证,该方法Top-1准确率提升近4.9%,Top-3准确率提升近5.8%,显著削减了codebook搜索成本,优化了mmWave系统通信质量及响应时延。

Millimeter wave(mmWave)is a significant technology for 5G application,featuring higher achievable rate as well as lower response delay.As usual,mm Wave executes beam prediction based on whole codebook.It,however,consumes tremendous time cost when matching the best beam vector via exhaustive search.Thus,it has became a hot topic on how to slash codebook sea-rching cost while ensuring high throughput.The paper proposes an approach for mmWave beam prediction using channel matrix patterns.Via multi-sub-carriers integration and visualization,our research converts structured data to image samples.Then,our research adopts EfficientNetV2-S for feature extraction.When leveraging transfer learning,our research freezes backbone para meters and only trains the last layer neurons.In this way,it dramatically cuts training cost while ensuring prediction accuracy.The statistics suggest that,our research improves Top-1 accuracy 4.9%and Top-3 accuracy 5.8%on average,which is a remarkable contribution to mmWave system optimization.