基于数字孪生信道的大规模MIMO信道低开销预测研究

Low-Overhead Channel Prediction for Massive MIMO Based on Channel Twin

准确的CSI对于大规模多输入多输出系统的性能优化至关重要,而其庞大的天线数量会给CSI获取带来巨大的开销与时延。提出了一种基于Crossformer网络的信道预测方法,利用该方法可以在数字孪生信道中获得准确有效的CSI,从而实现对物理信道的精准指导。具体地,将使用间隔多个时隙的历史CSI数据与孪生环境中终端的位置坐标作为神经网络的输入,首先通过维度分段模块将输入的多元时间序列嵌入到二维矢量数组中,以保留时间、空间以及跨维度信息,然后通过二阶注意层来捕获多维度间的依赖关系,最后通过分层编码器-解码器结构利用挖掘到的信道空时依赖关系和环境与信道跨维度依赖关系联合预测未来多个时隙的CSI。最后将所提出的信道预测框架与传统的LSTM与Transformer模型的预测性能进行比较,验证该方法的优越性。

Accurate Channel State Information(CSI)is critical for optimizing the performance of massive Multiple Input Multiple Output(MIMO)systems.However,the large number of antennas in these systems results in significant overhead and latency for CSI acquisition.This paper proposes a novel channel prediction method based on a Crossformer network,leveraging digital twin channels to achieve precise and efficient CSI prediction,thereby guiding the physical channels accurately.The method utilizes historical CSI data from multiple previous time slots and terminal position coordinates in the twin environment as inputs to the neural network.Initially,a dimensional segmentation module embeds the multivariate time series inputs into a two-dimensional vector array to preserve temporal,spatial,and cross-dimensional information.Subsequently,a second-order attention layer captures dependencies across multiple dimensions.Finally,a hierarchical encoder-decoder structure exploits the spatiotemporal dependencies of the channel and cross-dimensional dependencies between the environment and the channel to jointly predict the CSI for multiple future time slots.The proposed channel prediction framework is evaluated against traditional LSTM and Transformer models,demonstrating its superior performance.