基于深度学习的FDD大规模MIMO系统下行CSI反馈技术

Downlink CSI Feedback for FDD Massive MIMO System Based on Deep Learning

准确的CSI反馈是FDD大规模MIMO系统中的关键技术之一。为了解决现有基于深度学习的压缩反馈会因分解复值CSI矩阵而造成信息损失的问题,提出了一种改进的下行CSI反馈技术,基于全卷积模型的CVCNN算法,利用复卷积层和反卷积层分别对下行CSI进行压缩和解压缩,避免了分解复值带来的损耗。实验结果表明,相较于传统的基准CsiNet算法,所提出的CVCNN算法提升了下行CSI的重构精度,同时降低了存储和计算开销。

Accurate CSI feedback is one of the key technologies for FDD massive MIMO systems. Existing deep learning-based compression feedback will cause information loss due to decomposing complex-valued CSI matrices. To solve the problem, an improved downlink CSI feedback technology is proposed. Specifically, the CVCNN algorithm based on the fully convolutional model utilizes complex convolutional layers and deconvolutional layers to compress and decompress downlink CSI respectively, which avoids the loss caused by decomposing complex values. The experimental results show that the proposed CVCNN algorithm improves the reconstruction accuracy of downlink CSI compared with the traditional benchmark algorithm CsiNet, while reducing the storage and computational overhead.