【摘要】 在大规模MIMO系统中，精确获取下行信道状态信息（CSI）对基站而言至关重要。然而，目前基于深度学习的CSI预测模型普遍依赖训练数据与测试数据之间分布的一致性或相似性假设，在面对新环境时，性能往往会显著下降甚至失效。针对大规模MIMO系统中CSI预测模型泛化性不足的问题，提出了一种基于元学习的训练框架，以提升CSI预测模型在新环境中的预测精度。针对高度动态时变的无线环境，所提元框架通过在内层更新中引入动量优化器来丰富元梯度的表征能力，进而帮助模型获取更高效的元知识来提升对新环境的泛化能力，仿真实验结果表明，相较于传统的迁移学习和元学习的方法，所提方法需要更少的迁移样本和步数即可适应新环境，展现出卓越的泛化能力。更多还原

【Abstract】 In massive multiple-input multiple-output(MIMO) systems, accurate acquisition of downlink channel state information(CSI) is crucial for base stations. However, existing deep learning-based CSI prediction models generally rely on the assumption of consistency or similarity between the training and testing data distributions, leading to significant performance degradation or even failure when facing new environments. To address the limited generalization of CSI prediction models in massive MIMO systems, a meta-learning-based training framework is proposed to improve prediction accuracy in new environments. Considering highly dynamic and time-varying wireless environments, the proposed meta-framework introduces a momentum optimizer into the innerloop update to enrich the representation capability of meta-gradients, thereby enabling the model to acquire more efficient metaknowledge and enhance generalization to new environments. Simulation results show that, compared with conventional transfer learning and meta-learning approaches, the proposed method requires fewer transfer samples and adaptation steps to achieve satisfactory performance in new environments, demonstrating superior generalization capability.更多还原