多输入多输出（Multiple Input Multiple Output, MIMO）-正交频分复用（Orthogonal Frequency Division Multiplexing, OFDM）与可重构智能超表面（Reconfigurable Intelligent Surface, RIS）技术凭借分集增益、抗干扰特性、频谱效率优势及对无线传播环境的智能重塑能力，成为高可靠、高传输效率的全新无线系统解决方案。MIMO-OFDM系统中信号恢复解调、预编码及波束成形等核心功能以及RIS反射单元的幅度优化和相位调整，都高度依赖于精确的信道估计。因此如何准确高效地获取信道状态信息（Channel State Information, CSI）是支撑相关技术实现的关键前提。如何在降低计算复杂度和导频资源消耗的同时保证信道估计精度，成为当前制约大规模MIMO系统实际部署的重要挑战。为此本文聚焦基于深度学习的MIMO信道估计方法，主要工作包括以下两方面：

        （1）针对MIMO-OFDM通信系统中信道估计精度低、导频开销大的问题，本文提出一种基于导频设计和信道估计的联合优化方案（AE-DRSN）。方案由基于Concrete自编码器（Auto-encoder, AE）的导频优化模块和基于深度残差收缩网络（Deep Residual Shrinkage Network, DRSN）的信道估计模块组成。导频优化模块在时频网格中搜索信息量最大的导频位置，再将优化后的导频位置信息输入信道估计模块完成信道的精确估计，两模块间通过组合损失函数进行联合优化。仿真结果表明了与传统的LS、MMSE算法相比，该方案在少量导频开销下仍能实现高精度信道估计，验证了方案的有效性。

        （2）针对RIS辅助无线通信系统信道估计精度低、计算复杂度高的问题，本文设计了一种基于图神经网络（Graph Neural Network, GNN）的信道估计方案。方案利用图结构中节点特征及边关联信息，借助GNN的特征提取能力构建出复杂信道模型，有效提高级联信道估计精度。方案利用最小二乘估计获取有源元件的信道信息，进行插值计算得到所有RIS元件的信道信息，将初步估计结果输入GNN进行精细化重构，实现精确估计。仿真数据显示，相较于传统CNN算法本方案信道估计性能提升可达4 dB。

        Multiple Input Multiple Output (MIMO)-Orthogonal Frequency Division Multiplexing (OFDM) and Reconfigurable Intelligent Surface (RIS) technologies have emerged as new wireless solutions for highly reliable and efficient communication systems, leveraging their diversity gain, anti-interference characteristics, spectral efficiency advantages, and intelligent reconfiguration capability of wireless propagation environments. In MIMO-OFDM systems, critical functionalities such as signal recovery and demodulation, precoding, and beamforming, as well as the amplitude optimization and phase adjustment of RIS reflective units, heavily rely on accurate channel estimation. Therefore, the accurate and efficient acquisition of Channel State Information (CSI) is a critical prerequisite for the realization of these technologies. However, balancing estimation accuracy with reduced computational complexity and pilot resource consumption remains a significant challenge for the practical deployment of large-scale MIMO systems. To address these challenges, this thesis focuses on deep learning-based channel estimation methods for MIMO systems, with main contributions in the following two aspects:

        (1) To address the issues of low channel estimation accuracy and high pilot overhead in MIMO-OFDM communication systems, this thesis proposes a joint optimization scheme (AE-DRSN) based on pilot design and channel estimation. The scheme consists of the Concrete Auto-encoder (AE)-based pilot optimization module and the Deep Residual Shrinkage Network (DRSN)-based estimation module. The pilot optimization module searches for the most informative pilot positions in the time-frequency grid, and then the optimized pilot position information is input into the channel estimation module to achieve precise channel estimation. The two modules are jointly optimized through a combined loss function. Simulation results demonstrate that compared to traditional LS and MMSE algorithms, the proposed scheme achieves high-accurate channel estimation with significantly reduced pilot overhead, validating the effectiveness of the approach.

        (2) To address the challenge of low channel estimation accuracy and high computational complexity in RIS-assisted wireless communication systems, this thesis designs a Graph Neural Network (GNN)-based channel estimation scheme. The scheme effectively exploits node features and edge correlation information in the graph structure, utilizing the feature extraction capabilities of GNN to construct complex channel model, thereby effectively improving the cascaded channel estimation accuracy. The scheme utilizes Least Squares estimation to obtain the channel information of active elements, applies interpolation algorithms to compute the channel information of all RIS elements, and finally inputs the preliminary estimation results into the GNN for fine-grained reconstruction to achieve accurate estimation. Simulation results demonstrate that, compared to traditional CNN algorithms, the channel estimation performance of the proposed scheme improves by up to 4 dB.