大规模多输入多输出（Multiple Input Multiple Output，MIMO）是第五代移动通信的关键技术之一，可在不占用额外频谱资源的前提下扩大信道容量，提高频谱效率与能量效率。为充分利用大规模MIMO的优势，基站端需要获取尽可能精确的下行信道状态信息（Channel State Information，CSI）从而完成波束赋形。在频分双工系统中，上下行信道不具备互易性，因此下行CSI需要由用户端进行信道估计并反馈给基站端。然而在大规模MIMO系统中，CSI的复杂度随着天线规模的增加而显著增加，极大提升了信道反馈的难度和开销。本论文基于深度学习技术，围绕大规模MIMO系统信道反馈这一难点问题，提出了一系列创新性的智能优化方案，服务于反馈精度与系统性的提升和反馈开销与复杂度的降低。首先，本文提出了基于多分辨率思想和聚合卷积思想的反馈网络性能优化方法，可提升信道反馈的精度。基于CSI特征尺度多样化的特点，本文设计了多分辨率反馈网络CRNet。相比前序工作CsiNet，所提CRNet以更低的复杂度实现了更高的反馈精度；基于CSI特征形态多样化的特点，本文设计了聚合反馈网络ACRNet，实现了论文公开时间点上的最佳反馈性能。其次，本文提出了基于网络二值化和码字模仿学习的反馈网络部署优化方法，可实现低成本部署。基于对反馈网络复杂度的综合分析，本文提出了反馈网络全连接层二值化的策略。所提超轻量级反馈网络BCsiNet可在资源敏感的用户端实现30倍以上的参数量压缩；更进一步，为减少反馈网络轻量化后的性能损失，本文将特征蒸馏思想引入信道反馈。所提出的码字模仿策略和模仿-探索两阶段训练策略可在不增加部署成本的前提下提升反馈性能。最后，本文提出了子连接结构下和广义空间调制下的信道估计、信道反馈与波束赋形联合优化方法，可提高系统频谱效率。为解决分模块独立优化算法系统性不足的问题，本文考虑了信道估计、信道反馈与波束赋形的端到端联合优化。所提出的子连接结构下联合优化网络EFBAttnNet和广义空间调制下联合优化网络GsmEFBNet均实现了领先于传统分模块独立优化算法的频谱效率。综上，在大规模MIMO信道反馈问题上，本文所提系列智能优化方案有效提升了反馈精度，降低了反馈复杂度，并通过多模块联合优化提升了反馈设计的系统性，从而有力推动了未来通信系统中智能信道反馈技术的发展和应用。

Massive multiple-input multiple-output (MIMO) is one of the key technologies of the fifth-generation (5G) wireless communication systems. It can expand channel capacity and improve spectrum efficiency and energy efficiency without occupying additional spectrum resources. In order to fully utilize the advantages of massive MIMO, the base station (BS) needs to obtain accurate downlink channel state information (CSI) for beamforming design. In frequency division duplexing (FDD) systems, the uplink and downlink channels do not have reciprocity, so the downlink CSI needs to be estimated at the user equipment (UE) and fed back to the BS. However, the complexity of CSI increases significantly as the number of antennas increases in massive MIMO systems. The complicated CSI greatly increases the difficulty and overhead of the feedback. In this paper, a series of innovative intelligent optimization schemes are proposed around the challenging CSI feedback task in massive MIMO systems. The feedback accuracy and system integrity are improved while the feedback overhead and complexity are reduced.Firstly, this paper proposes feedback network performance optimization methods based on multi-resolution strategy and aggregated convolution strategy, which can improve the accuracy of CSI feedback. Considering the diversity of the CSI characteristic scales, a multi-resolution network named channel reconstruction network (CRNet) is designed. Compared with the previous state-of-the-art (SOTA) network CsiNet, the proposed CRNet significantly improves the feedback accuracy with less computational complexity. Besides, considering the diversity of the CSI characteristic patterns, a novel network named aggregated channel reconstruction network (ACRNet) is designed, achieving SOTA feedback performance at the time of publication of the paper.Secondly, this paper proposes feedback network deployment optimization methods based on network binarization and codeword mimic learning, which can achieve low-cost deployment. Based on a comprehensive analysis of the complexity of the feedback network, this paper proposes a binarization strategy for the fully connected layers of feedback networks. An extremely lightweight feedback network named binary CsiNet (BCsiNet) is designed, so that the network parameters are compressed by more than 30 times at the resource-sensitive UE. Moreover, the feature distillation strategy is introduced to the CSI feedback task to alleviate the performance loss brought by network lightweighting. In particular, the codeword mimic learning and the mimic-explore two-stage training strategy are carefully designed to improve the feedback capacity of the lightweight network without any extra inference cost.Finally, this paper proposes a joint optimization method of channel estimation, feedback and beamforming under subarray structure and generalized spatial modulation scheme, which can improve the system spectral efficiency. In order to improve the system integrity of the module-wise independent optimization scheme, an end-to-end joint optimization of channel estimation, feedback and beamforming is considered. The proposed joint optimization network EFBAttnNet under the sub-connection structure and the joint optimization network GsmEFBNet under the generalized spatial modulation scheme both achieve higher spectral efficiency compared with traditional block-based communication systems whose modules are independently optimized.In conclusion, a series of DL-aided algorithms are proposed in this paper for the massive MIMO CSI feedback task. With the help of these innovative algorithms, the feedback accuracy is improved, the feedback computational complexity is reduced, and the feedback system integrity is enhanced with the multi-module joint optimization. Overall, this paper promotes the development and application of DL-aided CSI feedback technology in future communication systems.