随着移动互联网和物联网技术的快速发展，移动通信系统的应用场景对能量效率、数据传输速率、频谱效率等各项性能指标提出了更高的要求。多输入多输出（Multiple Input Multiple Output，MIMO）结合正交频分复用（Orthogonal Frequency Division Multiplexing，OFDM）和智能反射面（Intelligent Reflecting Surface，IRS）等关键技术则在更高效、更可靠的数据传输方面具有重要的性能优势。而信道估计是发挥MIMO-OFDM和IRS等技术潜力的先决条件，也是实现预编码、信道均衡、资源分配、信号检测、室内定位、物理层安全等技术的重要基础。信道估计精度和算法复杂度是评价信道估计的重要性能指标，因此，设计高精度、低复杂度以及较小开销的信道估计算法具有重要的研究意义，基于此，本文的主要工作如下：

针对MIMO毫米波系统的信道估计问题，为了有效降低导频开销和提高信道估计精度，本文设计了一种基于压缩感知结合深度学习的信道估计算法（CS-DnCNN）。该算法首先通过设计感知矩阵进行导频优化以减少导频开销，然后利用压缩感知重构算法中的基追踪降噪（Basis Pursuit De-Noising，BPDN）算法进行信道初步估计，最后通过去噪卷积神经网络（Denoising Convolutional Neural Network，DnCNN）进行去噪得到信道精确估计。仿真结果表明，设计的CS-DnCNN算法的估计性能优于最小二乘（Least Square，LS）、正交匹配追踪（Orthogonal Matching Pursuit，OMP）、BPDN算法。本文所设计的算法对噪声具有较高的鲁棒性，提高了信道状态信息的估计精度。

针对IRS系统复杂度较高的问题，本文设计了一种基于超分辨率卷积神经网络（Super Resolution Convolutional Neural Network，SRCNN）的IRS辅助系统的信道估计算法（Chan-SRCNN）。该算法将信道建模为低分辨率图像到高分辨率图像的恢复问题，采用了混合有源/无源IRS架构，首先利用LS算法获取有源元件的信道估计值，再通过插值得到信道初步估计，最后利用Chan-SRCNN深度学习网络将其恢复为信道精确估计。仿真结果表明，所设计的算法的归一化均方误差（Normalized Mean Squared Error，NMSE）性能优于LS、OMP、同步正交匹配追踪（Simultaneous Orthogonal Matching Pursuit，SOMP）、深度神经网络（Deep Neural Network，DNN）等信道估计算法，证明了方案的可行性。

针对IRS辅助系统在信道估计时，估计精度较低的问题，本文进一步改进了Chan-SRCNN算法，设计了一种基于Wasserstein生成对抗网络（Wasserstein Generative Adversarial Network，WGAN）的信道估计算法（Chan-SRWGAN）。该算法将改进的Chan-SRCNN网络结构作为生成器模型，DNN作为判别器模型。此外，本文进一步优化了WGAN-GP的损失函数，设计了联合损失函数。通过改进网络结构和设计损失函数来提高信道估计的精度。仿真结果表明，所设计的算法性能优于LS、OMP、SOMP、DNN以及Chan-SRCNN等信道估计算法，适用于更复杂的应用场景。

With the rapid development of mobile Internet and Internet of Things technology, the application scenarios of mobile communication systems put forward higher requirements for energy efficiency, data transmission rate, spectrum efficiency and other performance indicators. Key technologies such as MIMO-OFDM and intelligent reflecting surface (IRS) have important performance advantages in more efficient and reliable data transmission. Channel estimation is a prerequisite for the potential of MIMO-OFDM and smart reflectors, and it is also an important basis for the realization of precoding, channel equalization, resource allocation, signal detection, indoor positioning, physical layer security and other technologies. Channel estimation accuracy and algorithm complexity are important performance indicators for evaluating channel estimation. Therefore, it is of great significance to design channel estimation algorithms with high accuracy, low complexity and low overhead. Based on this, The main work of this paper is as follows:

Aiming at the channel estimation problem of MIMO millimeter wave system, in order to effectively reduce the pilot overhead and improve the channel estimation accuracy. In this paper, we design a channel estimation algorithm (CS-DnCNN) based on compressed sensing combined with deep learning. The algorithm firstly reduces the pilot overhead by designing the sense matrix for pilot optimization, then uses the basis pursuit de-noising (BPDN) algorithm in the compressed sense reconstruction algorithm for the initial estimation of the channel, and then uses the denoising convolutional neural network (DnCNN) for the initial estimation of the channel. Network for denoising to get the accurate channel estimation. Simulation results show that the estimation performance of the designed CS-DnCNN algorithm is better than least square (LS), orthogonal matching pursuit (OMP), and BPDN algorithms. The algorithm designed in this paper has high robustness to noise and improves the estimation accuracy of channel state information.

Aiming at the problem of high complexity of IRS system, this paper designs a channel estimation algorithm (Chan-SRCNN) based on super resolution convolutional neural network (SRCNN) for IRS auxiliary system. The algorithm models the channel as a low resolution image to high resolution image recovery problem and adopts a hybrid active/passive IRS architecture, which firstly uses the LS algorithm to obtain the channel estimates of the active element, then interpolates them to obtain the preliminary estimates of the channel, and finally recovers them into the exact estimates of the channel using the Chan-SRCNN deep learning network. The simulation results show that the normalized mean squared error (NMSE) performance of the designed algorithm is better than that of LS, OMP, simultaneous orthogonal matching pursuit (SOMP) and deep neural network (DNN) channel estimation algorithms, which proves the feasibility of the scheme.

Aiming at the problem of low estimation accuracy of IRS auxiliary system during channel estimation, this paper further improves the Chan-SRCNN algorithm and designs a channel estimation algorithm based on wasserstein generative adversarial network (WGAN) (Chan- SRWGAN). The algorithm uses the improved Chan-SRCNN network structure as the generator model and DNN as the discriminator model. In addition, this paper further optimizes the loss function of WGAN-GP and designs the joint loss function. The accuracy of channel estimation is improved by improving the network structure and designing the loss function. The simulation results show that the performance of the proposed algorithm is better than that of LS, OMP, SOMP, DNN and Chan-SRCNN channel estimation algorithms, which is suitable for more complex application scenarios.