针对中性氢21cm信号的前景噪音扣除问题，本文主要开展深度学习方法应用研究。中性氢21cm谱线为观测宇宙大尺度结构提供了新的途径，可以还原宇宙演化历史信息，填补宇宙再电离时期的探测空白，帮助检验宇宙演化模型，进而揭示两暗（暗物质、暗能量）之谜。然而，中性氢21cm信号非常微弱，淹没在比其高出4-5个量级的宇宙前景噪音中，因此，前景噪音污染是探测中性氢21cm信号面临的主要挑战。同时传统的PCA方法会造成中性氢21cm信号的损失，导致中性氢21cm信号无法准确恢复，能否准确扣除前景噪音成为提取信号的关键。

针对此问题，本文发展了一套基于UNet神经网络模型的前景噪音扣除方法，并创新性提出基于频间相减图像的训练模式，为扣除前景噪音提供了新的思路。本文在第一章介绍了研究背景及意义，第二章介绍了所涉及到的主成分分析（PCA）和UNet神经网络模型的原理，在第三、四和五章，介绍了所开展的三个研究内容：

（1）开展中性氢巡天观测模拟研究：仿真构建21cm观测天图，系统性考虑各种前景噪音辐射机制，包括银河系同步辐射、自由-自由发射、分子发射线、尘埃辐射等，并基于平方公里阵列（SKA）望远镜加入仪器噪音的影响。

（2）开展UNet模型实验性研究：探索UNet模型扣除噪音的方法，系统性验证UNet模型对不同程度噪声的扣除能力，并与传统PCA方法的结果进行对比，同时围绕近几年提出的PCA-UNet结合方法进行验证性研究。

（3）发展基于UNet模型的新方法：提出基于频间相减图像的训练模式，利用21cm信号和前景噪音在不同频率间的相关性模式，即前者的随机性和后者的强相关性，频率间的温度相减图像可以有效去除大部分前景噪音，同时可以保留两者频率的信号信息，进而与纯21cm信号图像组成训练数据对，可以有效恢复信号图像。

本文研究发现，UNet模型可以恢复一定程度的21cm信号，但是在前景噪声的亮温度T>10²mK或者仪器噪声的标准差σ>0.2时，UNet模型的去噪能力将系统性减弱，无法胜任真实观测条件下的去噪任务；PCA-UNet的结合方法仅能恢复较好的21cm信号自相关功率谱，但是在交叉相关功率谱方面能力欠佳。相比而言，基于频间相减的图像，前景噪声的亮温度量级降低至10⁰-10¹mK之间，UNet模型重建的21cm信号与真实21cm信号之间的交叉相关功率谱比值达到0.98左右，包含仪器效应的交叉相关比值在大尺度上可以达到0.95，因此，相比传统方法，频间相减图像与纯21cm图像组成的训练数据对，可以有效提升UNet模型效果。

To address the problem of foreground noise subtraction of the neutral hydrogen 21cm signal, this thesis focuses on conducting research on the application of machine learning methods. The neutral hydrogen 21cm spectrum provides a new way to observe the large-scale structure of the universe, which can restore information on the evolutionary history of the universe, fill the detection gap of the reionization period of the universe, help to test the evolutionary model of the universe, and then reveal the mystery of the two darks (dark matter and dark energy). However, the temperature of the neutral hydrogen 21cm signal is very faint and drowned in the cosmic foreground noise which is 4-5 orders of magnitude higher than it. Therefore, foreground noise contamination is the main challenge to detecting the neutral hydrogen 21cm signal. Meanwhile, the traditional PCA method will cause the loss of neutral hydrogen 21cm signal, resulting in the neutral hydrogen 21cm signal not being recovered accurately, so whether the foreground noise can be accurately deducted becomes the key to extracting the signal.

To address this problem, this thesis develops a foreground noise subtraction method based on the UNet neural network model and innovatively proposes a training model based on inter-frequency phase subtraction images, which provides a new idea for subtracting foreground noise. This paper introduces the background and significance of the research in Chapter 1, the principles of the two computational methods involved, principal component analysis (PCA) and UNet neural network model, in Chapter 2, and in Chapters 3, 4, and 5, the main research carried out includes:

(1) Simulation of neutral hydrogen survey observations: simulation to construct a 21-cm observation sky map, systematic consideration of various foreground noise radiation mechanisms, including Galactic synchrotron radiation, free-free emission, molecular emission lines, dust radiation, etc., and inclusion of the effect of instrument noise based on the Square Kilometer Array (SKA) telescope.

(2) Research on the experimental of the UNet model: explore the methods of noise deduction by the UNet model, systematically verify the ability of the UNet model to deduct different levels of noise, compare the results with those of traditional PCA methods, and conduct validation studies around the combined PCA-UNet method proposed in recent years.

(3) Development of a new method based on the UNet model: a training model based on inter-frequency phase subtracted images is proposed, using the correlation pattern between 21cm signal and foreground noise at different frequencies, i.e., the randomness of the former and the strong correlation of the latter, and the inter-frequency temperature phase subtracted images can effectively remove most of the foreground noise, while the signal information of both frequencies can be retained, which in turn can be composed with pure 21cm signal images. The training data pair can effectively recover the signal image.

In this thesis, we find that the UNet model can recover a certain degree of 21cm signal, but at the bright temperature T>10²mK of foreground noise or the standard deviation σ>0.2 of the instrumental noise, the denoising ability of the UNet model will be systematically weakened and cannot perform the task of denoising under real observation conditions, the combined PCA-UNet method can only recover a better. The combined PCA-UNet method can only recover the autocorrelation power spectrum of the 21cm signal but has poor capability in the cross-correlation power spectrum. In contrast, the bright temperature magnitude of the foreground noise is reduced to between 10⁰-10¹mK for inter-frequency phase subtracted-based images, and the cross-correlation power spectrum ratio between the 21cm signal reconstructed by the UNet model and the real 21cm signal reaches about 0.98, and the cross-correlation ratio including instrumental effects can reach 0.95 on large scales, thus, compared with the conventional method. The training data pair consisting of inter-frequency subtracted images and pure 21cm images can effectively improve the UNet model effect.