# 简介

## 问题背景

• 轨道上的架空电线杆等物体会周期性地阻塞直射路径（LOS），这会引起信道的不断变化，也就意味着更短的相干时间。
• 高速的列车运动同样会带来不可忽略的多普勒频移，从而导致了额外的载波间干扰 (ICI).

In contrast to urban cellular systems, many challenging problems exist inHSRscenarios. The electromagnetic shielding from the carriage results in a strong penetration loss of signals, and it is difficult to build a direct link between the base station (BS) and users [12], [13]. The overhead line poles along the rail periodically block the line-of-sight (LoS) path and act as a static and periodic scatters [14]. The high mobility of trains not only results in a sharp decrease in coherence time (CT) but also leads to a nonnegligible Doppler shift, which causes intercarrier interference (ICI) and serious performance loss.

## 文章贡献

• 通过FFT来快速获取AOA信息， 并通过波束校准来补偿由多普勒频移带来的DFO。
• 用卡尔曼滤波器 (KF-filter)来自适应信道的时变
• 提出了一种低复杂度的hybrid beamforming算法。
1. An AD beam alignment method is applied to acquire a precise AOA, which can be obtained by applying a fast Fourier transform (FFT) and phase rotation. By utilizing beam alignment, DFO can be directly compensated. 2) Based on temporal coherence, a KF-based multipath channel tracking scheme is proposed, which is self-adaptive to CSI variance. The tracking scheme is further expanded to the AD to significantly decrease the pilot overhead and computational complexity. 3) In the data transmission phase, a hybrid beamforming scheme is proposed to compress the channel dimension and keep low complexity, which is composed of DFO compensation, angular beamforming, and BD precoding. Compared with the linear method, the hybrid beamforming scheme requires considerably less computations while retaining high throughput.

# 系统模型

$\mathbf{h}_{k, n, f_{D}}=e^{j \phi_{n}} \sum_{l=1}^{L} e^{j \phi_{f_{D, l, n}}} \cdot \psi_{l} \alpha_{k, l, n} \mathbf{a}\left(\theta_{k, l, n}\right) \odot \mathbf{b}_{k, l, n}$

• $\phi_{n}$: the phase caused by the carrier frequency.
• $\phi_{f_{D}, l, n}$: 由多普勒频移引起的相位偏移绝对量. absolute phase offset resulting from Doppler shift. 其中 $\phi_{f_{D, l, n}}=2 \pi f_{D, l, n} n T$$T$是一个相干时间的长度。多普勒频移为 $f_{D, l, n}=\frac{v}{\lambda} \cos \left(\theta_{k, l, n}\right)$$v$代表动车速度。$\theta k, l, n$代表了第$l$条路径的AOA角度，图中有画出。
• $\psi_{l}$代表权重系数（LOS和NLOS的比重不同）: $l$=1代表LOS路径，$\mathcal{K}$代表
$\psi_{l}=\left\{\begin{array}{ll}{\sqrt{\frac{\kappa}{\kappa+1}},} & {l=1} \\ {\sqrt{\frac{1}{(\mathcal{K}+1)(L-1)}},} & {l \geq 2}\end{array}\right.$
• $\alpha_{k, l, n}$$l$径的信道路径增益，服从0均值高斯分布
• $\mathbf{a}\left(\theta_{k, l, n}\right)=\left[1, e^{j 2 \pi \frac{d}{\lambda} \cos \left(\theta_{k, l, n}\right)}, \ldots, e^{j 2 \pi(M-1) \frac{d}{\lambda} \cos \left(\theta_{k, l, n}\right)}\right]$
• $\mathbf{b}_{k, l, n}$： 相对多普勒频移， $\left[1, e^{j 2 \pi f_{D, l} t}, \cdots, e^{j 2 \pi(M-1) f_{D, l} t}\right]^{T}$$t=\frac{d \cos \left(\theta_{k, l, n}\right)}{c}$， 代表由于不同天线到同一终端的距离差引起的传播时间差异， $d$代表天线间隔。

# 多普勒频移补偿和波束分配

$\mathbf{h}_{k, n, f_{D}}=\sum_{l=1}^{L} e^{j\left(\phi_{n}+\phi_{f_{D}, l, n}\right)} \cdot \psi_{l} \alpha_{k, l, n} \mathbf{a}^{\prime}\left(s_{l, n}\right)$

$s_{l, n}=\frac{d}{\lambda}\left(1+\frac{v}{c} \cos \left(\theta_{k, l, n}\right)\right) \cos \left(\theta_{k, l, n}\right), \mathbf{a}^{\prime}\left(x\right)=\left[1, e^{j 2 \pi x}, \ldots, e^{j 2 \pi(M-1)x}\right]$.

$\mathbf{h}_{k, n, f_{D}}=\sum_{l=1}^{L} e^{j\left(\phi_{n}+\phi_{f_{D}, l, n}\right)} \cdot \psi_{l} \alpha_{k, l, n} \mathbf{a}\left(\theta^{\prime} _{k, l, n}\right)$,

# 通过模拟波束成形补偿多普勒频移

• 由于绝对多普勒频移是固定值，可以简单的用偏转量的共轭值补偿。
$\mathbf{\Phi}_{n, u}=\operatorname{diag}\left(e^{j \phi_{f_{D_{1}, 1, n}}, \ldots, e^{j \phi_{f_{D_{K}, 1, n}}}}\right)$
• 把相对多普勒频移等效为到达角的偏移量， 就可以通过对实际到达角进行波束成形，从而无形中补偿了相对多普勒频移，下面就对此进行展开。
• 如前所述，本节中作者只考虑LOS信道来简化操作，即只靠虑 $l=1$径。

## 通过傅里叶变换获取信道能量角度域分布

$\mathbf{a}\left(\theta\right)=\left[1, e^{j 2 \pi \frac{d}{\lambda} \cos \left(\theta\right)}, \ldots, e^{j 2 \pi(M-1) \frac{d}{\lambda} \cos \left(\theta\right)}\right]$,

$\mathbf{y}_{k, n}=\mathbf{h}_{k, n, f_{D}} \mathbf{x}_{k}+\mathbf{n}_{\mathbf{k}, \mathbf{n}}$

$\mathbf{F}=\frac{1}{\sqrt{M}}\left[\begin{array}{cccc}{1} & {1} & {\cdots} & {1} \\ {1} & {e^{-j 2 \pi \Delta f}} & {\dots} & {e^{-j 2 \pi(M-1) \Delta f}} \\ {\vdots} & {\vdots} & {\vdots} & {\vdots} \\ {1} & {e^{-j 2 \pi(M-1) \Delta f}} & {\dots} & {e^{-j 2 \pi(M-1)^{2} \Delta f}}\end{array}\right]$

$\tilde{\mathbf{h}}_{k, n, f_{D}}=\mathbf{F} \hat{\mathbf{h}}_{k, n, f_{D}}$

$\mathbf{O}\left(\phi_{k, n, u}\right)=\operatorname{diag}\left(1, e^{j \phi_{k, n, u}}, \ldots, e^{j(M-1) \phi_{k, n, u}}\right)$

$\tilde{\mathbf{h}}_{k, n, f_{D}}^{r}=\mathbf{F} \mathbf{O}\left(\phi_{k, n, u}\right) \hat{\mathbf{h}}_{k, n, f_{D}}$

• 使用过采样傅里叶变换，如做NM点傅里叶变换，可以得到更高的角度分辨率。比如我们发现$\tilde{\mathbf{h}}_{k, n, f_{D}}^{r}$的第i个元素最大，那么就对其周围进行过采样FFT，从而提升精度，得到$\phi_{k, n, u}^{0}$ (不是全局最优，但是比只用FFT矩阵精度更高）。
• 通过和差化积等推导，可以得到:
$\frac{2 \pi\left(i_{k}-1\right)}{M}-2 \pi s_{1, n}-\phi_{k, n, u}^{0}=0$
其中 $\phi_{k, n, u}^{0}$ 可以由收到的训练序列求最大值得到，代入可以求得 $s_{1, n}$，从而求得到达角信息，相当于做好了信道估计。具体推导过程可参阅论文。

# 用卡尔曼滤波来追踪信道时变

$\mathbf{h}_{k, n}=\sum_{l=1}^{L} \psi_{l} \alpha_{k, l, n} \mathbf{a}\left(\theta_{k, l, n}^{\prime}\right)$

## AOA到达角

$\tan \left(\theta^{\prime}_{k, 1, n}+\pi \Delta f\right)=d_{0} /\left(d_{m}-v N_{m} T\right)$

$N_{m}=\left\{\begin{array}{c}{\frac{1}{v T}\left(d_{m}-\frac{d_{0}}{\tan \left(\theta^{\prime} k, n+\pi \Delta f\right)}\right), \theta^{\prime}_{k, 1, n} \in\left(\theta_{1}, \frac{\pi}{2}\right]} \\ {\frac{1}{v T}\left(\frac{d_{0}}{\tan \left(\pi-\theta^{\prime} k_{1, n}-\pi \Delta f\right)}-d_{m}\right), \theta_{k, 1, n}^{\prime} \in\left(\frac{\pi}{2}, \theta_{2}\right)}\end{array}\right.$

## 信道增益

$\alpha_{k, l, n+1}=\rho_{l} \alpha_{k, l, n}+\sqrt{1-\rho^{2}} \beta_{k, l, n}$

$\rho_l$: 时间相关系数，表示相邻相关时间之间相关的方差，是一个常数。
$\beta_{k, l, n}$：一个独立分布的高斯变量，$\beta_{k, l, n} \sim \mathcal{C} \mathcal{N}\left(0, \sigma_{k, l, n}^{2}\right)$

$\rho_{l}=\frac{\mathcal{K}}{\mathcal{K}+1} \exp \left[j 2 \pi f_{D} T \cos (\gamma)+j \frac{2 \pi d}{\lambda} \cos (\alpha)\right]$

## 使用卡尔曼滤波追踪信道时变

$\left(\tilde{\mathbf{h}}_{k, n+1}^{r_{0}}\right)_{i_{k}}=\rho\left(\tilde{\mathbf{h}}_{k, n}^{r_{0}}\right)_{i_{k}}+\sqrt{1-\rho^{2}}\left(\tilde{\mathbf{c}}_{k, n+1}^{r_{0}}\right)_{i_{k}}$

$\begin{array}{l}{\left(\tilde{\mathbf{h}}_{k, n+1 | n}^{r_{0}}\right)_{i_{k}}=\rho\left(\tilde{\mathbf{h}}_{k, n | n}^{r_{0}}\right)_{i_{k}}} \\ {p_{k, n+1 | n}=\rho^{2} p_{k, n | n}+\left(1-\rho^{2}\right) \mathbf{f}_{i_{k}}^{T} \mathbf{O}\left(\phi_{k, n, u}^{0}\right) \mathbf{R}_{k, n} \mathbf{O}^{H}\left(\phi_{k, n, u}^{0}\right) \mathbf{f}_{i k}^{*}} \\ {\left(\tilde{\mathbf{h}}_{k, n+1 | n+1}^{r_{0}}\right)_{i_{k}}=\left(\tilde{\mathbf{h}}_{k, n+1 | n}^{r_{0}}\right)_{i_{k}}+k g\left(\tilde{\mathbf{y}}_{k, n+1}^{\prime}-\tau p_{\tau}\left(\tilde{\mathbf{h}}_{k, n+1}^{r_{0}}\right)_{i_{k}}\right)} \\ {k g=\frac{p_{k, n+1 | n}}{\tau p_{\tau} p_{k, n+1 | n}+\sigma_{n}^{2}}} \\ {p_{k, n+1 | n+1}=\left(1-k g \tau p_{\tau}\right) p_{k, n+1 | n}}\end{array}$

# 混合波束成形

$\mathbf{P}_{n, u}=\tilde{\mathbf{H}}_{n}^{\left\{B_{r 0}\right\}}\left[\left(\tilde{\mathbf{H}}_{n}^{\left\{B_{r 0}\right\}}\right)^{H} \tilde{\mathbf{H}}_{n}^{\left\{B_{r 0}\right\}}\right]^{-1}$

# 疑问

To retain the orthogonality of users, the length of the pilot sequence for a time-division duplex (TDD) mMIMO system should be larger than the number of users, whereas for a frequency-division duplex system, the length will be even larger than the number of antennas, which is unacceptable [32].