首先定义如下MVDR(最小方差无畸变响应)算法的适用条件:远场,线阵,窄带
空间阵列模型:
假设接收了空间中有M个复信号,以不同的的角度射入阵列。接收天线的个数为N,则在t时刻阵列接收信号有如下表达式:
Y
(
t
)
N
×
1
=
∑
i
=
1
M
x
i
(
t
)
a
(
θ
i
)
+
N
oise
(
t
)
=
∑
i
=
1
M
x
i
(
t
)
[
1
,
e
−
j
2
π
d
sin
(
θ
)
λ
…
.
,
e
−
j
2
π
(
N
−
1
)
d
sin
(
θ
)
λ
]
N
×
1
+
N
oise
(
t
)
\begin{aligned} Y(\mathrm{t})_{N \times 1} &=\sum_{i=1}^{M} x_{i}(t) a\left(\theta_{i}\right)+N \text { oise }(\mathrm{t}) \\ &=\sum_{i=1}^{M} x_{i}(t)\left[1, e^{-j 2 \pi \frac{d \sin (\theta)}{\lambda}} \ldots ., e^{-j 2 \pi \frac{(N-1) d \sin (\theta)}{\lambda}}\right]_{N \times 1}+N \text { oise }(\mathrm{t}) \end{aligned}
Y(t)N×1=i=1∑Mxi(t)a(θi)+N oise (t)=i=1∑Mxi(t)[1,e−j2πλdsin(θ)….,e−j2πλ(N−1)dsin(θ)]N×1+N oise (t)
a
(
θ
)
a\left(\theta\right)
a(θ)为阵列导向向量,反应复信号到达不同天线的相位差/时延
假设我们只对
θ
d
\theta_{d}
θd方向的来波
x
d
(
t
)
x_{d}(t)
xd(t)感兴趣,并希望抑制其他方向
θ
j
\theta_{j}
θj的干扰信号
x
j
(
t
)
x_{j}(t)
xj(t)。
则阵列接收信号可以写为
Y
(
t
)
=
x
d
(
t
)
⋅
a
(
θ
d
)
+
∑
j
=
1
M
−
1
x
j
(
t
)
⋅
a
(
θ
j
)
+
N
(
t
)
Y(t)=x_{d}(t) \cdot a\left(\theta_{d}\right)+\sum_{j=1}^{M-1} x_{j}(t) \cdot a\left(\theta_{j}\right)+N(t)
Y(t)=xd(t)⋅a(θd)+j=1∑M−1xj(t)⋅a(θj)+N(t)
先假设有一复自适应权W作用于阵列接收信号,并希望经过自适应权后的输出
d
^
(
t
)
\hat{d}(t)
d^(t)尽可能逼近
x
d
(
t
)
x_{d}(t)
xd(t)。
d
^
(
t
)
1
×
1
=
W
N
×
1
H
⋅
Y
(
t
)
N
×
1
\hat{d}(t)_{1 \times 1}=W_{N \times 1}^{H} \cdot Y(t)_{N \times 1}
d^(t)1×1=WN×1H⋅Y(t)N×1
现在推导MVDR是如何实现这一目的的。
首先经过自适应滤波后,输出信号的平均功率可以写为:
P
=
E
(
∣
d
^
(
t
)
∣
2
)
=
E
(
d
^
(
t
)
∗
⋅
d
^
(
t
)
)
=
W
H
E
(
Y
∗
Y
T
)
W
=
W
H
(
a
(
θ
d
)
E
(
∣
x
d
(
t
)
∣
2
)
a
(
θ
d
)
H
+
∑
j
=
1
M
−
1
a
(
θ
j
)
E
(
∣
x
j
(
t
)
∣
2
)
a
(
θ
j
)
H
+
E
(
∣
N
(
t
)
∣
2
)
)
W
=
W
H
(
σ
d
2
a
(
θ
d
)
a
(
θ
d
)
H
+
∑
j
=
1
M
−
1
σ
j
2
a
(
θ
j
)
a
(
θ
j
)
H
+
σ
n
2
I
)
W
=
W
H
R
W
\begin{aligned} P &=E\left(|\hat{d}(t)|^{2}\right)=E\left(\hat{d}(t)^{*} \cdot \hat{d}(t)\right) \\ &=W^{H} E\left(Y^{*} Y^{T}\right) W^{} \\ &=W^{H}\left(a\left(\theta_{d}\right) E\left(\left|x_{d}(t)\right|^{2}\right) a\left(\theta_{d}\right)^{H}+\sum_{j=1}^{M-1} a\left(\theta_{j}\right) E\left(\left|x_{j}(t)\right|^{2}\right) a\left(\theta_{j}\right)^{H}+E\left(|N(t)|^{2}\right)\right) W \\ &=W^{H}\left(\sigma_{d}^{2} a\left(\theta_{d}\right) a\left(\theta_{d}\right)^{H}+\sum_{j=1}^{M-1} \sigma_{j}^{2} a\left(\theta_{j}\right) a\left(\theta_{j}\right)^{H}+\sigma_{n}^{2} I\right) W \\ &=W^{H} R W \end{aligned}
P=E(∣d^(t)∣2)=E(d^(t)∗⋅d^(t))=WHE(Y∗YT)W=WH(a(θd)E(∣xd(t)∣2)a(θd)H+j=1∑M−1a(θj)E(∣xj(t)∣2)a(θj)H+E(∣N(t)∣2))W=WH(σd2a(θd)a(θd)H+j=1∑M−1σj2a(θj)a(θj)H+σn2I)W=WHRW
为了保留感兴趣的信号,去掉干扰信号以及噪声,我们希望最小化输出平均功率,但是保持
W
H
a
(
θ
d
)
=
1
W^{H} a\left(\theta_{d}\right)=1
WHa(θd)=1
在上式第二行到第三行推导中,假设期望信号,干扰,噪声之间统计独立。
同时由于零均值信号的方差就是平均功率(零均值化是信号中的一般预处理操作),所以最小化平均功率就等于最小方差,这也是MVDR名字的含义。
介绍一些常见的复数的求导法则:
∂ w H R w ∂ w = R T w ∗ \frac{\partial \mathbf{w}^{H} \mathbf{R} \mathbf{w}}{\partial \mathbf{w}}=\mathbf{R}^{T} \mathbf{w}^{*} ∂w∂wHRw=RTw∗ ∂ w H R w ∂ w H = R w \frac{\partial \mathbf{w}^{H} \mathbf{R} \mathbf{w}}{\partial \mathbf{w}^{H}}=\mathbf{R}\mathbf{w} ∂wH∂wHRw=Rw ∂ Tr ( w w H ) ∂ w = w ∗ ∂ Tr ( w w H ) ∂ w ∗ = w \begin{aligned} &\frac{\partial \operatorname{Tr}\left(\mathbf{w} \mathbf{w}^{H}\right)}{\partial \mathbf{w}}=\mathbf{w}^{*} &\frac{\partial \operatorname{Tr}\left(\mathbf{w} \mathbf{w}^{H}\right)}{\partial \mathbf{w}^{*}}=\mathbf{w} \end{aligned} ∂w∂Tr(wwH)=w∗∂w∗∂Tr(wwH)=w ∂ Tr ( a H w ) ∂ w = a ∗ \frac{\partial \operatorname{Tr}(a^{H} w)}{\partial w}=a^{*} ∂w∂Tr(aHw)=a∗ ∂ Tr ( w H a ) ∂ w = 0 \frac{\partial \operatorname{Tr}(w^{H} a)}{\partial w}=0 ∂w∂Tr(wHa)=0 ∂ Tr ( w H a ) ∂ w H = a \frac{\partial \operatorname{Tr}(w^{H} a)}{\partial w^{H}}=a ∂wH∂Tr(wHa)=a
参考:Matrix Cookbook
知乎:关于复求导和实求导
知乎:复数矩阵求导的转置和共轭转置问题
stackproblem
推导最优权
为了普适性, θ d \theta_{d} θd写为 θ \theta θ
{
min
W
H
R
W
s
.
t
W
H
a
(
θ
)
=
1
,
a
H
(
θ
)
W
=
1
\left\{\begin{array}{l}\min \mathbf{W}^{\mathrm{H}} \mathbf{R} \mathbf{W} \\ \mathbf{s . t} \mathbf{W}^{\mathrm{H}} \mathbf{a}(\mathbf{\theta})=\mathbf{1}, \mathbf{a}^{\mathrm{H}}(\boldsymbol{\theta}) \mathbf{W}=\mathbf{1}\end{array}\right.
{minWHRWs.tWHa(θ)=1,aH(θ)W=1
W为最优权,R为阵列接收信号的协方差矩阵
构造拉格朗日函数:
L ( W ) = W H R W − λ ( W H a ( θ ) − 1 ) − λ ∗ ( a H ( θ ) W − 1 ) L(W)=W^{H} R W-\lambda\left(W^{H} a(\theta)-1\right)-\lambda^{*}\left(a^{H}(\theta) W-1\right) L(W)=WHRW−λ(WHa(θ)−1)−λ∗(aH(θ)W−1)
*表示共轭
对
W
H
W^{H}
WH 求复梯度并令其为零得:
∂
L
(
W
)
∂
W
H
=
R
W
−
λ
a
(
θ
)
−
0
=
0
\frac{\partial L(W)}{\partial W^{H}}= R W-\lambda a(\theta)-0=0
∂WH∂L(W)=RW−λa(θ)−0=0
可得
{
R
W
=
λ
a
(
θ
)
(
1
)
W
H
a
(
θ
)
=
1
(
2
)
\left\{\begin{array}{l} R W=\lambda a(\theta) (1)\\ W^{H} a(\theta)=1 (2) \end{array}\right.
{RW=λa(θ)(1)WHa(θ)=1(2)
由(1)得:
W
=
λ
R
−
1
a
(
θ
)
W=\lambda R^{-1} a(\theta)
W=λR−1a(θ)
两边同乘导向向量的共轭转置
a
H
(
θ
)
W
=
λ
a
H
(
θ
)
R
−
1
a
(
θ
)
a^{H}(\theta) W=\lambda a^{H}(\theta) R^{-1} a(\theta)
aH(θ)W=λaH(θ)R−1a(θ)
1
=
λ
a
H
(
θ
)
R
−
1
a
(
θ
)
\mathbf{1}=\lambda a^{H}(\theta) R^{-1} a(\theta)
1=λaH(θ)R−1a(θ)
λ
=
1
a
H
(
θ
)
R
−
1
a
(
θ
)
\lambda=\frac{1}{a^{H}(\theta) R^{-1} a(\theta)}
λ=aH(θ)R−1a(θ)1
将
λ
\lambda
λ带入上面可得最优权
W
o
p
t
=
R
−
1
a
(
θ
)
a
H
(
θ
)
R
−
1
a
(
θ
)
W_{\mathrm{opt}}=\frac{R^{-1} a(\theta)}{a^{H}(\theta) R^{-1} a(\theta)}
Wopt=aH(θ)R−1a(θ)R−1a(θ)
同时还可以引申出空间谱估计。将上式
W
o
p
t
W_{\mathrm{opt}}
Wopt带入P的表达式
W
H
R
W
\mathbf{W}^{\mathrm{H}} \mathbf{R} \mathbf{W}
WHRW可得
P
(
θ
)
=
1
a
(
θ
)
H
R
−
1
a
(
θ
)
=
σ
θ
P(\theta)=\frac{1}{a(\theta)^{H} R^{-1} a(\theta)}=\sigma_{\theta}
P(θ)=a(θ)HR−1a(θ)1=σθ