模型
输入输出为平稳随机过程,求最优权向量使得输出估计结果的均方误差最小
d
^
(
n
)
=
w
H
u
(
n
)
m
i
n
{
J
(
w
)
{
e
(
n
)
=
d
(
n
)
−
d
^
(
n
)
J
(
w
)
=
=
E
{
∣
e
(
n
)
∣
2
}
\begin {align} &\hat{d}(n)=w^Hu(n) \\ &min\{J(w)\\ &\left\{\begin{aligned} &e(n)=d(n)-\hat{d}(n) \\ &J(w)==E\{|e(n)|^2\} \end{aligned} \right. \end {align}
d^(n)=wHu(n)min{J(w){e(n)=d(n)−d^(n)J(w)==E{∣e(n)∣2}
方法
最优权向量:
Wiener-Hopf方程:
R
w
o
=
p
R
=
E
{
u
(
n
)
u
H
(
n
)
}
=
[
r
(
0
)
r
(
1
)
⋯
r
(
M
−
1
)
r
(
−
1
)
r
(
0
)
⋯
r
(
M
−
2
)
⋮
⋯
⋱
⋯
r
(
−
M
+
1
)
⋯
⋯
r
(
0
)
]
p
=
E
{
u
(
n
)
d
∗
(
n
)
}
=
[
E
{
u
(
n
)
d
∗
(
n
)
}
E
{
u
(
n
−
1
)
d
∗
(
n
)
}
⋮
E
{
u
(
n
−
M
+
1
)
d
∗
(
n
)
}
]
=
[
p
(
0
)
p
(
−
1
)
⋮
p
(
−
M
+
1
)
]
\begin {align} &Rw_o=p\\ &R=E\{u(n)u^H(n)\}=\begin{bmatrix} r(0) & r(1) & \cdots & r(M-1) \\r(-1)&r(0) & \cdots & r(M-2) \\ \vdots &\cdots &\ddots &\cdots \\ r(-M+1)&\cdots&\cdots&r(0)\end{bmatrix}\\ &p=E\{u(n)d^*(n)\}=\begin{bmatrix} E\{u(n)d^*(n)\} \\ E\{u(n-1)d^*(n)\}\\ \vdots\\ E\{u(n-M+1)d^*(n)\} \end{bmatrix} =\begin{bmatrix} p(0)\\ p(-1)\\ \vdots\\ p(-M+1) \end{bmatrix} \end {align}
Rwo=pR=E{u(n)uH(n)}=
r(0)r(−1)⋮r(−M+1)r(1)r(0)⋯⋯⋯⋯⋱⋯r(M−1)r(M−2)⋯r(0)
p=E{u(n)d∗(n)}=
E{u(n)d∗(n)}E{u(n−1)d∗(n)}⋮E{u(n−M+1)d∗(n)}
=
p(0)p(−1)⋮p(−M+1)
算法
1.最陡下降算法
(1)核心
Δ w = − 1 2 μ Δ J ( w ( n ) ) \Delta w = -\frac{1}{2}\mu\Delta J(w(n)) Δw=−21μΔJ(w(n))
(2)计算
{ w ( n + 1 ) = w ( n ) + μ [ p − R w ( n ) ] 0 < μ < 2 λ m a x \begin{align} \left\{\begin{aligned} &w(n+1)=w(n)+\mu [p-Rw(n)]\\ &0<\mu<\frac{2}{\lambda_{max}} \end{aligned} \right. \end{align} ⎩ ⎨ ⎧w(n+1)=w(n)+μ[p−Rw(n)]0<μ<λmax2
(3)性能
当
0
<
μ
<
2
λ
m
a
x
0<\mu<\frac{2}{\lambda_{max}}
0<μ<λmax2时,最陡下降算法收敛:
lim
n
−
>
∞
w
(
n
)
=
w
o
\lim_{n->\infty} w(n)=w_o
n−>∞limw(n)=wo
2.LMS算法
(1)核心
R ^ = 1 N ∑ i = 1 N u ( i ) u H ( i ) p ^ = 1 N ∑ i = 1 N u ( i ) d ∗ ( i ) \begin{align} &\hat{R}=\frac{1}{N}\sum_{i=1}^{N}u(i)u^H(i)\\ &\hat{p}=\frac{1}{N}\sum_{i=1}^{N}u(i)d^*(i) \end{align} R^=N1i=1∑Nu(i)uH(i)p^=N1i=1∑Nu(i)d∗(i)
(2)计算
w ^ ( n + 1 ) = w ^ ( n ) + μ u ( n ) e ∗ ( n ) \hat{w}(n+1) = \hat{w}(n)+\mu u(n)e^*(n) w^(n+1)=w^(n)+μu(n)e∗(n)
(3)性能
滤波器权向量1阶收敛,但均方误差大于最小均方误差
1阶收敛
当
0
<
μ
<
2
λ
m
a
x
0<\mu<\frac{2}{\lambda_{max}}
0<μ<λmax2时,最陡下降算法收敛:
lim
n
−
>
∞
E
{
w
^
(
n
)
}
=
w
o
\lim_{n->\infty} E\{\hat{w}(n)\}=w_o
n−>∞limE{w^(n)}=wo
均方误差的稳态性能
步长因子满足
0
<
μ
<
2
∑
i
=
1
M
λ
i
=
2
M
r
(
0
)
0<\mu<\frac{2}{\sum\limits_{i=1}^{M}{\lambda_i}}=\frac{2}{Mr(0)}
0<μ<i=1∑Mλi2=Mr(0)2
a.剩余均方误差 J e x J_{ex} Jex
J e x ( n ) = E J ^ ( n ) − J m i n J e x ( ∞ ) ≈ μ J m i n ∑ i = 1 M λ i 2 − μ ∑ i = 1 M λ i \begin{align} &J_{ex}(n) = E{\hat{J}(n)}-J_{min}\\ &J_{ex}(\infty)\approx \mu J_{min}\frac{\sum\limits^{M}_{i=1}\lambda_i}{2-\mu\sum\limits_{i=1}^{M}{\lambda_i}} \end{align} Jex(n)=EJ^(n)−JminJex(∞)≈μJmin2−μi=1∑Mλii=1∑Mλi
b.失调参数 M M M
M = J e x ( ∞ ) J m i n M = μ ∑ i = 1 M λ i 2 − μ ∑ i = 1 M λ i ≈ μ 2 ∑ i = 1 M λ i = μ 2 M r ( 0 ) = μ 2 M λ a v = M 4 τ a v \begin{align} &M=\frac{J_{ex}(\infty)}{J_{min}}\\ &M=\frac{\mu \sum\limits^{M}_{i=1}\lambda_i}{2-\mu\sum\limits_{i=1}^{M}{\lambda_i}} \approx\frac{\mu}{2}\sum\limits_{i=1}^{M}\lambda_i=\frac{\mu}{2}Mr(0)=\frac{\mu}{2}M\lambda_{av}=\frac{M}{4\tau_{av}} \end{align} M=JminJex(∞)M=2−μi=1∑Mλiμi=1∑Mλi≈2μi=1∑Mλi=2μMr(0)=2μMλav=4τavM
c.平均时间常数 τ a v \tau_{av} τav
τ a v = 1 2 μ λ a v λ a v = 1 M ∑ i = 1 M λ i \tau_{av}=\frac{1}{2\mu\lambda_{av}} \lambda_{av}=\frac{1}{M}\sum\limits_{i=1}^{M}\lambda_i τav=2μλav1λav=M1i=1∑Mλi
d.特征值拓展/特征值比 χ \chi χ
χ ( R ) = λ m a x λ m i n \chi(R)=\frac{\lambda_{max}}{\lambda_{min}} χ(R)=λminλmax