连续时间
降阶的LQG控制问题几乎和全阶的LQG控制 问题相同。令
x
^
r
(
t
)
{\displaystyle {\hat {\mathbf {x} }}_{r}(t)}
表示降阶LQG控制器的状态,唯一的差异是LQG控制器的状态维度
n
r
=
d
i
m
(
x
^
r
(
t
)
)
{\displaystyle n_{r}=dim({\hat {\mathbf {x} }}_{r}(t))}
是事先定义好的值,比受控系统的状态维度
n
=
d
i
m
(
x
(
t
)
)
{\displaystyle n=dim({\mathbf {x} }(t))}
要少。
降阶LQG控制器可以表示为下式:
x
^
˙
r
(
t
)
=
A
r
(
t
)
x
^
r
(
t
)
+
B
r
(
t
)
u
(
t
)
+
K
r
(
t
)
(
y
(
t
)
−
C
r
(
t
)
x
^
r
(
t
)
)
,
x
^
r
(
0
)
=
x
r
(
0
)
,
{\displaystyle {\dot {\hat {\mathbf {x} }}}_{r}(t)=A_{r}(t){\hat {\mathbf {x} }}_{r}(t)+B_{r}(t){\mathbf {u} }(t)+K_{r}(t)\left({\mathbf {y} }(t)-C_{r}(t){\hat {\mathbf {x} }}_{r}(t)\right),{\hat {\mathbf {x} }}_{r}(0)={\mathbf {x} }_{r}(0),}
u
(
t
)
=
−
L
r
(
t
)
x
^
r
(
t
)
.
{\displaystyle {\mathbf {u} }(t)=-L_{r}(t){\hat {\mathbf {x} }}_{r}(t).}
上述公式刻意写的类似传统全阶LQG控制器的形式,降阶的LQG控制问题也可以改写为下式:
x
^
˙
r
(
t
)
=
F
r
(
t
)
x
^
r
(
t
)
+
K
r
(
t
)
y
(
t
)
,
x
^
r
(
0
)
=
x
r
(
0
)
,
{\displaystyle {\dot {\hat {\mathbf {x} }}}_{r}(t)=F_{r}(t){\hat {\mathbf {x} }}_{r}(t)+K_{r}(t){\mathbf {y} }(t),{\hat {\mathbf {x} }}_{r}(0)={\mathbf {x} }_{r}(0),}
u
(
t
)
=
−
L
r
(
t
)
x
^
r
(
t
)
,
{\displaystyle {\mathbf {u} }(t)=-L_{r}(t){\hat {\mathbf {x} }}_{r}(t),}
其中
F
r
(
t
)
=
A
r
(
t
)
−
B
r
(
t
)
L
r
(
t
)
−
K
r
(
t
)
C
r
(
t
)
.
{\displaystyle F_{r}(t)=A_{r}(t)-B_{r}(t)L_{r}(t)-K_{r}(t)C_{r}(t).}
降阶LQG控制器的矩阵
F
r
(
t
)
,
K
r
(
t
)
,
L
r
(
t
)
{\displaystyle F_{r}(t),K_{r}(t),L_{r}(t)}
和
x
r
(
0
)
{\displaystyle {\mathbf {x} }_{r}(0)}
是由所谓的最佳投影方程 (optimal projection equations、OPE)来决定[ 3] 。
n
{\displaystyle n}
维的最佳投影方阵
τ
(
t
)
{\displaystyle \tau (t)}
是OPE 的核心。此矩阵的秩在所有状态下几乎都等于
n
r
{\displaystyle n_{r}}
。相关投影为斜投影(oblique projection):
τ
2
(
t
)
=
τ
(
t
)
{\displaystyle \tau ^{2}(t)=\tau (t)}
。最佳投影方程包括四个矩阵微分方程。前二个是LQG控制器对应的矩阵Riccati微分方程的扩展。在方程式中
τ
⊥
(
t
)
{\displaystyle \tau _{\perp }(t)}
表示
I
n
−
τ
(
t
)
{\displaystyle I_{n}-\tau (t)}
,而
I
n
{\displaystyle I_{n}}
为
n
{\displaystyle n}
维的单位矩阵
P
˙
(
t
)
=
A
(
t
)
P
(
t
)
+
P
(
t
)
A
′
(
t
)
−
P
(
t
)
C
′
(
t
)
W
−
1
(
t
)
C
(
t
)
P
(
t
)
+
V
(
t
)
+
τ
⊥
(
t
)
P
(
t
)
C
′
(
t
)
W
−
1
(
t
)
C
(
t
)
P
(
t
)
τ
⊥
′
(
t
)
,
P
(
0
)
=
E
(
x
(
0
)
x
′
(
0
)
)
,
−
S
˙
(
t
)
=
A
′
(
t
)
S
(
t
)
+
S
(
t
)
A
(
t
)
−
S
(
t
)
B
(
t
)
R
−
1
(
t
)
B
′
(
t
)
S
(
t
)
+
Q
(
t
)
+
τ
⊥
′
(
t
)
S
(
t
)
B
(
t
)
R
−
1
(
t
)
B
′
(
t
)
S
(
t
)
τ
⊥
(
t
)
,
{\displaystyle {\begin{aligned}{\dot {P}}(t)={}&A(t)P(t)+P(t)A'(t)-P(t)C'(t)W^{-1}(t)C(t)P(t)+V(t)\\[6pt]&{}+\tau _{\perp }(t)P(t)C'(t)W^{-1}(t)C(t)P(t)\tau '_{\perp }(t),\\[6pt]P(0)={}&E\left({\mathbf {x} }(0){\mathbf {x} }'(0)\right),\\[6pt]&{}-{\dot {S}}(t)=A'(t)S(t)+S(t)A(t)-S(t)B(t)R^{-1}(t)B'(t)S(t)+Q(t)\\[6pt]&{}+\tau '_{\perp }(t)S(t)B(t)R^{-1}(t)B'(t)S(t)\tau _{\perp }(t),\end{aligned}}}
S
(
T
)
=
F
.
{\displaystyle S(T)=F.}
若LQG的维度没有减少,也就是
n
=
n
r
{\displaystyle n=n_{r}}
,则
τ
(
t
)
=
I
n
,
τ
⊥
(
t
)
=
0
{\displaystyle \tau (t)=I_{n},\tau _{\perp }(t)=0}
,上述二个方程就是二个没有耦合的矩阵Riccati微分方程,对应全阶的LQG控制器。若
n
r
<
n
{\displaystyle n_{r}<n}
,则两个方程会有斜投影项
τ
(
t
)
.
{\displaystyle \tau (t).}
。这也是为何降阶的LQG控制器无法分离的原因,斜投影
τ
(
t
)
{\displaystyle \tau (t)}
是由另外二个矩阵微分方程所决定,其中也和秩的条件(rank conditions)有关。这四个矩阵微分方程组成了最佳投影方程。为了要列出另外二个矩阵微分方程,先定义以下二个矩阵:
Ψ
1
(
t
)
=
(
A
(
t
)
−
B
(
t
)
R
−
1
(
t
)
B
′
(
t
)
S
(
t
)
)
P
^
(
t
)
+
P
^
(
t
)
(
A
(
t
)
−
B
(
t
)
R
−
1
(
t
)
B
′
(
t
)
S
(
t
)
)
′
{\displaystyle \Psi _{1}(t)=(A(t)-B(t)R^{-1}(t)B'(t)S(t)){\hat {P}}(t)+{\hat {P}}(t)(A(t)-B(t)R^{-1}(t)B'(t)S(t))'}
+
P
(
t
)
C
′
(
t
)
W
−
1
(
t
)
C
(
t
)
P
(
t
)
,
{\displaystyle {}+P(t)C'(t)W^{-1}(t)C(t)P(t),}
Ψ
2
(
t
)
=
(
A
(
t
)
−
P
(
t
)
C
′
(
t
)
W
−
1
(
t
)
C
(
t
)
)
′
S
^
(
t
)
+
S
^
(
t
)
(
A
(
t
)
−
P
(
t
)
C
′
(
t
)
W
−
1
(
t
)
C
(
t
)
)
{\displaystyle \Psi _{2}(t)=(A(t)-P(t)C'(t)W^{-1}(t)C(t))'{\hat {S}}(t)+{\hat {S}}(t)(A(t)-P(t)C'(t)W^{-1}(t)C(t))}
+
S
(
t
)
B
(
t
)
R
−
1
(
t
)
B
′
(
t
)
S
(
t
)
.
{\displaystyle {}+S(t)B(t)R^{-1}(t)B'(t)S(t).}
则最后二个矩阵微分方程如下:
P
^
˙
(
t
)
=
1
/
2
(
τ
(
t
)
Ψ
1
(
t
)
+
Ψ
1
(
t
)
τ
′
(
t
)
)
,
P
^
(
0
)
=
E
(
x
(
0
)
)
E
(
x
(
0
)
)
′
,
rank
(
P
^
(
t
)
)
=
n
r
{\displaystyle {\dot {\hat {P}}}(t)=1/2\left(\tau (t)\Psi _{1}(t)+\Psi _{1}(t)\tau '(t)\right),{\hat {P}}(0)=E({\mathbf {x} }(0))E({\mathbf {x} }(0))',\operatorname {rank} ({\hat {P}}(t))=n_{r}}
almost everywhere,
−
S
^
˙
(
t
)
=
1
/
2
(
τ
′
(
t
)
Ψ
2
(
t
)
+
Ψ
2
(
t
)
τ
(
t
)
)
,
S
^
(
T
)
=
0
,
rank
(
S
^
(
t
)
)
=
n
r
{\displaystyle -{\dot {\hat {S}}}(t)=1/2\left(\tau '(t)\Psi _{2}(t)+\Psi _{2}(t)\tau (t)\right),{\hat {S}}(T)=0,\operatorname {rank} ({\hat {S}}(t))=n_{r}}
almost everywhere,
其中
τ
(
t
)
=
P
^
(
t
)
S
^
(
t
)
(
P
^
(
t
)
S
^
(
t
)
)
∗
.
{\displaystyle \tau (t)={\hat {P}}(t){\hat {S}}(t)\left({\hat {P}}(t){\hat {S}}(t)\right)^{*}.}
此处的 * 表示群广义逆矩阵(group generalized inverse)或Drazin逆矩阵 ,是唯一的,定义如下
A
∗
=
A
(
A
3
)
+
A
.
{\displaystyle A^{*}=A(A^{3})^{+}A.}
其中 + 是摩尔-彭若斯广义逆 .
矩阵
P
(
t
)
,
S
(
t
)
,
P
^
(
t
)
,
S
^
(
t
)
{\displaystyle P(t),S(t),{\hat {P}}(t),{\hat {S}}(t)}
都需要是非负对称矩阵。可以建构最佳投影方程的解,而此解可以决定降阶LQG控制器矩阵
F
r
(
t
)
,
K
r
(
t
)
,
L
r
(
t
)
{\displaystyle F_{r}(t),K_{r}(t),L_{r}(t)}
和
x
r
(
0
)
{\displaystyle {\mathbf {x} }_{r}(0)}
:
F
r
(
t
)
=
H
(
t
)
(
A
(
t
)
−
P
(
t
)
C
′
(
t
)
W
−
1
(
t
)
C
(
t
)
−
B
(
t
)
R
−
1
(
t
)
B
′
(
t
)
S
(
t
)
)
G
(
t
)
+
H
˙
(
t
)
G
′
(
t
)
,
{\displaystyle F_{r}(t)=H(t)\left(A(t)-P(t)C'(t)W^{-1}(t)C(t)-B(t)R^{-1}(t)B'(t)S(t)\right)G(t)+{\dot {H}}(t)G'(t),}
K
r
(
t
)
=
H
(
t
)
P
(
t
)
C
′
(
t
)
W
−
1
(
t
)
,
{\displaystyle K_{r}(t)=H(t)P(t)C'(t)W^{-1}(t),}
L
r
(
t
)
=
R
−
1
(
t
)
B
′
(
t
)
S
(
t
)
G
′
(
t
)
,
{\displaystyle L_{r}(t)=R^{-1}(t)B'(t)S(t)G'(t),}
x
r
(
0
)
=
H
(
0
)
E
(
x
(
0
)
)
.
{\displaystyle {\mathbf {x} }_{r}(0)=H(0)E({\mathbf {x} }(0)).}
上式中的矩阵
G
(
t
)
,
H
(
t
)
{\displaystyle G(t),H(t)}
是符合以下性质的矩阵:
G
′
(
t
)
H
(
t
)
=
τ
(
t
)
,
G
(
t
)
H
′
(
t
)
=
I
n
r
{\displaystyle G'(t)H(t)=\tau (t),G(t)H'(t)=I_{n_{r}}}
几乎在所有状态下。
可以由
P
^
(
t
)
S
^
(
t
)
{\displaystyle {\hat {P}}(t){\hat {S}}(t)}
的投影分解中得到[ 4] :
若降阶LQG问题中的所有矩阵都是非时变的,且最终时间(horizon)
T
{\displaystyle T}
趋近无限大,则最佳降阶LQG控制器和最佳投影方程也都会是非时变的[ 1] 。此情形下,最佳投影方程左侧的微分项会为零。
离散时间
离散时间的情形类似连续时间的例子,要处理的是将
n
{\displaystyle n}
阶传统离散时间全阶LQG问题转换为事先已知固定阶数的
n
r
<
n
{\displaystyle n_{r}<n}
阶降阶LQG控制器。为了要表示离散时间的OPE,先引入以下二个矩阵:
Ψ
i
1
=
(
A
i
−
B
i
(
B
i
′
S
i
+
1
B
i
+
R
i
)
−
1
B
i
′
S
i
+
1
A
i
)
)
P
^
i
(
A
i
−
B
i
(
B
i
′
S
i
+
1
B
i
+
R
i
)
−
1
B
i
′
S
i
+
1
A
i
)
)
′
{\displaystyle \Psi _{i}^{1}=\left(A_{i}-B_{i}(B'_{i}S_{i+1}B_{i}+R_{i})^{-1}B'_{i}S_{i+1}A_{i})\right){\hat {P}}_{i}\left(A_{i}-B_{i}(B'_{i}S_{i+1}B_{i}+R_{i})^{-1}B'_{i}S_{i+1}A_{i})\right)'}
+
A
i
P
i
C
i
′
(
C
i
P
i
C
i
′
+
W
i
)
−
1
C
i
P
i
A
i
′
{\displaystyle {}+A_{i}P_{i}C'_{i}(C_{i}P_{i}C'_{i}+W_{i})^{-1}C_{i}P_{i}A'_{i}}
Ψ
i
+
1
2
=
(
A
i
−
A
i
P
i
C
i
′
(
C
i
P
i
C
i
′
+
W
i
)
−
1
C
i
)
′
S
^
i
+
1
(
A
i
−
A
i
P
i
C
i
′
(
C
i
P
i
C
i
′
+
W
i
)
−
1
C
i
)
{\displaystyle \Psi _{i+1}^{2}=\left(A_{i}-A_{i}P_{i}C'_{i}(C_{i}P_{i}C'_{i}+W_{i})^{-1}C_{i}\right)'{\hat {S}}_{i+1}\left(A_{i}-A_{i}P_{i}C'_{i}(C_{i}P_{i}C'_{i}+W_{i})^{-1}C_{i}\right)}
+
A
i
′
S
i
+
1
B
i
(
B
i
′
S
i
+
1
B
i
+
R
i
)
−
1
B
i
′
S
i
+
1
A
i
{\displaystyle {}+A'_{i}S_{i+1}B_{i}(B'_{i}S_{i+1}B_{i}+R_{i})^{-1}B'_{i}S_{i+1}A_{i}}
则离散时间OPE为
P
i
+
1
=
A
i
(
P
i
−
P
i
C
i
′
(
C
i
P
i
C
i
′
+
W
i
)
−
1
C
i
P
i
)
A
i
′
+
V
i
+
τ
⊥
i
+
1
Ψ
i
1
τ
⊥
i
+
1
′
,
P
0
=
E
(
x
0
x
′
0
)
{\displaystyle P_{i+1}=A_{i}\left(P_{i}-P_{i}C'_{i}\left(C_{i}P_{i}C'_{i}+W_{i}\right)^{-1}C_{i}P_{i}\right)A'_{i}+V_{i}+\tau _{\perp i+1}\Psi _{i}^{1}\tau '_{\perp i+1},P_{0}=E\left({\mathbf {x} }_{0}{\mathbf {x'} }_{0}\right)}
.
S
i
=
A
i
′
(
S
i
+
1
−
S
i
+
1
B
i
(
B
i
′
S
i
+
1
B
i
+
R
i
)
−
1
B
i
′
S
i
+
1
)
A
i
+
Q
i
+
τ
⊥
i
′
Ψ
i
+
1
2
τ
⊥
i
,
S
N
=
F
{\displaystyle S_{i}=A'_{i}\left(S_{i+1}-S_{i+1}B_{i}\left(B'_{i}S_{i+1}B_{i}+R_{i}\right)^{-1}B'_{i}S_{i+1}\right)A_{i}+Q_{i}+\tau '_{\perp i}\Psi _{i+1}^{2}\tau _{\perp i},S_{N}=F}
.
P
^
i
+
1
=
1
/
2
(
τ
i
+
1
Ψ
i
1
+
Ψ
i
1
τ
i
+
1
′
)
,
P
^
0
=
E
(
x
(
0
)
)
E
(
x
(
0
)
)
′
,
rank
(
P
^
i
)
=
n
r
{\displaystyle {\hat {P}}_{i+1}=1/2(\tau _{i+1}\Psi _{i}^{1}+\Psi _{i}^{1}\tau '_{i+1}),{\hat {P}}_{0}=E({\mathbf {x} }(0))E({\mathbf {x} }(0))',\operatorname {rank} ({\hat {P}}_{i})=n_{r}}
almost everywhere,
S
^
i
=
1
/
2
(
τ
i
′
Ψ
i
+
1
2
+
Ψ
i
+
1
2
τ
i
)
,
S
^
N
=
0
,
rank
(
S
^
i
)
=
n
r
{\displaystyle {\hat {S}}_{i}=1/2(\tau '_{i}\Psi _{i+1}^{2}+\Psi _{i+1}^{2}\tau _{i}),{\hat {S}}_{N}=0,\operatorname {rank} ({\hat {S}}_{i})=n_{r}}
almost everywhere.
斜投影(oblique projection)矩阵为
τ
i
=
P
^
i
S
^
i
(
P
^
i
S
^
i
)
∗
.
{\displaystyle \tau _{i}={\hat {P}}_{i}{\hat {S}}_{i}\left({\hat {P}}_{i}{\hat {S}}_{i}\right)^{*}.}
非负对称矩阵
P
i
,
S
i
,
P
^
i
,
S
^
i
{\displaystyle P_{i},S_{i},{\hat {P}}_{i},{\hat {S}}_{i}}
是离散时间OPE的解,也决定了降阶LQG控制器的矩阵
F
i
r
,
K
i
r
,
L
i
r
{\displaystyle F_{i}^{r},K_{i}^{r},L_{i}^{r}}
and
x
0
r
{\displaystyle {\mathbf {x} }_{0}^{r}}
:
F
i
r
=
H
i
+
1
(
A
i
−
P
i
C
i
′
(
C
i
P
i
C
i
′
+
W
i
)
−
1
C
i
−
B
i
(
B
i
′
S
i
+
1
B
i
+
R
i
)
−
1
B
i
′
S
i
+
1
)
G
i
′
,
{\displaystyle F_{i}^{r}=H_{i+1}\left(A_{i}-P_{i}C'_{i}\left(C_{i}P_{i}C'_{i}+W_{i}\right)^{-1}C_{i}-B_{i}\left(B'_{i}S_{i+1}B_{i}+R_{i}\right)^{-1}B'_{i}S_{i+1}\right)G'_{i},}
K
i
r
=
H
i
+
1
P
i
C
i
′
(
C
i
P
i
C
i
′
+
W
i
)
−
1
,
{\displaystyle K_{i}^{r}=H_{i+1}P_{i}C'_{i}\left(C_{i}P_{i}C'_{i}+W_{i}\right)^{-1},}
L
i
r
=
(
B
i
′
S
i
+
1
B
i
+
R
i
)
−
1
B
i
′
S
i
+
1
G
i
′
,
{\displaystyle L_{i}^{r}=\left(B'_{i}S_{i+1}B_{i}+R_{i}\right)^{-1}B'_{i}S_{i+1}G'_{i},}
x
0
r
=
H
0
E
(
x
0
)
.
{\displaystyle {\mathbf {x} }_{0}^{r}=H_{0}E({\mathbf {x} }_{0}).}
在上述的方程中,矩阵
G
i
,
H
i
{\displaystyle G_{i},H_{i}}
是有以下性质的矩阵:
G
i
′
H
i
=
τ
i
,
G
i
H
i
′
=
I
n
r
{\displaystyle G'_{i}H_{i}=\tau _{i},G_{i}H'_{i}=I_{n_{r}}}
几乎在所有状态下。
这些矩阵可以从
P
^
i
S
^
i
{\displaystyle {\hat {P}}_{i}{\hat {S}}_{i}}
的投影因式分解中求得[ 4] 。
如同在连续时间中的例子一样,若问题中所有的矩阵都是非时变,且且最终时间(horizon)
T
{\displaystyle T}
趋近无限大,降阶LQG控制器就会是非时变的。因此离散时间OPE会收敛到稳态解,决定非时变的降阶LOG控制器[ 2] 。
离散时间OPE也可以应用在状态维度,输入维度或是输出维度可变的离散时间系统(具有时变维度的离散时间系统)[ 6] 。若在数位控制器中的取样是不同步的,就可能会出现这类的系统。