前面两篇博客里我们分析了文献[1]在关于Lyapunov函数的两种假设条件下证明积分形式系统扩张状态观测器(ESO)收敛性的套路,文献[1]还用齐次性说明了后一种假设条件的适应性,以表明有一定的应用性。仿照文献[1]的套路,根据自己大论文的研究方向加一个所谓的工程应用背景(其实还是一个微分方程模型),只要模型形式能够想办法改写成积分形式,然后控制器设计上再做点小改动,核心思想还是在Lyapunov函数上做文章,然后推导出一个所谓的存在性收敛条件,就是一篇自己的论文了,投个一般的SCI期刊(对理论要求不强的)还是比较有希望的,有毕业压力的同学可以考虑这个方法。总之就是,借助文献[1]的套路,既能用一坨数学公式唬住审稿人,增加中稿概率,又不增加实际难度,不需要花费太多脑力,何乐而不为?
同样地,其实可以结合自己的研究方向找上几篇理论研究的论文(最好是近几年的),打印出来仔细和反复研读,弄明白作者的套路,须知任何论文都是有套路的,特别是高产作者的论文,然后就可以依葫芦画瓢写自己的论文了。当然,找的论文一定要是自己能看明白的,这里说的看明白不是说能够一下子把原文的公式推导看明白,而是说能够看明白作者证明的思路,弄明白思路后其实很容易看明白公式推导,毕竟复杂公式的外表背后往往是简单而直接的思想。然后就能开始自己的公式推导了,自己的公式推导不一定要和原文一样(最好外表看起来不一样,也就是换汤不换药),只需要吸收其套路,结合自己的应用背景做一些改动更好。从这个角度讲,如果导师对自己是放养状态的话,用这种方式凑齐毕业所需的小论文是不难的。
言归正传,文献[2]是对文献[1]的进一步扩展,分析了下三角系统的ESO收敛性,其中的套路一样可以借鉴,这里对其进行介绍,考虑到文献[2]内容和前面两篇博客没有多大关系,公式也重新编号。
首先看看文献[2]中所谓的下三角系统是何方神圣:
{
x
˙
1
(
t
)
=
x
2
(
t
)
+
g
1
(
u
(
t
)
,
x
1
(
t
)
)
,
x
˙
2
(
t
)
=
x
3
(
t
)
+
g
2
(
u
(
t
)
,
x
1
(
t
)
,
x
2
(
t
)
)
,
⋮
x
˙
n
(
t
)
=
f
(
t
,
x
(
t
)
,
w
(
t
)
)
+
g
n
(
u
(
t
)
,
x
(
t
)
)
,
y
(
t
)
=
x
1
(
t
)
,
(1)
\left\{\begin{aligned} &\dot{x}_1(t)=x_2(t)+g_1(u(t),x_1(t)),\\ &\dot{x}_2(t)=x_3(t)+g_2(u(t),x_1(t),x_2(t)),\\ &\vdots\\ &\dot{x}_n(t)=f(t,x(t),w(t))+g_n(u(t),x(t)),\\ &y(t)=x_1(t), \end{aligned}\right.\tag{1}
⎩⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎧x˙1(t)=x2(t)+g1(u(t),x1(t)),x˙2(t)=x3(t)+g2(u(t),x1(t),x2(t)),⋮x˙n(t)=f(t,x(t),w(t))+gn(u(t),x(t)),y(t)=x1(t),(1)
其中,
g
i
(
⋅
)
∈
C
(
R
i
+
m
,
R
)
g_i(\cdot)\in C(\mathbb{R}^{i+m},\mathbb{R})
gi(⋅)∈C(Ri+m,R)为已知非线性函数,
f
(
t
,
⋅
)
∈
C
(
R
n
+
s
+
1
,
R
)
f(t,\cdot)\in C(\mathbb{R}^{n+s+1},\mathbb{R})
f(t,⋅)∈C(Rn+s+1,R)为未知非线性函数,
x
(
t
)
=
(
x
1
(
t
)
,
x
2
(
t
)
,
…
,
x
n
(
t
)
)
x(t)=(x_1(t),x_2(t),\ldots,x_n(t))
x(t)=(x1(t),x2(t),…,xn(t))为系统状态,
u
∈
R
m
u\in\mathbb{R}^m
u∈Rm为控制输入,
y
(
t
)
=
x
1
(
t
)
y(t)=x_1(t)
y(t)=x1(t)为输出,
w
∈
C
(
R
‾
+
,
R
)
w\in C(\overline{\mathbb{R}}^+,\mathbb{R})
w∈C(R+,R)为外部干扰。可见这里所谓的下三角系统其实就是文献[1]中的积分形式系统加上一些摄动项,准确来说,对于每一级状态的微分中对应的
g
i
(
⋅
)
g_i(\cdot)
gi(⋅),其自变量的最后一个分量只列写到该状态本身。系统(1)可以认为是文献[1]套路的延续。类似地,定义扩张状态
x
n
+
1
(
t
)
≜
f
(
t
,
x
(
t
)
,
w
(
t
)
)
(2)
x_{n+1}(t)\triangleq f(t,x(t),w(t))\tag{2}
xn+1(t)≜f(t,x(t),w(t))(2)
然后给出ESO的形式:
{
x
^
˙
1
(
t
)
=
x
^
2
(
t
)
+
1
r
n
−
1
h
1
(
r
n
(
y
(
t
)
−
x
^
1
(
t
)
)
)
+
g
1
(
u
(
t
)
,
x
^
1
(
t
)
)
,
⋮
x
^
˙
n
(
t
)
=
x
^
n
+
1
(
t
)
+
h
n
(
r
n
(
y
(
t
)
−
x
^
1
(
t
)
)
)
+
g
n
(
u
(
t
)
,
x
^
1
(
t
)
,
…
,
x
^
n
(
t
)
)
,
x
^
˙
n
+
1
(
t
)
=
r
h
n
+
1
(
r
n
(
y
(
t
)
−
x
^
1
(
t
)
)
)
,
(3)
\left\{\begin{aligned} &\dot{\hat{x}}_1(t)=\hat{x}_2(t)+\frac{1}{r^{n-1}}h_1(r^n(y(t)-\hat{x}_1(t)))+g_1(u(t),\hat{x}_1(t)),\\ &\vdots\\ &\dot{\hat{x}}_n(t)=\hat{x}_{n+1}(t)+h_n(r^n(y(t)-\hat{x}_1(t)))+g_n(u(t),\hat{x}_1(t),\ldots,\hat{x}_n(t)),\\ &\dot{\hat{x}}_{n+1}(t)=rh_{n+1}(r^n(y(t)-\hat{x}_1(t))), \end{aligned}\right.\tag{3}
⎩⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎧x^˙1(t)=x^2(t)+rn−11h1(rn(y(t)−x^1(t)))+g1(u(t),x^1(t)),⋮x^˙n(t)=x^n+1(t)+hn(rn(y(t)−x^1(t)))+gn(u(t),x^1(t),…,x^n(t)),x^˙n+1(t)=rhn+1(rn(y(t)−x^1(t))),(3)
其中,
r
r
r为高增益参数(其实就是文献[1]中
ε
\varepsilon
ε的倒数,即
r
=
1
ε
r=\frac{1}{\varepsilon}
r=ε1),
h
i
∈
C
(
R
,
R
)
h_i\in C(\mathbb{R},\mathbb{R})
hi∈C(R,R),
i
=
1
,
2
,
…
,
n
+
1
i=1,2,\ldots,n+1
i=1,2,…,n+1为设计函数。为了推导出ESO的收敛性,和文献[1]类似,同样需要一些假设:
假设1:
g
i
:
R
i
+
m
→
R
g_i:\mathbb{R}^{i+m}\rightarrow\mathbb{R}
gi:Ri+m→R满足
∣
g
i
(
u
,
v
1
,
…
,
v
i
)
−
g
i
(
u
,
v
~
1
,
…
,
v
~
i
)
∣
≤
Γ
(
u
)
∥
(
v
1
−
v
~
1
,
…
,
v
i
−
v
~
i
)
∥
θ
i
,
Γ
∈
C
(
R
m
,
R
)
,
(4)
\vert g_i(u,v_1,\ldots,v_i)-g_i(u,\tilde{v}_1,\ldots,\tilde{v}_i)\vert\leq \varGamma(u)\Vert(v_1-\tilde{v}_1,\ldots,v_i-\tilde{v}_i)\Vert^{\theta_i},\;\;\varGamma\in C(\mathbb{R}^m,\mathbb{R}),\tag{4}
∣gi(u,v1,…,vi)−gi(u,v~1,…,v~i)∣≤Γ(u)∥(v1−v~1,…,vi−v~i)∥θi,Γ∈C(Rm,R),(4)
其中, θ i ∈ ( ( n − i ) / ( n + 1 − i ) , 1 ] \theta_i\in((n-i)/(n+1-i),1] θi∈((n−i)/(n+1−i),1], i = 1 , 2 , … , n i=1,2,\ldots,n i=1,2,…,n.
假设2:
f
∈
C
1
(
R
n
+
2
,
R
)
f\in C^1(\mathbb{R}^{n+2},\mathbb{R})
f∈C1(Rn+2,R)满足
∣
f
(
t
,
x
,
w
)
∣
+
∣
∂
f
(
t
,
x
,
w
)
∂
t
∣
+
∣
∂
f
(
t
,
x
,
w
)
∂
x
i
∣
+
∣
∂
f
(
t
,
x
,
w
)
∂
w
∣
≤
ϖ
1
(
x
)
+
ϖ
1
(
w
)
,
\vert f(t,x,w)\vert+\left\vert\frac{\partial f(t,x,w)}{\partial t}\right\vert+ \left\vert\frac{\partial f(t,x,w)}{\partial x_i}\right\vert+ \left\vert\frac{\partial f(t,x,w)}{\partial w}\right\vert\leq\varpi_1(x)+\varpi_1(w),
∣f(t,x,w)∣+∣∣∣∣∂t∂f(t,x,w)∣∣∣∣+∣∣∣∣∂xi∂f(t,x,w)∣∣∣∣+∣∣∣∣∂w∂f(t,x,w)∣∣∣∣≤ϖ1(x)+ϖ1(w),
其中,
i
=
1
,
2
,
…
,
n
i=1,2,\ldots,n
i=1,2,…,n,
ϖ
1
∈
C
(
R
n
,
R
‾
+
)
\varpi_1\in C(\mathbb{R}^n,\overline{\mathbb{R}}^+)
ϖ1∈C(Rn,R+),
ϖ
2
∈
C
(
R
n
,
R
‾
+
)
\varpi_2\in C(\mathbb{R}^n,\overline{\mathbb{R}}^+)
ϖ2∈C(Rn,R+)为两个已知函数。
假设3: sup t ∈ [ 0 , ∞ ) ( ∣ w ( t ) ∣ + ∣ w ˙ ( t ) ∣ + ∥ u ( t ) ∥ ) < ∞ \sup\limits_{t\in[0,\infty)}(\vert w(t)\vert+\vert\dot{w}(t)\vert+\Vert u(t)\Vert)<\infty t∈[0,∞)sup(∣w(t)∣+∣w˙(t)∣+∥u(t)∥)<∞.
假设4:所有 h i ∈ C ( R , R ) h_i\in C(\mathbb{R},\mathbb{R}) hi∈C(R,R)满足如下Lyapunov条件:存在正常数 R , N > 0 R,N>0 R,N>0和连续径向无界的正定函数 V , W ∈ C ( R n + 1 , R + ) \mathcal{V},\mathcal{W}\in C(\mathbb{R}^{n+1},\mathbb{R}^+) V,W∈C(Rn+1,R+)使得
- ∑ i = 1 n ( v i + 1 − h i ( v 1 ) ) ∂ V ( v ) v i − h n + 1 ( v 1 ) ∂ V ( v ) v n ≤ − W ( v ) \sum\limits_{i=1}^n(v_{i+1}-h_i(v_1))\frac{\partial \mathcal{V}(v)}{\mathcal{v}_i}-h_{n+1}(v_1)\frac{\partial \mathcal{V}(v)}{\mathcal{v}_n}\leq -\mathcal{W}(v) i=1∑n(vi+1−hi(v1))vi∂V(v)−hn+1(v1)vn∂V(v)≤−W(v), ∀ v = ( v 1 , v 2 , … , v n + 1 ) ∈ R n + 1 \forall v=(v_1,v_2,\ldots,v_{n+1})\in \mathbb{R}^{n+1} ∀v=(v1,v2,…,vn+1)∈Rn+1;
- max i = 1 , … , n { ∥ ( v 1 , … , v i ) ∥ θ i ∣ ∂ ( V ) ( v ) ∂ v i ∣ } ≤ N W ( v ) \max\limits_{i=1,\ldots,n}\left\{\Vert(v_1,\ldots,v_i)\Vert^{\theta_i}\left\vert\frac{\partial \mathcal(V)(v)}{\partial v_i}\right\vert\right\}\leq N\mathcal{W}(v) i=1,…,nmax{∥(v1,…,vi)∥θi∣∣∣∂vi∂(V)(v)∣∣∣}≤NW(v), ∣ ∂ V ( v ) ∂ v n + 1 ∣ ≤ N W ( v ) \left\vert\frac{\partial\mathcal{V}(v)}{\partial v_{n+1}}\right\vert\leq N\mathcal{W}(v) ∣∣∣∂vn+1∂V(v)∣∣∣≤NW(v), v ∈ R n + 1 v\in\mathbb{R}^{n+1} v∈Rn+1, ∥ v ∥ ≥ R \Vert v\Vert\geq R ∥v∥≥R.
我们对比一下这4个假设和博客1以及博客2中的假设条件,可以发现:
- 假设1涉及到下三角系统特有的项 g i g_i gi,因此和博客1中没有对应的假设;
- 假设2和博客1中的假设1类似,是关于扩张状态界的假设;
- 假设3和博客1中的假设2类似,是关于外部干扰的假设;
- 假设4和博客1中的假设3或者博客2中的假设4类似,是关于Lyapunov函数的假设;
可见大部分论文的创新不是一蹴而就,总是在之前的工作上修修补补,不然怎么形成数量呢?对于假设1,这里用了一个所谓的Hölder连续性条件来唬人,其实并不改变后面的证明实质,而且当 θ i = 1 \theta_i=1 θi=1时就是我们在高等数学课上多次见到的Lipschitz条件,所以这里就可以直接认为假设1实际上认定了 g i ( ⋅ ) g_i(\cdot) gi(⋅)的某种连续性。假设1-4也都是为了后面利用Lyapunov函数证明收敛性方面而作出的,也就是说,先有证明,后有假设。然后就有下面的定理:
定理1:若假设1-4成立,系统(1)的解全局有界,则ESO(3)的状态实际收敛于系统(1)的状态和扩张状态:对任意
σ
>
0
\sigma>0
σ>0,存在正常数
r
0
>
0
r_0>0
r0>0使得
∣
x
^
i
(
t
)
−
x
i
(
t
)
∣
<
σ
,
∀
t
>
t
r
,
r
>
r
0
,
i
=
1
,
2
,
…
,
n
+
1
,
(5)
\vert \hat{x}_i(t)-x_i(t)\vert <\sigma,\;\forall t>t_r,\;r>r_0,\;i=1,2,\ldots,n+1,\tag{5}
∣x^i(t)−xi(t)∣<σ,∀t>tr,r>r0,i=1,2,…,n+1,(5)
其中,
t
r
t_r
tr为与
r
r
r相关的常数。
证明:定理1的结论和博客2中的定理有点像,只不过这里把之前用的 ε \varepsilon ε换成了 r r r,但不改变问题的实质。我们可以看到,尽管这里是研究所谓的下三角系统,但是可以看到,设计的ESO有什么区别呢?无法是你原系统(1)有特有项 g i g_i gi,我ESO(3)里面也照样给你来一个特有项 g i g_i gi,只不过ESO里面的 g i g_i gi自变量中状态取估计值,后果无法是ESO估计误差里面多了一些摄动项,没关系,结合文献[1]的套路,一样用Lyapunov函数把这些摄动项搞定,而为了搞定这些项,就有了前面的4个假设,这是定理证明的主要思路。
同样利用与文献[1]类似的坐标变换
η
i
(
t
)
=
r
n
+
1
−
i
(
x
i
(
t
)
−
x
^
i
(
t
)
)
,
i
=
1
,
2
,
…
,
n
+
1
,
η
(
t
)
=
(
η
1
(
t
)
,
…
,
η
n
+
1
(
t
)
)
T
,
(6)
\begin{aligned} &\eta_i(t)=r^{n+1-i}(x_i(t)-\hat{x}_i(t)),\;i=1,2,\ldots,n+1,\\ &\eta(t)=(\eta_1(t),\ldots,\eta_{n+1}(t))^\mathrm{T}, \end{aligned}\tag{6}
ηi(t)=rn+1−i(xi(t)−x^i(t)),i=1,2,…,n+1,η(t)=(η1(t),…,ηn+1(t))T,(6)
其中,
x
i
(
t
)
x_i(t)
xi(t)为系统(1)的解,
x
^
i
(
t
)
\hat{x}_i(t)
x^i(t)为ESO(3)的解,从而有
{
η
˙
1
(
t
)
=
r
(
η
2
(
t
)
−
h
1
(
η
1
(
t
)
)
)
+
r
n
(
g
1
(
u
(
t
)
,
x
1
(
t
)
)
−
g
1
(
u
(
t
)
,
x
^
1
(
t
)
)
)
,
⋮
η
˙
n
(
t
)
=
r
(
η
n
+
1
(
t
)
−
h
n
(
η
1
(
t
)
)
)
+
r
(
g
n
(
u
(
t
)
,
x
)
,
…
,
x
n
(
t
)
)
−
g
n
(
u
(
t
)
,
x
^
1
(
t
)
,
…
,
x
^
n
(
t
)
)
)
,
η
˙
n
+
1
(
t
;
r
)
=
−
r
h
n
+
1
(
η
1
(
t
)
)
+
x
˙
n
+
1
(
t
)
.
(7)
\left\{\begin{aligned} &\dot{\eta}_1(t)=r(\eta_2(t)-h_1(\eta_1(t)))+r^n(g_1(u(t),x_1(t))-g_1(u(t),\hat{x}_1(t))),\\ &\vdots\\ &\dot{\eta}_n(t)=r(\eta_{n+1}(t)-h_n(\eta_1(t)))+r(g_n(u(t),x),\ldots,\\ &\qquad\qquad x_n(t))-g_n(u(t),\hat{x}_1(t),\ldots,\hat{x}_n(t))),\\ &\dot{\eta}_{n+1}(t;r)=-rh_{n+1}(\eta_1(t))+\dot{x}_{n+1}(t). \end{aligned}\right.\tag{7}
⎩⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎧η˙1(t)=r(η2(t)−h1(η1(t)))+rn(g1(u(t),x1(t))−g1(u(t),x^1(t))),⋮η˙n(t)=r(ηn+1(t)−hn(η1(t)))+r(gn(u(t),x),…,xn(t))−gn(u(t),x^1(t),…,x^n(t))),η˙n+1(t;r)=−rhn+1(η1(t))+x˙n+1(t).(7)
下面开始分析摄动项,首先根据假设1,将式(6)代入式(7)可得
r
n
+
1
−
i
∣
g
i
(
u
(
t
)
,
x
1
(
t
)
,
…
,
x
i
(
t
)
)
−
g
i
(
u
(
t
)
,
x
^
1
(
t
)
,
…
,
x
^
i
(
t
)
)
∣
≤
Γ
(
u
(
t
)
)
r
n
+
1
−
i
∥
η
1
(
t
)
/
r
n
,
…
,
η
i
(
t
)
/
r
n
+
1
−
i
∥
θ
i
≤
Γ
(
u
(
t
)
)
r
(
n
+
1
−
i
)
(
1
−
θ
i
)
∥
η
1
(
t
)
/
r
n
,
…
,
η
i
(
t
)
/
r
n
+
1
−
i
∥
θ
i
,
∀
r
>
1.
(8)
\begin{aligned} r^{n+1-i}&\vert g_i(u(t),x_1(t),\ldots,x_i(t))-g_i(u(t),\hat{x}_1(t),\ldots,\hat{x}_i(t))\vert\\ &\leq\varGamma(u(t))r^{n+1-i} \Vert\eta_1(t)/r^n,\ldots,\eta_i(t)/r^{n+1-i}\Vert ^{\theta_i} \\ &\leq \varGamma(u(t))r^{(n+1-i)(1-\theta_i)}\Vert\eta_1(t)/r^n,\ldots,\eta_i(t)/r^{n+1-i}\Vert ^{\theta_i},\;\forall r>1. \end{aligned}\tag{8}
rn+1−i∣gi(u(t),x1(t),…,xi(t))−gi(u(t),x^1(t),…,x^i(t))∣≤Γ(u(t))rn+1−i∥η1(t)/rn,…,ηi(t)/rn+1−i∥θi≤Γ(u(t))r(n+1−i)(1−θi)∥η1(t)/rn,…,ηi(t)/rn+1−i∥θi,∀r>1.(8)
根据假设2,
∣
x
˙
n
+
1
(
t
)
∣
≤
(
ϖ
1
(
x
(
t
)
)
+
ϖ
2
(
w
(
t
)
)
)
(
1
+
w
˙
(
t
)
+
∑
i
=
1
n
−
1
∣
x
i
+
1
(
t
)
∣
+
∑
i
=
1
n
∣
g
i
(
u
(
t
)
,
x
1
(
t
)
,
…
,
x
i
(
t
)
)
∣
+
ϖ
1
(
x
(
t
)
)
+
ϖ
2
(
w
(
t
)
)
)
.
(9)
\begin{aligned} \vert\dot{x}_{n+1}(t)\vert \leq& (\varpi_1(x(t))+\varpi_2(w(t)))\left(1+\dot{w}(t)+\sum_{i=1}^{n-1}\vert x_{i+1}(t)\vert\right.\\ &\left. +\sum_{i=1}^n\vert g_i(u(t),x_1(t),\ldots,x_i(t))\vert+\varpi_1(x(t))+\varpi_2(w(t))\right). \end{aligned}\tag{9}
∣x˙n+1(t)∣≤(ϖ1(x(t))+ϖ2(w(t)))(1+w˙(t)+i=1∑n−1∣xi+1(t)∣+i=1∑n∣gi(u(t),x1(t),…,xi(t))∣+ϖ1(x(t))+ϖ2(w(t))).(9)
令
V
:
R
n
+
1
→
R
\mathcal{V}:\mathbb{R}^{n+1}\rightarrow\mathbb{R}
V:Rn+1→R为满足假设4的Lyapunov函数,
V
(
η
(
t
)
)
\mathcal{V}(\eta(t))
V(η(t))沿系统(7)的导数为
d
V
d
t
∣
a
l
o
n
g
(
7
)
=
∑
i
=
1
n
(
r
(
η
i
+
1
(
t
)
−
h
i
(
η
1
(
t
)
)
)
+
r
n
+
1
−
i
[
g
i
(
u
(
t
)
,
x
1
(
t
)
,
…
,
x
i
(
t
)
)
−
g
i
(
u
(
t
)
,
x
^
1
(
t
)
,
…
,
x
^
i
(
t
)
)
]
)
∂
V
(
η
(
t
)
)
∂
η
i
+
(
−
r
h
n
+
1
(
η
1
(
t
)
)
)
+
x
˙
n
+
1
(
t
)
)
∂
V
(
η
(
t
)
)
∂
η
n
+
1
.
(10)
\begin{aligned} \left.\frac{\mathrm{d}\mathcal{V}}{\mathrm{d}t}\right\vert_{\mathrm{along}\; (7)}=&\sum_{i=1}^n\left(r(\eta_{i+1}(t)-h_i(\eta_1(t)))+r^{n+1-i}[g_i(u(t),x_1(t),\ldots,x_i(t))\right.\\ &\left.-g_i(u(t),\hat{x}_1(t),\ldots,\hat{x}_i(t))] \right)\frac{\partial\mathcal{V}(\eta(t))}{\partial\eta_i}+(-rh_{n+1}(\eta_1(t)))\\ &+\dot{x}_{n+1}(t))\frac{\partial\mathcal{V}(\eta(t))}{\partial\eta_{n+1}}. \end{aligned}\tag{10}
dtdV∣∣∣∣along(7)=i=1∑n(r(ηi+1(t)−hi(η1(t)))+rn+1−i[gi(u(t),x1(t),…,xi(t))−gi(u(t),x^1(t),…,x^i(t))])∂ηi∂V(η(t))+(−rhn+1(η1(t)))+x˙n+1(t))∂ηn+1∂V(η(t)).(10)
由假设4的条件1、式(8)和式(9)可得
d
V
d
t
∣
a
l
o
n
g
(
7
)
≤
−
r
W
(
η
(
t
)
)
+
∣
x
˙
n
+
1
(
t
)
∣
∣
∂
V
(
η
(
t
)
)
∂
η
n
+
1
∣
+
∑
i
=
1
n
Γ
(
u
(
t
)
)
r
(
n
+
1
−
i
)
(
1
−
θ
i
)
∥
η
1
(
t
)
,
…
,
η
i
(
t
)
∥
θ
i
∣
∂
V
(
η
(
t
)
)
∂
η
i
∣
.
(11)
\begin{aligned}\left.\frac{\mathrm{d}\mathcal{V}}{\mathrm{d}t}\right\vert_{\mathrm{along}\; (7)} \leq&-r\mathcal{W}(\eta(t))+\vert \dot{x}_{n+1}(t)\vert\left\vert\frac{\partial\mathcal{V}(\eta(t))}{\partial\eta_{n+1}}\right\vert\\ &+\sum_{i=1}^n\varGamma(u(t))r^{(n+1-i)(1-\theta_i)}\Vert\eta_1(t),\ldots,\eta_i(t)\Vert^{\theta_i}\left\vert\frac{\partial\mathcal{V}(\eta(t))}{\partial\eta_i}\right\vert.\end{aligned}\tag{11}
dtdV∣∣∣∣along(7)≤−rW(η(t))+∣x˙n+1(t)∣∣∣∣∣∂ηn+1∂V(η(t))∣∣∣∣+i=1∑nΓ(u(t))r(n+1−i)(1−θi)∥η1(t),…,ηi(t)∥θi∣∣∣∣∂ηi∂V(η(t))∣∣∣∣.(11)
式(11)右端最后两项比较烦人,是证明稳定性的障碍,因此结合前面的假设条件去掉其影响,具体来说,根据假设2和假设4的条件2,当
∥
η
(
t
)
∥
≥
R
\Vert\eta(t)\Vert\geq R
∥η(t)∥≥R时有
∥
η
1
(
t
)
,
…
,
η
i
(
t
)
∥
θ
i
∣
∂
V
(
η
(
t
)
)
∂
η
i
∣
≤
N
(
W
)
(
η
(
t
)
)
,
∣
∂
V
(
η
(
t
)
)
∂
η
n
+
1
∣
≤
N
W
(
η
(
t
)
)
(12)
\begin{aligned} &\Vert\eta_1(t),\ldots,\eta_i(t)\Vert^{\theta_i}\left\vert\frac{\partial\mathcal{V}(\eta(t))}{\partial\eta_i}\right\vert\leq N\mathcal(W)(\eta(t)),\\ &\left\vert\frac{\partial\mathcal{V}(\eta(t))}{\partial\eta_{n+1}}\right\vert\leq N\mathcal{W}(\eta(t)) \end{aligned}\tag{12}
∥η1(t),…,ηi(t)∥θi∣∣∣∣∂ηi∂V(η(t))∣∣∣∣≤N(W)(η(t)),∣∣∣∣∂ηn+1∂V(η(t))∣∣∣∣≤NW(η(t))(12)
令
Λ
=
max
1
≤
i
≤
n
(
n
+
1
−
i
)
(
1
−
θ
i
)
,
N
11
=
sup
t
∈
[
0
,
∞
)
n
N
Γ
(
u
(
t
)
)
,
N
12
=
N
sup
t
∈
[
0
,
∞
)
[
ϖ
1
(
x
(
t
)
)
+
ϖ
2
(
w
(
t
)
)
]
(
1
+
∣
w
˙
(
t
)
∣
+
∑
i
=
1
n
∣
x
i
+
1
(
t
)
∣
+
∑
i
=
1
n
∣
g
i
(
u
(
t
)
,
x
1
(
t
)
,
…
,
x
i
(
t
)
)
∣
+
ϖ
1
(
x
(
t
)
)
+
ϖ
2
(
w
(
t
)
)
)
(13)
\begin{aligned} \varLambda=&\max_{1\leq i\leq n}(n+1-i)(1-\theta_i),\;N_{11}=\sup_{t\in[0,\infty)}nN\varGamma(u(t)),\\ N_{12}=&N\sup_{t\in[0,\infty)}[\varpi_1(x(t))+\varpi_2(w(t))] \left(1+\vert\dot{w}(t)\vert+\sum_{i=1}^n\vert x_{i+1}(t)\vert\right.\\ &\left.+\sum_{i=1}^n\vert g_i(u(t),x_1(t),\ldots,x_i(t))\vert+\varpi_1(x(t))+\varpi_2(w(t))\right) \end{aligned}\tag{13}
Λ=N12=1≤i≤nmax(n+1−i)(1−θi),N11=t∈[0,∞)supnNΓ(u(t)),Nt∈[0,∞)sup[ϖ1(x(t))+ϖ2(w(t))](1+∣w˙(t)∣+i=1∑n∣xi+1(t)∣+i=1∑n∣gi(u(t),x1(t),…,xi(t))∣+ϖ1(x(t))+ϖ2(w(t)))(13)
由假设1,
Λ
∈
(
0
,
1
)
\varLambda\in(0,1)
Λ∈(0,1)。由于
w
(
t
)
w(t)
w(t),
w
˙
(
t
)
\dot{w}(t)
w˙(t),
u
(
t
)
u(t)
u(t)和
x
(
t
)
x(t)
x(t)一致有界,因此
N
11
≤
∞
N_{11}\leq\infty
N11≤∞,
N
12
≤
∞
N_{12}\leq\infty
N12≤∞。仔细看的话,其实如果
r
>
1
r>1
r>1,那么有
∣
x
˙
n
+
1
(
t
)
∣
=
∣
∂
f
(
t
,
x
,
w
)
∂
t
+
∂
f
(
t
,
x
,
w
)
∂
x
i
x
˙
i
+
∂
f
(
t
,
x
,
w
)
∂
w
w
˙
i
∣
≤
∣
∂
f
(
t
,
x
,
w
)
∂
t
∣
+
∣
∂
f
(
t
,
x
,
w
)
∂
x
i
x
˙
i
∣
+
∣
∂
f
(
t
,
x
,
w
)
∂
w
w
˙
i
∣
≤
ϖ
1
(
x
)
+
ϖ
1
(
w
)
+
(
ϖ
1
(
x
)
+
ϖ
1
(
w
)
)
∑
i
=
1
n
−
1
(
∣
x
i
+
1
+
∣
g
i
∣
)
+
(
ϖ
1
(
x
)
+
ϖ
1
(
w
)
)
(
∣
g
n
∣
+
(
ϖ
1
(
x
)
+
ϖ
1
(
w
)
)
)
+
(
ϖ
1
(
x
)
+
ϖ
1
(
w
)
)
∣
w
˙
∣
\begin{aligned} \vert \dot{x}_{n+1}(t)\vert=&\left\vert\frac{\partial f(t,x,w)}{\partial t}+ \frac{\partial f(t,x,w)}{\partial x_i}\dot{x}_i+ \frac{\partial f(t,x,w)}{\partial w}\dot{w}_i\right\vert\\ \leq&\left\vert\frac{\partial f(t,x,w)}{\partial t}\right\vert+ \left\vert\frac{\partial f(t,x,w)}{\partial x_i}\dot{x}_i\right\vert+ \left\vert\frac{\partial f(t,x,w)}{\partial w}\dot{w}_i\right\vert\\ \leq&\varpi_1(x)+\varpi_1(w)+(\varpi_1(x)+\varpi_1(w))\sum_{i=1}^{n-1}(\vert x_{i+1}+\vert g_i\vert)\\&+(\varpi_1(x)+\varpi_1(w))(\vert g_n\vert+(\varpi_1(x)+\varpi_1(w)))+(\varpi_1(x)+\varpi_1(w))\vert\dot{w}\vert \end{aligned}
∣x˙n+1(t)∣=≤≤∣∣∣∣∂t∂f(t,x,w)+∂xi∂f(t,x,w)x˙i+∂w∂f(t,x,w)w˙i∣∣∣∣∣∣∣∣∂t∂f(t,x,w)∣∣∣∣+∣∣∣∣∂xi∂f(t,x,w)x˙i∣∣∣∣+∣∣∣∣∂w∂f(t,x,w)w˙i∣∣∣∣ϖ1(x)+ϖ1(w)+(ϖ1(x)+ϖ1(w))i=1∑n−1(∣xi+1+∣gi∣)+(ϖ1(x)+ϖ1(w))(∣gn∣+(ϖ1(x)+ϖ1(w)))+(ϖ1(x)+ϖ1(w))∣w˙∣
∑ i = 1 n Γ ( u ( t ) ) r ( n + 1 − i ) ( 1 − θ i ) ∥ η 1 ( t ) , … , η i ( t ) ∥ θ i ∣ ∂ V ( η ( t ) ) ∂ η i ∣ ≤ n Γ ( u ( t ) ) r Λ ∥ η 1 ( t ) , … , η i ( t ) ∥ θ i ∣ ∂ V ( η ( t ) ) ∂ η i ∣ \begin{aligned} &\sum_{i=1}^n\varGamma(u(t))r^{(n+1-i)(1-\theta_i)}\Vert\eta_1(t),\ldots,\eta_i(t)\Vert^{\theta_i}\left\vert\frac{\partial\mathcal{V}(\eta(t))}{\partial\eta_i}\right\vert\\ \leq& n\varGamma(u(t))r^\varLambda \Vert\eta_1(t),\ldots,\eta_i(t)\Vert^{\theta_i}\left\vert\frac{\partial\mathcal{V}(\eta(t))}{\partial\eta_i}\right\vert \end{aligned} ≤i=1∑nΓ(u(t))r(n+1−i)(1−θi)∥η1(t),…,ηi(t)∥θi∣∣∣∣∂ηi∂V(η(t))∣∣∣∣nΓ(u(t))rΛ∥η1(t),…,ηi(t)∥θi∣∣∣∣∂ηi∂V(η(t))∣∣∣∣
可见式
N
11
r
Λ
W
(
η
)
N_{11}r^\varLambda\mathcal{W}(\eta)
N11rΛW(η)和
N
12
W
(
η
)
N_{12}\mathcal{W}(\eta)
N12W(η)对应式(11)中右端最后两项的界,因此如果
∥
η
(
t
)
∥
≥
R
\Vert\eta(t)\Vert\geq R
∥η(t)∥≥R且
r
>
1
r>1
r>1,则有
d
V
d
t
∣
a
l
o
n
g
(
7
)
≤
−
(
r
−
N
11
r
Λ
−
N
12
)
W
(
η
(
t
)
)
.
(14)
\begin{aligned}\left.\frac{\mathrm{d}\mathcal{V}}{\mathrm{d}t}\right\vert_{\mathrm{along}\; (7)}\leq -(r-N_{11}r^\varLambda-N_{12})\mathcal{W}(\eta(t)).\end{aligned}\tag{14}
dtdV∣∣∣∣along(7)≤−(r−N11rΛ−N12)W(η(t)).(14)
由于
V
(
⋅
)
\mathcal{V}(\cdot)
V(⋅)和
W
(
⋅
)
\mathcal{W}(\cdot)
W(⋅)径向无界且连续正定,根据文献[3]的引理4.3(在博客2中有介绍)可知存在
K
∞
\mathcal{K}_\infty
K∞类函数
κ
i
:
R
‾
+
→
R
‾
+
(
i
=
1
,
2
,
3
,
4
)
\kappa_i:\overline{\mathbb{R}}^+\rightarrow\overline{\mathbb{R}}^+(i=1,2,3,4)
κi:R+→R+(i=1,2,3,4)使得
κ
1
(
∥
v
∥
)
≤
V
(
v
)
≤
κ
2
(
∥
v
∥
)
,
κ
3
(
∥
v
∥
)
≤
W
(
v
)
≤
κ
4
(
∥
v
∥
)
,
∀
v
∈
R
n
+
1
(15)
\begin{aligned} &\kappa_1(\Vert v\Vert)\leq\mathcal{V}(v)\leq\kappa_2(\Vert v\Vert),\\ &\kappa_3(\Vert v\Vert)\leq\mathcal{W}(v)\leq\kappa_4(\Vert v\Vert),\;\;\forall v\in\mathbb{R}^{n+1} \end{aligned}\tag{15}
κ1(∥v∥)≤V(v)≤κ2(∥v∥),κ3(∥v∥)≤W(v)≤κ4(∥v∥),∀v∈Rn+1(15)
若
V
(
η
(
r
;
t
)
)
≥
κ
2
(
R
)
\mathcal{V}(\eta(r;t))\geq\kappa_2(R)
V(η(r;t))≥κ2(R),则
∥
η
(
t
)
∥
≥
κ
2
−
1
(
V
(
η
(
r
;
t
)
)
)
≥
R
\Vert\eta(t)\Vert\geq\kappa_2^{-1}(\mathcal{V}(\eta(r;t)))\geq R
∥η(t)∥≥κ2−1(V(η(r;t)))≥R,且
W
(
η
(
t
)
)
≥
κ
3
(
∥
η
(
t
)
∥
)
≥
κ
3
(
R
)
\mathcal{W}(\eta(t))\geq\kappa_3(\Vert\eta(t)\Vert)\geq\kappa_3(R)
W(η(t))≥κ3(∥η(t)∥)≥κ3(R)。令
r
>
r
1
≜
max
{
1
,
(
3
N
11
)
1
/
(
1
−
Λ
)
,
(
N
12
/
N
11
)
1
/
Λ
}
,
r>r_1\triangleq\max\{1,(3N_{11})^{1/(1-\varLambda)},(N_{12}/N_{11})^{1/\varLambda}\},
r>r1≜max{1,(3N11)1/(1−Λ),(N12/N11)1/Λ},
则由式(14)可知
V
(
η
(
t
)
)
\mathcal{V(\eta(t))}
V(η(t))沿系统(7)的导数满足
d
V
d
t
∣
a
l
o
n
g
(
7
)
≤
−
N
11
κ
3
(
R
)
r
Λ
<
0
,
∀
r
>
r
1
,
∥
η
(
t
)
∥
≥
R
.
(16)
\begin{aligned}\left.\frac{\mathrm{d}\mathcal{V}}{\mathrm{d}t}\right\vert_{\mathrm{along}\; (7)}\leq -N_{11}\kappa_3(R)r^\varLambda<0,\forall r>r_1,\;\Vert\eta(t)\Vert\geq R. \end{aligned}\tag{16}
dtdV∣∣∣∣along(7)≤−N11κ3(R)rΛ<0,∀r>r1,∥η(t)∥≥R.(16)
因此,对每一个
r
>
r
1
r>r_1
r>r1,均存在与
r
r
r相关的常数
t
r
1
t_{r_1}
tr1使得当
t
>
t
r
1
t>t_{r_1}
t>tr1时,
V
(
η
(
t
)
)
≤
κ
2
(
R
)
\mathcal{V}(\eta(t))\leq\kappa_2(R)
V(η(t))≤κ2(R).
到目前为止,证明了
V
(
η
(
t
)
)
\mathcal{V(\eta(t))}
V(η(t))的有界性,进一步利用式(11)可得
d
V
d
t
∣
a
l
o
n
g
(
7
)
≤
−
r
W
(
η
(
t
)
)
+
M
11
r
Λ
+
M
12
,
∀
r
>
r
1
,
t
>
t
r
1
.
(17)
\begin{aligned}\left.\frac{\mathrm{d}\mathcal{V}}{\mathrm{d}t}\right\vert_{\mathrm{along}\; (7)}\leq -r\mathcal{W}(\eta(t))+M_{11}r^\varLambda+M_{12},\forall r>r_1,\;t>t_{r_1}.\end{aligned}\tag{17}
dtdV∣∣∣∣along(7)≤−rW(η(t))+M11rΛ+M12,∀r>r1,t>tr1.(17)
其中,
M
11
=
sup
t
∈
[
0
,
∞
)
Γ
(
u
(
t
)
)
∑
i
=
1
n
sup
v
∈
{
v
∈
R
n
+
1
:
V
(
v
)
≤
κ
2
(
R
)
}
∥
v
∥
i
θ
∥
∂
V
(
v
)
∂
v
∥
,
M
12
=
N
12
N
sup
v
∈
{
v
∈
R
n
+
1
:
V
(
v
)
≤
κ
2
(
R
)
}
∣
∂
V
(
v
)
∂
v
n
+
1
∣
\begin{aligned} &M_{11}=\sup_{t\in[0,\infty)}\varGamma(u(t))\sum_{i=1}^n\sup_{v\in\{v\in\mathbb{R}^{n+1}:\mathcal{V}(v)\leq\kappa_2(R)\}}\Vert v\Vert^\theta_i\left\Vert\frac{\partial \mathcal{V}(v)}{\partial v}\right\Vert,\\ &M_{12}=\frac{N_{12}}{N}\sup_{v\in\{v\in\mathbb{R}^{n+1}:\mathcal{V}(v)\leq\kappa_2(R)\}}\left\vert\frac{\partial \mathcal{V}(v)}{\partial v_{n+1}}\right\vert \end{aligned}
M11=t∈[0,∞)supΓ(u(t))i=1∑nv∈{v∈Rn+1:V(v)≤κ2(R)}sup∥v∥iθ∥∥∥∥∂v∂V(v)∥∥∥∥,M12=NN12v∈{v∈Rn+1:V(v)≤κ2(R)}sup∣∣∣∣∂vn+1∂V(v)∣∣∣∣
根据
V
(
v
)
\mathcal{V}(v)
V(v)的径向无界性,
{
v
∈
R
n
+
1
:
V
(
v
)
≤
κ
2
(
R
)
}
⊂
R
n
+
1
\{v\in\mathbb{R}^{n+1}:\mathcal{V}(v)\leq\kappa_2(R)\}\subset \mathbb{R}^{n+1}
{v∈Rn+1:V(v)≤κ2(R)}⊂Rn+1有界,结合
∇
V
\nabla\mathcal{V}
∇V的连续性可知
M
11
<
∞
M_{11}<\infty
M11<∞,
M
12
<
∞
M_{12}<\infty
M12<∞。
对任意给定
σ
>
0
\sigma>0
σ>0,由式(15)和(17)可知,若
r
>
r
0
≜
max
{
r
1
,
(
3
M
11
κ
3
(
σ
)
)
1
1
−
Λ
,
(
M
12
M
11
1
Λ
)
}
,
r>r_0\triangleq \max\left\{r_1,\left(\frac{3M_{11}}{\kappa_3(\sigma)}\right)^{\frac{1}{1-\varLambda}},\left(\frac{M_{12}}{M_{11}}^\frac{1}{\varLambda}\right)\right\},
r>r0≜max{r1,(κ3(σ)3M11)1−Λ1,(M11M12Λ1)},
且
∥
η
(
t
)
∥
≥
σ
\Vert\eta(t)\Vert\geq\sigma
∥η(t)∥≥σ,则
W
(
η
(
t
)
)
≥
κ
3
(
σ
)
W(\eta(t))\geq\kappa_3(\sigma)
W(η(t))≥κ3(σ),此时结合式(17)有
d
V
d
t
∣
a
l
o
n
g
(
7
)
≤
−
r
κ
3
(
σ
)
+
2
M
11
r
Λ
≤
−
M
11
r
Λ
<
0
(18)
\left.\frac{\mathrm{d}\mathcal{V}}{\mathrm{d}t}\right\vert_{\mathrm{along}\; (7)}\leq -r\kappa_3(\sigma)+2M_{11}r^\varLambda\leq-M_{11}r^\varLambda<0\tag{18}
dtdV∣∣∣∣along(7)≤−rκ3(σ)+2M11rΛ≤−M11rΛ<0(18)
因此,存在
t
r
>
t
r
1
t_r>t_{r_1}
tr>tr1使得对任意
r
>
r
0
r>r_0
r>r0和所有
t
>
t
r
t>t_r
t>tr,
∥
η
(
t
)
∥
<
σ
\Vert\eta(t)\Vert<\sigma
∥η(t)∥<σ成立。由式(6)可得
∣
x
i
(
t
)
−
x
^
i
(
t
)
∣
=
∣
η
i
(
t
)
∣
r
n
+
1
−
i
≤
σ
,
i
=
1
,
2
,
…
,
n
+
1.
(19)
\vert x_i(t)-\hat{x}_i(t)\vert=\frac{\vert\eta_i(t)\vert}{r^{n+1-i}}\leq\sigma,\;i=1,2,\ldots,n+1.\tag{19}
∣xi(t)−x^i(t)∣=rn+1−i∣ηi(t)∣≤σ,i=1,2,…,n+1.(19)
定理1证明完毕。
和文献[1]的套路类似,这里也是反复在Lyapunov函数的导数上作文章,想方设法弄出负定项来证明收敛性,反正 r r r理论上可以任意大。这里的语言组织表面上没有利用反证法,实际上在证明 V ( η ( t ) ) \mathcal{V(\eta(t))} V(η(t))的有界性时其实还是反证法,换汤不换药。接着,文献[2]将 w w w趋于稳态值和 f f f形式作为定理1的特例进行了收敛性分析,不过证明过程没有实质区别。进一步,文献[2]利用齐次性分析了齐次ESO的收敛性,套路依旧类似。此外,文献[2]还讨论了ESO增益过高时导致的峰值现象,并建议 r r r取为时变形式,这个属于理论上的细节了,就不展开说明了,不是这里讨论套路的重点。
到目前为止,关于ESO收敛性分析套路说明完毕。总之,这里想强调的是,千万不要被表面上复杂的公式唬住,而是应认真分析公式背后的思想(就像很多控制理论方面的文章其实就是一个Lyapunov函数推来推去,不管是加了ESO也好,加了神经网络也罢,都是一路货色),明白了思想,如果不是进一步追求理论深度的话,其实是容易结合自己的研究方向弄出缓解毕业压力的新文章的,这是本文的主要观点。
参考文献
[1]Guo B Z, Zhao Z. On the convergence of an extended state observer for nonlinear systems with uncertainty[J]. Systems & Control Letters, 2011, 60(6): 420-430.
[2]Zhao Z L, Guo B Z. Extended state observer for uncertain lower triangular nonlinear systems[J]. Systems & Control Letters, 2015, 85: 100-108.
[3]Khalil H K. Nonlinear systems[M]. Prentice-Hall, 2001.