Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
1490 JOURNAL OF COMPUTERS, VOL. 8, NO. 6, JUNE 2013<br />
control. So, λ is selected larger than the magnitude of the<br />
uncerta<strong>in</strong>ty and it will affect the convergence rate of the<br />
track<strong>in</strong>g error, and Ξ is chosen very small to best<br />
approximate the sign function and it will affect the size<br />
of the residual set to which the track<strong>in</strong>g error will<br />
converge. The sign function is not used here to avoid<br />
problems associated with it as chatter<strong>in</strong>g and solutions<br />
existence.<br />
By add<strong>in</strong>g and subtract<strong>in</strong>g ν <strong>in</strong> (3), we obta<strong>in</strong><br />
T<br />
T<br />
bPe<br />
e ⎛ ⎞<br />
= ( A0 −bK ) e −bλ<br />
tanh ⎜ ⎟−b[ h( ξη , , u)<br />
−v]<br />
(5)<br />
⎝ Ξ ⎠<br />
From the fact that the signal v does not explicitly<br />
depend upon the control <strong>in</strong>put u and Assumption 1, the<br />
partial derivative of h(, ξη , u)<br />
− v with respect to the<br />
<strong>in</strong>put u satisfies<br />
( ξη )<br />
∂ h(, , u) −v ∂h(, ξη, u)<br />
= > 0<br />
∂u<br />
∂u<br />
Thus accord<strong>in</strong>g to the implicit function theorem, there<br />
exists some ideal controller u * ( ξην , , ) satisfy<strong>in</strong>g the<br />
follow<strong>in</strong>g equality for all (, ξη,) v ∈ × ×<br />
R<br />
R<br />
:<br />
ξ<br />
η<br />
(6)<br />
*<br />
h(, ξη, u (, ξη ,)) v − v = 0<br />
(7)<br />
Therefore, if the control <strong>in</strong>put u is chosen as the ideal<br />
controller u<br />
* (, ξη ,) v , the closed-loop error dynamic (5) is<br />
reduced to<br />
T<br />
T<br />
b Pe<br />
e ⎛ ⎞<br />
= ( A0 −bK ) e −bλ<br />
tanh ⎜ ⎟ (8)<br />
⎝ Ξ ⎠<br />
Consider<strong>in</strong>g the follow<strong>in</strong>g positive function to the<br />
error dynamic:<br />
V<br />
T<br />
= e Pe<br />
(9)<br />
Us<strong>in</strong>g (4) and (8), the time derivative of (9) becomes<br />
T<br />
T T b Pe<br />
V ⎛ ⎞<br />
=−e Qe −2λb Pe tanh⎜ ⎟ (10)<br />
⎝ Ξ ⎠<br />
T<br />
⎛b Pe ⎞<br />
Because the term b T Pe and tanh ⎜ ⎟ always<br />
⎝ Ξ ⎠<br />
have same sign, we conclude that V<br />
≤ 0 , and only<br />
when e = 0 , V = 0 , which means lim | e | = 0 .<br />
t→∞<br />
III. ZERO DYNAMICS<br />
If system (1) is controlled by the <strong>in</strong>put u, the state<br />
vector η is completely unobservable from the output,<br />
then the subsystem<br />
dη<br />
q(0, , u(0, , v(0, )))<br />
dt = η η η (11)<br />
is addressed as the zero dynamic.<br />
Assumption 3: Zero dynamics (11) is exponentially<br />
stable, and the function q( ξη , , u)<br />
is Lipschitz <strong>in</strong>ξ . There<br />
exists Lipschitz constants L ξ<br />
and L<br />
q<br />
such that<br />
q(, ξη, u) − q(0, η, u ≤ L ξ + Lq<br />
(12)<br />
where u = u(0, η η<br />
, v(0, η ))) .<br />
By Lyapunov converse theorem, there is a Lyapunov<br />
function V ( ) 0<br />
η which satisfies<br />
η<br />
2 2<br />
1<br />
V0( )<br />
2<br />
ξ<br />
σ η ≤ η ≤ σ η (13)<br />
∂V0<br />
( η)<br />
q(0, η)<br />
≤−σ3<br />
η<br />
∂η<br />
∂V ( η)<br />
0<br />
∂η<br />
≤ σ<br />
Where σ , i = 1,2,3,4 are positive constant.<br />
i<br />
4<br />
η<br />
IV. DESIGN OF CONTROLLER<br />
2<br />
(14)<br />
(15)<br />
In control eng<strong>in</strong>eer<strong>in</strong>g, radial basis function (RBF)<br />
NNs are usually used as a tool for model<strong>in</strong>g nonl<strong>in</strong>ear<br />
functions because of their good capabilities <strong>in</strong> function<br />
approximation. RBFNN represents a class of l<strong>in</strong>early<br />
parameterized approximations and can be replaced by any<br />
other l<strong>in</strong>early parameterized approximations such as<br />
spl<strong>in</strong>e functions or fuzzy systems. Moreover, nonl<strong>in</strong>early<br />
parameterized approximations, such as multilayer neural<br />
network (MNN), can be l<strong>in</strong>earized as l<strong>in</strong>early<br />
parameterized approximations, with the higher order<br />
terms of Taylor series expansions be<strong>in</strong>g taken as part of<br />
the model<strong>in</strong>g error.<br />
In this paper, the follow<strong>in</strong>g RBF NN based on GGAP-<br />
RBF [20] algorithm which can avoid to select <strong>in</strong>itial<br />
neural network parameters and nodes number of hidden<br />
T<br />
layer artificially uz ( ) = φ ( z)<br />
θ is used to approximate<br />
the ideal controller<br />
1<br />
* T T<br />
u ( z ) , where z = ⎡ ⎣ ξ , η , v⎤<br />
⎦ ,<br />
φ( z) = ( φ ( z), , φ ( z)) T is the basic function vector, and<br />
M<br />
θ = ( θ1, , θ ) T<br />
M<br />
is the adjustable parameter. It has been<br />
proven that neural network can approximate any smooth<br />
q<br />
function over a compact set ΩZ<br />
⊂ R to arbitrarily any<br />
accuracy as<br />
( ) = φ ( ) θ + δ( ) (16)<br />
* T *<br />
u z z z<br />
with bounded function approximation error δ ( z)<br />
satisfy<strong>in</strong>g δ ( z)<br />
≤ δ .Where<br />
vector which m<strong>in</strong>imizes the function δ ( z)<br />
*<br />
θ is an ideal parameter<br />
T<br />
. In this paper,<br />
we assume that the used neural network does not violate<br />
the universal approximation property on the compact set<br />
Ω<br />
Z<br />
, which is assumed large enough so that the variable<br />
z rema<strong>in</strong>s <strong>in</strong>side it under closed-loop control.<br />
Let us def<strong>in</strong>e the control error between the controllers<br />
uz ( ) and u<br />
* ( z ) as<br />
© 2013 ACADEMY PUBLISHER