Designing SDRE-Based Controller For A Class of Nonlinear Singularly Perturbed Systems

Seyed Mostafa Ghadami, Roya Amjadifard & Hamid Khaloozadeh
Designing SDRE-Based Controller for a Class of Nonlinear

Singularly Perturbed Systems
Seyed Mostafa Ghadami
m.gadami@srbiau.ac.ir
Department of Electrical Engineering,

Science and Research Branch,
Islamic Azad University, Tehran, Iran
Roya Amjadifard
amjadifard@tmu.ac.ir
Assistant Professor,
Kharazmi University,
Tehran, Iran
Hamid Khaloozadeh
h_khaloozadeh@kntu.ac.ir
Professor,
K.N. Toosi University of Technology,
Tehran, Iran
Abstract
Designing a controller for nonlinear systems is difficult to be applied. Thus, it is usually based on
a linearization around their equilibrium points. The state dependent Riccati equation control
approach is an optimization method that has the simplicity of the classical linear quadratic control
method. On the other hand, the singular perturbation theory is used for the decomposition of a
high-order system into two lower-order systems. In this study, the finite-horizon optimization of a
class of nonlinear singularly perturbed systems based on the singular perturbation theory and the
state dependent Riccati equation technique together is addressed. In the proposed method, first,
the Hamiltonian equations are described as a state-dependent Hamiltonian matrix, from which,
the reduced-order subsystems are obtained. Then, these subsystems are converted into outerlayer, initial layer correction and final layer correction equations, from which, the separated state
dependent Riccati equations are derived. The optimal control law is, then, obtained by computing
the Riccati matrices.
Keywords: Singularly Perturbed Systems, State-Dependent Riccati Equation, Nonlinear Optimal
Control, Finite-Horizon Optimization Problem, Single Link Flexible Joint Robot Manipulator.
1. INTRODUCTION
Designing regulator systems is an important class of optimal control problems in which optimal
control law leads to the Hamilton-Jacobi-Belman (HJB) equation. Various techniques have been
suggested to solve this equation. One of these techniques, which are used for optimizing in
infinite horizon, is based on the state-dependent Riccati equation (SDRE). In this technique,
unlike linearization methods, a description of the system as state-dependent coefficients (SDCs)
and in the form f(x)=A(x)x must be provided. In this representation, A(x) is not unique. Therefore,
the solutions of the SDRE would be dependent on the choice of matrix A(x). With suitable choice
of the matrix, the solution to the equation is optimal; otherwise, the equation has suboptimal
solutions. Bank and Mhana [1] proposed a suitable method for the selection of SDCs. imen [2]
provided the condition for the solvability and local asymptotic stability of the SDRE closed-loop
system for a class of nonlinear systems. Khaloozadeh and Abdolahi converted the nonlinear
regulation [3] and tracking [4] problems in the finite-horizon to a state-dependent quasi-Riccati
equation. They also provided an iterative method based on the Piccard theorem, which obtains a
solution at a low convergence rate but good precision. On the other hand, the system discussed
in this study is a class of nonlinear singularly perturbed systems. Naidu and Calise [5] dealt with
International Journal of Robotics and Automation (IJRA), Volume (4) : Issue (1) : 2013
the use of the singular perturbation theory and the two time scale (TTS) method in satellite and
interplanetary trajectories, missiles, launch vehicles and hypersonic flight, space robotics. For LTI
singularly perturbed systems, Su et al. [6] and Gajic et al. [7] performed the exact slow-fast
decomposition of the linear quadratic (LQ) singularly perturbed optimal control problem in infinite
horizon by deriving separate Riccati equations. Also, Gajic et al. [8] did the same for the case of
finite horizon. Amjadifard et al. [9, 10] addressed the robust disturbance attenuation of a class of
nonlinear singularly perturbed systems and robust regulation of a class of nonlinear singularly
perturbed systems [11], and also position and velocity control of a flexible joint robot manipulator
via fuzzy controller based on singular perturbation analysis [12]. Fridman [13, 14] dealt with the
infinite horizon nonlinear quadratic optimal control problem for a class of non-standard nonlinear
singularly perturbed systems by invariant manifolds of the Hamiltonian system and its
decomposition into linear-algebraic Riccati equations.
In this study, we extend results of [13, 14] to the finite horizon by slow-fast manifolds of the
Hamiltonian system and its decomposition into SDREs. Our contribution is that, we used the
singular perturbation theory and SDRE method together. In the proposed method, first, the statedependent Hamiltonian matrix is derived for the system under study. Then, this matrix is
separated into the reduced-order slow and fast subsystems. Using the singular perturbation
theory, the state equations and SDREs are converted into outer layer, initial layer correction and
final layer correction equations, which are then solved to obtain the optimal control law. The block
diagram of the proposed method is shown in Figure 1.
Slow
Hamiltonian
matrix
Description
of the
system as
SDCs
Slow state
equations
Slow
SDREs
Optimal
control
law
State
dependent
Hamiltonian
matrix
Fast state
equations
Fast
Hamiltonian
matrix
Fast
SDREs
FIGURE 1: The design procedure stages in the proposed method.
The remainder of this study is organized as follows. Section 2 explains the structure of the
singularly perturbed system for optimization. Section 3 involves in the description of steps of the
design procedure in the proposed method. Section 4 presents the simulation results of the
system used in the proposed method. Finally, the study culminates with indication of remarks in
section 5.
2. PROBLEM FORMULATION
The following nonlinear singularly perturbed system is assumed:
Ex f ( x) B( x)u, x(t0 ) x0 ,
(1)
x1
ni
, xi R , i = 1,2 are the states of system, and x=0n is the equilibrium point of the
x
2
where x(t )
system (n=n1+n2). This system is full state observable, autonomous, nonlinear in the states, and
f (x , x )
B (x , x )
affine in the input. Moreover, f ( x) 1 1 2 , f i R ni , B( x) 1 1 2 , Bi R ni , i = 1,2
f 2 ( x1 , x2 )
B2 ( x1 , x2 )
are differentiable with respect to x1, x2 for a sufficient number of times. Furthermore, f(0n)=0n,
I n1 n1
0 n2 n1
n
B(x)0nm, xR and E
0 n1 n2
that >0 is a small parameter. Provided these, it
I n2 n2
m
is desired to obtain the optimal control law u(x)R

such that for k(x)R , k(0n)=0n and
n
m m
pointwise positive definite matrix R(x)R R , the following performance index is minimized.
(2)
tF
k T x k x uT Rx u dt
hxt F
t0
Suppose that k(x), R(x) are differentiable with respect to x1, x2 for a sufficient number of times.
Moreover, tF is chosen such that it is sufficiently large with respect to the dominant time constant
of the slow subsystem, and x(tF) is free.
3. THE PROPOSED METHOD

The singularly perturbed system (1) with performance index (2) is assumed. Defining the co-state
1 ( x1 , x2 )
ni
, i R , i = 1,2, the Hamiltonian function is obtained as (3):
(
x
,
x
)
2 1 2
vector ( x)
1
1
( x, u, ) k T ( x)k ( x) u T R( x)u 1T ( f1 ( x1 , x2 ) B1 ( x1 , x2 )u) 2T ( f 2 ( x1 , x2 ) B2 ( x1 , x2 )u).
2
2
(3)
According to the optimal control theory, the necessary conditions for optimization would be as
follow [2]:
(4a)
H T
x1 (
) f1 ( x1 , x2 ) B1 ( x1 , x2 )u, x1 (t 0 ),
1
H
x 2 ( )T f 2 ( x1 , x2 ) B2 ( x1 , x2 )u, x2 (t 0 ),
2
(4b)
T
T
T
f ( x) T
k
(
x
)
B
(
x
)
1 R( x)
T

k x
u
uT
1
x
x
x
x
2
x
1
1
1
1
1
(4c)
1 h
|t ,
, 1 x(t F )
2 x1 F
T
T
T
f ( x) T
H
k ( x)
1 R( x)
T B ( x)

k x
u
uT
2
x
x
x
2
x2
2
2
2
x2
, 2 x(t F )
1 h
|t ,
2 x2 F
H
R( x)u B1T ( x1, x2 )1 B2T ( x1, x2 )2 .
u
(4e)
3.1 Description of The System As SDCs (The first step)

A continuous nonlinear matrix-valued function A(x) always exists such that
f(x)=A(x)x
n
Where A(x):R R
nn
(4d)
(5)
is found by mathematical factorization and is, clearly, non-unique when
n>1. A suitable choice for matrix A(x) is Ax
1 f
x
0
| x x
d , where is a dummy variable that
was introduced in the integration [1]. Then, the relations (4) can be written as:
Ex A( x) x B( x)u, x(t0 )
(6a)
T
T
T
f ( x) T
H T
1 T R( x)
1 h
k ( x)
T B( x)
E ( )
, E x(t f ) |t f (6b)
k x
u
u
x
x
2
x
2 x
x
u R 1 x BT ( x)
(6c)
Considering that B(x) and R(x) are nonzero, the optimal control law is proportional to vector .
3.2
Description of The Hamiltonian Matrix As SDCs (The second step)
Assuming that K x
-1
1 k
x
0
| x x
d is available from k(x)=K(x)x and that Q(x)=KT(x)K(x) and
S(x)=B(x)R (x)B (x), the relations (6) can be rewritten as follow:
Ex A( x) x S ( x), x(t0 ) x0 ,
E Qx x AT ( x) [
i 1
A ( x) 1
xi i
x 2
i 1
(7a)
T
K ( x)
xi i
k x
x
R ( x)
T
1 T
1
Bx R x i i Rx B x
i 1
B ( x)
BxR x x } ],
m
i 1
(7b)
1 h
E x(t F ) |t F ,
2 x
Where,
A1i ( x)
x
1
Ai ( x)

x
Ani ( x)
x1
K1i ( x)
x
1
K i ( x)

x
K ni ( x)
x1
B1i ( x)
x
1
Bi ( x)

x
Bni ( x)
x1
A1i ( x)
xn
,
Ani ( x)
xn
K1i ( x)
xn
,
K ni ( x)
xn
B1i ( x)
xn
,
Bni ( x)
xn
(8a)
(8b)
(8c)
R1i ( x)
R1i ( x)
x
xn
1
Ri ( x)

.
x
R
(
x
)
R
(
x
)
mi
mi
x1
xn
(8d)
A( x) B( x) K ( x)
R( x)
and
are bounded in a
,
,
x
x
x
x
neighborhood of about the region. Then, the expression in the bracket will be ignored because
of being small. This approximation is asymptotically optimal, in that it converges to the optimal
control close to the origin as [2]. Thus, the relations (7) can be written as:
Assumption 1: A(x), B(x), Q(x), R(x),
S ( x) x
Ex A( x)
E Qx AT ( x)

(9)
Remark 1: Suppose that Tsi, TsF are dominant time constants of the slow subsystem for initial and
1
final layer correction, respectively. In other words, Tsi max
and
real eig slowJ i
TsF max
1
real eig slowJ F
where, Ji and JF are the Jacobian matrices of Hamiltonian system in

S ( x)
S ( x)
A( x)
A( x)
,JF
initial and final layer correction and, J i
t t .
T
T
t
t
0
Qx A ( x) x x0
Qx A ( x) xF0n
0n
0n
Note that (Tsi+TsF)/2 is the average time constant of the Hamiltonian system and the setting time
is fourfold of one, then a proper selection for tF is
tF > t0+2(Tsi+TsF)
(10)
3.3 The Singularly Perturbed SDRE in Finite Horizon

In the proposed method, co-sate vector , can be described as =P(x)x using the sweep method
P ( x , x ) P21T ( x1 , x2 )
ni n j
P( x) 11 1 2
[7] is the unique, non-symmetric,
, Pij R
P
(
x
,
x
)
P
(
x
,
x
)
22 1 2
21 1 2
positive-definite solution of the Riccati matrix equation. By differentiating with respect to time,
[3], where,
we can write:
(11)
P( x) x P ( x) x
By substituting (11) in (9) and with rearrangement of one, we have:
EP ( x) PT ( x) A( x) AT ( x) P( x) PT ( x)S ( x) P( x) Q( x) 0nn , Px(t F )
E 1
2
h
| x x d
0 x x
1
(12)
The relation (12) is called a SDRE for nonlinear singularly perturbed system in finite horizon. It
should be noted that the optimal control law is obtained by computing these Riccati matrices.
1/2
The solution conditions for SDRE are that {A(x),B(x)} be stabilizable and {A(x),(Q(x)) } be
n
detectable for xR . A sufficient test for the stabilizability condition of {A(x),B(x)} is to check that
n-1
the controllability matrix Mc= [B(x),A(x)B(x),,A (x)B(x)] has rank(Mc)=n,x. Similarly, a
1/2
1/2
sufficient test for detectability of {A(x), (Q(x)) } is that the observability matrix Mo=[(Q(x)) ,
1/2
1/2 n-1
(Q(x)) A(x),, (Q(x)) A (x)] has rank(Mo)=n, x [2]. Furthermore, the closed-loop matrix
A(x)-S(x)P(x) should be pointwise Hurwitz for x. Here, is any region such that the
1
Lyapunov function V ( x) xT P(x)d x is locally Lipschitz around the origin [2]. The SDRE in
(n1 n2 )( n1 n2 1)
differential equations that number of these equations is reduced
2
by using singular perturbation theory.
(12) consist
3.4
The Separated Hamiltonian Matrices
x1
,
1 x1 , x2
In the proposed method, by separating the slow and fast variables as X s
x2
Xf
, we can describe the optimization relations (9) in the form of the following
2 x1 , x2
singularly perturbed state-dependent Hamiltonian matrix:
X s

X f
H11( x1 , x2 )
H 21( x1 , x2 )
H12 ( x1 , x2 ) X s
,
H 22 ( x1 , x2 ) X f
(13)
Sij ( x1, x2 )
T and I, j=1,2. Thus, we assume that the 2n1
Qij ( x1, x2 ) A ji ( x1, x2 )
Aij ( x1, x2 )
Where, H ij x1, x2
eigenvalues of the system (13) are pointwise small and the remaining 2n2 eigenvalues are
pointwise large, corresponding to the slow and fast responses, respectively. The state and costate equations (13) constitute a singularly perturbed, two point boundary value problem
(TPBVP). Hence, the asymptotic solution is obtained as an outer solution in terms of the original
independent variable t, initial layer correction in terms of an initial stretched variable
and final layer correction in terms of a final stretched variable
tF t
composite solutions can be written as follow:
x1 (t , ) x1o (t , ) x1i ( , ) x1F ( , )
x2 (t , ) x2o (t , ) x2i ( , ) x2 F ( , )
P (t , ) P (t , ) P ( , ) P ( , )
so
si
sF
s
Pf (t , ) Pfo (t , ) Pfi ( , ) PfF ( , )

where t0 t t F ,0 t1
t F t0
,0 t 2
t t0
[5]. Thus, the
(14)
t F t0
. The first terms on the right hand sides of
the above relations represent the outer solution. The second and third terms represent boundarylayer corrections to the slow manifold near the initial and final times, respectively. Indices o, i and
F correspond to the outer layer, initial, and final correction layers. For any boundary condition on
the slow manifold, states and co-states are given by outer solution. For any boundary condition
out of the slow manifold, the trajectory rapidly approaches the slow manifold according to the fast
manifolds.
We now perform the slow-fast decomposition of the singularly perturbed state-dependent
Hamiltonian matrix, in which H22(x1,x2) must be non-singular for all x1, x2 (in what follows,
dependence upon x1, x2 is not represented, for convenience):
H11 H12 I
H12H 221
H 21 H 22 2 n1 2 n1
0 2 n 2 2 n 2
I 2 n 2 2 n 2

(15)
H11 H12 H 221H 21 0 2 n 2 n I
0 2 n1 2 n 2
1
2
2 n1 2 n1
1
22
0 2 n 2 2 n1
H 22 H 21 I 2 n 2 2 n 2
Stated differently:
I 2n 2n
1 1
0 2 n2 2 n1
H12H 221 X s H11 H12H 22 H 21
I 2 n2 2 n2 X f
0 2 n2 2 n1
0 2 n1 2 n2
H 22
I 2 n1 2 n1
1
H 22 H 21
New
co-sate
vector
can
be
described
as
new=Pnew(xnew)xnew,
(16)
0 2 n1 2 n2 X s
I 2 n2 2 n2 X f
where
xs
x new ,
x f
s xs , x f
Ps xs , x f Pa xs , x f
new
, s R n1 , f R n2 , and Pnew ( xnew )
. Then, the
f xs , x f
Pb xs , x f Pf xs , x f
new slow-fast variables are defined as follow:
xs
xf
s
Xs,
s xs , x f
(17a)
1
f
H 22 H 21X s X f ,
x
,
x
f s f
(17b)
Thus, (13) is converted to a new form:
X s H12 H 221 X f H11 H 12 H 221 H 21 s

X f H 22 f
(18)
Finally, the optimization equations in a singular perturbation model framework with the new
variables are obtained as:
s H11 H12 H 221 H 21 s H12 f
1
1
1
1
1
1
f H 22 H 21 H11 H12 H 22 H 21 s H 22 H 22 H 22 H 21 H 22 H 21 s H 22 H 22 H 21H12 f
Moreover, the separated state-dependent Hamiltonian matrices Hs(xs,xf) and

described in the form of the following:
(19)
H22(x1,x2) are
H s xs , x f H11x1 , x2 H12 x1 , x2 H 221 x1 , x2 H 21x1 , x2 O( )2n12n1

S s ( x1 , x2 ) n1n1
As ( x1 , x2 ) n1n1
O( )2n12n1 ,
T
Qs ( x1 , x2 ) n1n1 As ( x1 , x2 ) n1n1
A22 ( x1 , x2 ) n2n2
H f xs , x f H 22 x1 , x2 O( )2n22n2
Q22 ( x1 , x2 ) n2n2
S 22 ( x1 , x2 ) n2n2
O( )2n22n2 .
A22 ( x1 , x2 ) T n2n2
(20a)
(20b)
3.5 The slow-fast SDREs (The third step)

In the proposed method, using the singular perturbation theory, the subsystems (19) are
converted into outer-layer and boundary-layer correction subsystems. The separated SDRE
relations are, then derived and solved for obtaining the optimal control law.
Theorem 1: The singularly perturbed system (1) with performance index (2) is assumed. The
slow- fast state equations in the initial layer correction are obtained as follow:
x1o As ( x1o , x*2o x2i ) S s ( x1o , x*2o x2i ) Pso x1o , x1o |t 0 x1 (t0 ),
dx 2i
d
*
22 ( x1o , x 2o
(21a)
x 2i ) S 22 ( x1o , x * 2o x 2i ) P * 22o x * 2o x 2i
(21b)
*
*
*
*
*
21 ( x1o , x 2o x 2i ) S 21 ( x1o , x 2o x 2i ) Pso S 22 ( x1o , x 2o x 2i ) P 21o x1o , x 2i | t0 x 2 (t 0 ) x 2o (t 0 ),
Also, the slow- fast SDREs in the final layer correction are obtained as follow:
P P A ( x , x*2o ) A T ( x , x*2o ) P P S ( x , x*2o ) P
so
so so
1o
so
1o
so
so so
1o
Qso ( x1o , x
dPfF
so
2o )
(22a)
0 n1n1 , Pso |tF P11(t F ),
PfF A22o S 22o P*22o A22oT P*22o S 22o PfF PfF S 22o PfF , PfF |tF P22 (t F ) P*22o (t F ).
P11( x(t F )) P21T ( x(t F )) E 1
2
P21( x(t F )) P22 ( x(t F ))
where,
law is as follows:
(22b)
h
| x x d . Furthermore, the optimal control
0 x x
1
u R 1 ( x1o , x* 2o x2i ) B1T ( x1o , x* 2o x2i )Pso x1o B2T ( x1o , x* 2o x2i ) Pc x1o P* 22o PfF x* 2o x2i ) ,
(23)
where, Pso and PfF are the unique, symmetric, positive-definite solutions of (22), and
Pc P*22o PfF
I n2 n2 H 221 ( x1o , x*2o x2i ) H 21( x1o , x*2o x2i ) n1 n1 . The solution necessary
Pso
conditions of relations (21) and (22) are as follow:

*
*
*
*
{Aso(x1o,x 2o), Bso(x1o,x 2o)} and {A22o(x1o,x 2o), B2o(x1o,x 2o)} should be stabilizable for
{A (x ,x ),(Q (x ,x
x , x R R
x1o , x*2o R n1 R n2 .
so
1o
1o
*
2o
so
*
1/2
2o)) }
1o
n1
*
2o
n2
1/2
and {A22o(x1o,x 2o), (Q22o(x1o,x 2o)) }
should be detectable for
.
*
The outer equations (24) should have solutions (the slow manifolds) as x 2o(x1o,P11o),
*
*
P 21o(x1o,P11o) and P 22o(x1o,P11o)
(24a)
A21o S 21o P11o S22o P21o x1o A22o S 22o P22o x2o 0n ,
22o
P22o S 22o P21o A12o T P22o S 21o P11o P22o A21o Q21o 0 n1n2 ,
P22o A22o A22oT P22o P22o S22o P22o Q22o 0n2 n2 ,
(24b)
(24c)
It should be noted that in the above relations, all the elements of the state and Riccati matrices
are dependent on state variables, and have not been represented for simplicity.
Proofs of the theorems are given in appendix.
Remark 2: SDREs in (22) have n1n2 the less differential equations respect to (12).
4. EXAMPLE
Consider a single link flexible joint robot manipulator as it has been introduced in [11]. This link is
directly actuated by a D.C. electrical motor whose rotor is elastically coupled to the link. In this
example, the mathematical model of system is as follows:
(25)
Iq mgl sin(q ) k (q q ) 0
1
Iq2 q 2 k (q1 q2 ) u
FIGURE 2: Single link flexible joint robot manipulator
In Table 1 there is a complete list of notations of the mathematical model of a single link flexible
joint robot manipulator.
Notation
Description
q1
q2
u
I
J
mgl
K
angular positions of the link

angular positions of the motor
actuator force (motor torque)
the arm inertia
the motor inertia
the motor viscous friction
the nominal load in the rotor link
the stiffness coefficient of flexible joint
TABLE 1: Notations the mathematical model of a single link flexible joint robot manipulator.
Moreover, parameter values are given in Table 2.

parameter
Value of parameter
I
J
k
mgl
0.031(Kg.m )
2
0.004(Kg.m )
0.007
7.13
0.8 (N.m)
TABLE 2: Parameter values of the single link flexible joint robot manipulator.
x11 q1
x1

Defining x1 x12 q2 , x2 q 2 , x and =J, state equations are as follow:
x2
x13 q1
x13
0
10 0
x11
0
x

x2
3
0
12 mgl
u , x0 0 / s
k
k
0
x13
sin( x11) x11 x12 0

0/ s
I
I
I
1
kx11 kx12 x2
x 2
0
(26)
It is desired to obtain the optimal control law such that the following performance index is
minimized.
5
(27)

2
11
x122 x132 x2 2 u 2 dt
x13
x11
0
x
2
0, k ( x) x12 , R( x) 1, and
f ( x) mgl
,
B
(
x
)
In
this
example,
k
k
x13
0
sin( x11) x11 x12

I
I
I

kx11 kx12 x2
1
x2
h(x(tF))=0. Moreover, f(x), k(x) are differentiable with respect to x for a sufficient number of times
and x=04 is the equilibrium point of the system. Furthermore, t0=0, tF=5, P(x(tF))=044.
Step 1 (Description of the system as SDCs):
To solve the optimization problem, the nonlinear functions f(x), k(x) must first be represented as
SDCs. A suitable choice, considering [1], is as follows:
0
0 1 0
1
0
0 0 1
f
mgl sin(
x11) k k
A( x)
d
0 0
(28a)
x | x x
0
Ix11
I
I
k
k 0
1 0 0 0
1
0 1 0 0
k
(28b)
K ( x)
d
0 0 1 0
x | x x
0
0 0 0 1
Step 2 (Description of the Hamiltonian matrix as SDCs):
The separated Hamiltonian matrices can be derived:
0
0
1 0
0
0
k
k
1
0
0
2
2
2
1
1
1
k
mgl sin( x11) k
0 0
0
0
Ix11
I
I
(29a)
H s ( x1, x2 )
2
2
k
k
mgl sin( x11) k
1 k
0
0
Ix11
I
2 1
2 1
2 1
2
2
k
k
k
k
0
0
I
2 1
2 1
2 1
0
0
1 1
0
0
1
H 22 ( x1 , x2 )
(29b)
1
Step 3.1 (the outer equations):
The relations (24) have solutions as:
x * 2o
k ( x11o x12o ) Pso12 x11o Pso22 x12o Pso23x13o

2 1
(30a)
10
P*21o
k Pso12 k
2 1
P
k
k so22
2 1
Pso23
2
1
(30b)
(30c)
P*22o 2 1
k
Moreover, { Aso ( x1o , x * 2o )
2 1
mgl
sin(
x11o ) k
Ix11o
I
1

2
1

2
2 1 2k 2
2 2 1
1 2k
2
2 2 1
0
2 1 2k 2
2 1 2k 2
2 2 1
2 2 1
0
B2o ( x1o , x*2o ) 1, Q22o ( x1o , x*2o )
1
2
0}
0
1
0
k
1
2
0, Bso ( x1o , x * 2o )
, Q ( x , x * 2o )
so 1o
2
1
k
0
0
is stabilizable and detectable.
1
2
{A22o ( x1o , x * 2o ) ,
1} is also stabilizable and detectable.
Step 3.2 (the state equations):

According to (21), state variables relations in the initial layer correction are as follow:
x13o
10 0
k
(
x
x
)
P
x
P
x
P
x
11
o
12
o
so
12
11
o
so
22
12
o
so
23
13
o
, x (t ) 30
x1o
2
1 0
1
0 0 / s
mgl
k
k
sin( x11o ) x11o x12o

I
I
I
dx2i
7 k 10 Pso12 (t0 ) 3Pso22 (t0 ) 0
x2i 2 1, x2i (t0 )
d
2 1
(31a)
(31b)
Step 3.3 (the slow-fast SDREs):

The slow- fast SDREs in (22) have 3 the less equations respect to the original SDRE.
Considering (22), the SDRE relations in the final layer correction are as follow:
Ps _ 11o
T
Pso Ps _ 12o

T
Ps _ 13o
Ps _ 12o
P
s _ 22o
Ps _ 23o T
Ps _ 13o
Ps _ 23o , Pso (t F ) 033
Ps _ 33o
11
2mglPso13 sin( x11o )

2 P k 2kPso12 k 2 P 2 so12
1 so13
2
Ix11o
I
1
2
mglPso23 sin( x11o )
Pso23 Pso13
Pso12 Pso2 Pso12 Pso22 k
k
k
Ps _ 11o
Ix11o
I
2 1
2 1

mglPso33 sin( x11o ) Pso12 Pso23 kPso23
P
Ps _12o
Pso11 k so33
Ps _13o
2
Ix
I
1
11o
2
2
Pso23
2kPso22 k P so22
Ps _ 22o
2k
1
P

2
I
1
s _ 23o
Pso22 Pso23 kPso23
Pso33
Ps _ 33o
k
so12
I
2 1
2
P so23
2
P
so13
2
(32a)
dPfF
(32b)
2PfF 2 1 PfF 2 , PfF (t F ) 2 1
Step 3.4 (the optimal control law):

Moreover, the optimal control law is as follow:
2 1
( x11o x12o )
( Pso12 x11o Pso22 x12o Pso23 x13o ) ( PfF 2 1 ) x2i

2
1
(33)
The state equations and SDREs are two-point boundary value problem (TPBVP) and dependent
on state variables, but we have no state values in the whole interval [0,5]. To overcome this
problem we solve the above equations by an iterative procedure [3, 4]. Now, running the
simulation programs, Figures 3, 4 show the angular positions and velocities.
The angular positions(deg) and first angular velocity(deg/s)
The angular positions(deg) and first angular velocity(deg/s)
20
10
q1
10
q2
q2
dq1
q1
dq1
-10
-10
-20
-20
-30
-30
-40
-40
-50
-50
-60
-60
-70
0.5
1.5
2.5
3
Time(sec)
3.5
4.5
0.01
0.02
0.03
0.04 0.05 0.06

Time(sec)
0.07
0.08
0.09
0.1
FIGURE 3: The slow state variables (The angular positions of q1, q2 and angular velocity of q1 ).
12
Second angular velocity(deg/s)
Second angular velocity(deg/s)
100
90
dq2
dq2
80
80
70
60
60
50
40
40
20
30
0
20
-20
-40
10
0
0
0.5
1.5
2.5
Time(sec)
3.5
4.5
0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
Time(sec)
FIGURE 4: The fast state variable (angular velocity of q 2 ).
Also, Figures 5 and 6 show the Riccati gains.

The Riccati gains of P
20
15
10
5
0
-5
-10
-15
-20
0.5
1.5
2.5
Time(sec)
3.5
4.5
FIGURE 5: The Riccati gains of Ps.

0.35
0.25
0.3
0.25
0.2
0.2
0.15
0.15
0.1
0.1
0.05
0.05
0.5
1.5
2.5
3
Time(sec)
3.5
4.5
4.985
4.99
4.995
Time(sec)
FIGURE 6: The Riccati gains of Pf.
From Figures 3 and 5, it can be seen that for any initial and final conditions on the slow
manifold, for different values of , states are given by outer solution. On the other hand,
13
Figures 4 and 6 show that for any initial and final conditions out of the slow manifold, the
trajectories rapidly approach the slow manifold according to the fast manifolds. Moreover,
Figure 7 shows the optimal control law.
The optimal control law
1.2
0.8
0.6
0.4
0.2
-0.2
0.5
1.5
2.5
3
Time(sec)
3.5
4.5
FIGURE 7: The optimal control law u.
5. CONCLUSION
With the proposed method in this study, it is seen that the finite-horizon optimization problem of a
class of nonlinear singularly perturbed systems leads to SDREs for slow and fast state variables.
One of the advantages of SDRE method is that knowledge of the Jacobian of the nonlinearity in
the states, similar to HJB equation, is not necessary. Thus, the proposed method has not only
simplicity of the LQ method but also higher flexibility, due to adjustable changes in the Riccati
gains. On the other hand, one of the advantages of the singular perturbation theory is that it
reduces high-order systems into two lower-order subsystems due to the interaction between slow
and fast variables. Note that SDREs in the proposed method have n1n2 the less differential
equations respect to the original SDRE. Thus, the slow-fast SDREs have the simpler computing
than original SDRE and provide good approximations of one.
6. References
[1] S.P Banks and K.J. Mhana. Optimal Control and Stabilization for Nonlinear Systems.
IMA Journal of Mathematical Control and Information, vol. 9, pp. 179-196, 1992.
[2] T. imen. State-Dependent Riccati Equation (SDRE) Control: A Survey, in Proc. 17th
World Congress; the International Federation of Automatic Control Seoul, Korea, 2008,
pp. 3761-3775.
[3] H. Khaloozadeh and A. Abdolahi. A New Iterative Procedure for Optimal Nonlinear
Regulation Problem, in Proc. III International Conference on System Identification and
Control Problems, 2004, pp. 1256-1266.
[4] H. Khaloozadeh and A. Abdolahi. An Iterative Procedure for Optimal Nonlinear Tracking
Problem, in Proc. Seventh International Conference on Control, Automation, Robotics
and Vision, 2002, pp. 1508-1512.
[5] D.S. Naidu and A.J. Calise. Singular Perturbations and Time Scales in Guidance and
Control of Aerospace Systems: A Survey. Journal of Guidance, Control and Dynamics,
vol. 24, no.6, pp. 1057-1078, Nov.-Dec. 2001.
14
[6] W.C. Su, Z. Gajic and X. Shen. The Exact Slow-Fast Decomposition of the Algebraic
Riccati Equation of Singularly Perturbed Systems. IEEE Transactions on Automatic
Control, vol. 37, no. 9, pp. 1456-1459, Sep. 1992.
[7] Z. Gajic, X. Shen and M. Lim. High Accuracy Techniques for Singularly Perturbed
Control Systems-an Overview. PINSA, vol. 65, no. 2, pp. 117-127, March 1999.
[8] Z. Gajic, S. Koskie and C. Coumarbatch. On the Singularly Perturbed Matrix Differential
Riccati Equation, in Proc. CDC-ECC'05, 44th IEEE Conference on Decision and Control
and European Control Conference ECC 2005, Seville, Spain,2005, pp. 1417.
[9] R. Amjadifard and M.T. H. Beheshti. "Robust disturbance attenuation of a class of
nonlinear singularly perturbed systems." International Journal of Innovative Computing,
Information and Control (IJICIC), vol. 6, pp. 1349-4198, 2010.
[10] R. Amjadifard and M.T.H Beheshti. "Robust stabilization for a class of nonlinear
singularly perturbed systems." Journal of Dynamic Systems, Measurement and Control
(ASME), In Press, 2011.
[11] R. Amjadifard, M.J. Yazdanpanah and M.T.H. Beheshti. "Robust regulation of a class of
nonlinear singularly perturbed systems," IFAC, 2005.
[12] R. Amjadifard, S.E. Khadem and H. Khaloozadeh. "Position and velocity control of a
flexible Joint robot manipulator via fuzzy controller based on singular perturbation
analysis," IEEE International Fuzzy Systems Conference, 2001, pp. 348-351.
[13] E. Fridman. Exact Slow-Fast Decomposition of the Nonlinear Singularly Perturbed
Optimal Control Problem. System and Control Letters, vol. 40, pp. 121-131, Jun. 2000.
[14] E. Fridman. A Descriptor System Approach to Nonlinear Singularly Perturbed Optimal
Control Problem. Automatica, vol. 37, pp. 543-549, 2001.
Appendix A: The relation between the P(x) and Pnew(xnew)
In order to compute the optimal control law, the relations between the Riccati matrices
P ( x , x ) P21T ( x1, x2 )
Pa xs , x f
Ps xs , x f
P( x) 11 1 2
must be determined.
and Pnew ( xnew )
P
x
,
x
Pf xs , x f
P21( x1, x2 ) P22 ( x1, x2 )
b s f
l11 n2n1
l21 n2n1
Suppose that H 22 H 21
l12 n2n1
l22 n2n1 , according to (17), we have:
x f l11 l12 P11 x1 I l12 P21T x 2 ,
1
p s P11 I l12 P21T P21T (l11 l12 P11 ) O( 2 ) n n ,
1 1
T
1
2
p a ( P22l12 l 22 ) P21 (l11 l12 P11 ) O( ) n1n2 ,
T 1
T
pb I l12 P21 P21 ,
T
T 1
p f P22 l 22 P21 I l12 P21 ,
(A1)
Then, for =0, one can write:

I n1 n1
I n1 n1
P x , x xs
x1
P11x1 , x2
s s f
(A2a)
15
I n 2n 2
I n1n1
0
1
x1 n2 n1 x1 n2 n2 x 2
P x , x x f H 22 H 21
P11 x1 , x 2
P21 x1 , x 2
P22 x1 , x 2
f s f
Now, multiplying (A2b) by Pf xs , x f
(A2b)
I n2 n2 , the following relation is obtained.
Pf xs , x f I n n H 221H 21 n1 n1 P21x1 , x2 x1 Pf xs , x f P22 x1 , x2 x2 0 n

2
2
2
P11x1 , x2
In other words, we have:
(A3)
x1 xs O( )n1
(A4a)
P11( x1 , x2 ) Ps ( xs , x f ) O( )n1n1
(A4b)
P22 ( x1, x2 ) Pf ( xs , x f ) O( )n2 n2
(A4c)
P21( x1, x2 ) Pc xs , x f O( )n2 n1
Where, Pc xs , x f Pf xs , x f
(A4d)
I n2 n2 H 221H 21 n1 n1 . Also, for =0, we have:

P11x1 , x2
x1o xso
(A5a)
P11o ( x1o , x2o ) Pso ( xso , x fo )
(A5b)
P22o ( x1o , x2o ) Pfo ( xso , x fo )
(A5c)
P21o ( x1o , x2o ) Pco xso , x fo
(A5d)
Appendix B: Proof of Theorem 1

a)
The optimal control law
According to =P(x)x [3] and (A4), substituting Riccati matrices in (6c), the optimal control law
would result as in (23).
b)
The slow manifolds in boundary-layer correction
According to the singular perturbation theory, for =0, the fast variable should be derived with
respect to the slow variable. Substituting =0 in (19), the outer-layer equations are obtained as
follows:
(B1a)
H
H ,
s | 0 so
so
12o
fo
02n2 H22o fo .
(B1b)
Substituting (17b) in (B1b), the following relation is derived:
H2 1o X so H 22o X fo 02n2 .
(B2)
In other words, considering (14), we have:
A21o S 21o P11o S 22o P21o x1o A22o S 22o P22o x2o 0n2 ,
22o
P22o S 22o P21o A12o T P22o S 21o P11o P22o A21o Q21o 0 n1n2 ,
(B3a)
(B3b)
16
(B3c)
P22o A22o A22oT P22o P22o S 22o P22o Q22o 0n2 n2 ,

*
For which, x 2o(x1o,P11o), P 21o(x1o,P11o) and P 22o(x1o,P11o) are the solutions. The necessary
*
*
*
1/2
conditions for (B3) to be solvable, {A22o(x1o,x 2o), B2o(x1o,x 2o), (Q22o(x1o,x 2o)) } should be
pointwise stabilizable and detectable for x1o , x 2o R

In (B1a), H s
Hs
| 0
n1
R n2 [2].
for inside and out of the fast manifold, is separated as follows:
| 0
H11x1 , x2 H12 x1 , x2 H 221 x1 , x2 H 21x1 , x2 | 0

S s ( x1o , x2 )
As ( x1o , x2 )
Q ( x , x ) A ( x , x ) T ,
s 1o 2
s 1o 2
t0 t t0 t1 ,
(B4a)
Aso ( x1o , x*2o )

S so ( x1o , x* 2o )
T , t 0 t1 t t F .
*
*
Qso ( x1o , x 2o ) Aso ( x1o , x 2o )
(B4b)
Substituting (B4) in (B1a), we have:
x1o As ( x1o , x2 ) S s ( x1o , x2 ) Pso ( x1o , x2 )x1o , x1o |t0 x1 (t0 ),
t0 t t0 t1 ,
x1o
Aso ( x1o , x * 2o ) x1o S so ( x1o , x * 2o ) Pso ( x1o , x * 2o ) x1o

, Pso |tF P11(t F ),
P x P x
T
*
*
*
so 1o so 1o Qso ( x1o , x 2o ) x1o Aso ( x1o , x 2o ) Pso ( x1o , x 2o ) x1o
*
(B5a)
t 0 t1 t t F .
(B5b)
1/2
Thus, assuming that {Aso(x1o,x 2o), Bso(x1o,x 2o), (Qso(x1o,x 2o)) } is pointwise stabilizable-
detectable for x1o , x 2o R 1 R 2 [2], with rearrangement of (B5b), the SDRE of the slow
variable is obtained as (22a).
Remark 3: Note that under assumption of above, Pso is unique, symmetric, positive definite
solution of the SDRE (22a) that produces a locally asymptotically stable closed loop solution [2].
Thus the closed-loop matrix As(x1o,x2)-Ss(x1o,x2)Pso is pointwise Hurwitz for (x1o,x2)12.
Here, 12 is any region such that the Lyapunov function is locally Lipschitz around the origin.
c)
The fast manifold in initial layer correction

t t0
Since the time scale will be changed as
in the initial layer correction, the time derivative
(.)
d (.)
in this scale will be changed as
in forward time. Considering (4b), we have:
d
dt
(B6)
dx2
A22 ( x1o , x2 ) x2 A21( x1o , x2 ) x1o B2 ( x1o , x2 )u, x2 |t0 x2 (t 0 )
d
Substituting (23) in (B6), according to (A4) and (14), the fast state equation in initial layer is
obtained as (21b).
d)
The fast manifold in final layer correction

t t
Since the time scale will be changed as F
in the final layer correction, the time derivative
in this scale will be changed as
d (.)
d (.)

in backward time:
d
dt
d s
1
d H11 H12H 22 H 21 s H12 f
d
f H221H 21 H11 H12H 221H 21 s H 221H 22H 221H 21 H 221H 21 s H22 H 221H 21H12 f
d
(B7)
17
Substituting =0 in (B7), we have s 02n1 . Therefore, the final layer correction equation is
obtained as:
d f
d
H 22 ( x1o , x * 2o ) f , f | 0 f x1 (t F ), x 2 (t F ).
(B8)
Now, substituting (20b) and (17b) in (B8), we have:
dx f
A22 ( x1o , x * 2o ) x f S 22 ( x1o , x * 2o ) Pf x f

d

, Pf |tF P22 (t F ), t F t 2 t t F .
T
*
*
P dx f dPf x Q22 ( x1o , x 2o ) x f A22 ( x1o , x 2o ) Pf x f
f d d f
*
(B9)
1/2
Thus, assuming that {A22o(x1o,x 2o), B2o(x1o,x 2o), (Q22o(x1o,x 2o)) } is stabilizable-detectable for
x1o , x*2o R n1 R n2 [2], according to (A5) and (14), the SDRE of the fast variable is obtained as
(22b).
Remark 4: Note that under assumption of above, Pf is unique, symmetric, positive definite
solution of the SDRE (22b) that produces a locally asymptotically stable closed loop solution [2].
*
Thus, the closed-loop matrix A22(x1o,x2)-S22(x1o,x2)P 22o is pointwise Hurwitz for (x1o,x2)12.
Here, 12 is any region such that the Lyapunov function is locally Lipschitz around the origin.
18

Designing SDRE-Based Controller For A Class of Nonlinear Singularly Perturbed Systems

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Designing SDRE-Based Controller For A Class of Nonlinear Singularly Perturbed Systems

Uploaded by

Copyright:

Available Formats

Seyed Mostafa Ghadami, Roya Amjadifard & Hamid Khaloozadeh