J. of the Braz. Soc. of Mech. Sci. & Eng. Copyright © ©© © 2008 by ABCM OctoberDecember 2008, Vol. XXX, No. 4 / 279
Marat Rafikov
marat9119@yahoo.com.br
UFABC and UNIJUI
Dep. de Física, Estatística e Matemática
C.P. 560
98700000 Ijui, RS, Brazil
José Manoel Balthazar
Senior Member, ABCM
jmbaltha@rc.unesp.br
UNESP
Dep. de Estatística, Matemática Apli. e Comp.
C.P. 178
13500230 Rio Claro, SP, Brazil
Ângelo Marcelo Tusset
a_m_tusset@yahoo.com.br
UnC
Dep. de Ciência da Computação
89460000 Canoínhas, SC, Brazil
An Optimal Linear Control Design for
Nonlinear Systems
This paper studies the linear feedback control strategies for nonlinear systems. Asymptotic
stability of the closedloop nonlinear system is guaranteed by means of a Lyapunov
function, which can clearly be seen to be the solution of the HamiltonJacobiBellman
equation thus guaranteeing both stability and optimality. The formulated Theorem
expresses explicitly the form of minimized functional and gives the sufficient conditions
that allow using the linear feedback control for nonlinear system. The numerical
simulations the Duffing oscillator and the nonlinear automotive active suspension system
are provided to show the effectiveness of this method.
Keywords: optimal control, nonlinear system, duffing oscillator, active suspension system,
chaotic attractor
Introduction
1
Since the first publications (Krasovskii, 1959), (Kalman and
Beltram, 1960) and (Letov, 1961), in the early 1960s, the Lyapunov
function techniques have been used in studying optimal control
problems.
It is well known that the nonlinear optimal control problem can
be reduced to the HamiltonJacobiBellman partial differential
equation (Bryson and Ho, 1975). There are many difficulties in its
solution, in general case. There are, in current literature (see, for
example, Bardi and CapuzzoDolcetta, 1997), several methods that
may be used in obtaining a numerical solution of the Hamilton
JacobiBellman partial differential equation. In particular case, the
quadratic Lyapunov function is a solution of the HamiltonJacobi
Bellman equation for the linear system with the quadratic functional.
In recent years, the idea that a Lyapunov function for the nonlinear
system can be an analytical solution of the HamiltonJacobiBellman
equation has become popular.
In Bernstein (1993) a unified framework for continuoustime
nonlinearnonquadratic problems was presented in constructive
manner. The results of Bernstein (1993) are based on the fact that
steadystate solution of the HamiltonJacobiBellman equation is a
Lyapunov function for the nonlinear system thus guaranteeing both
stability and optimality.
In Haddad et al. (1998) the framework developed in Bernstein
(1993) was extended to the problem of optimal nonlinear robust
control. There are no systematic techniques for obtaining the
Lyapunov functions for general nonlinear systems in this case, but
this approach can be applied to systems for which the Lyapunov
functions can be found.
In Rafikov and Balthazar (2004) the nonquadratic nonlinear
Lyapunov function was proposed to resolve the optimal nonlinear
control design problem for the Rössler system.
This paper studies the linear feedback control strategies for
nonlinear systems.
Paper accepted July, 2008. Technical Editor: Marcelo Amorim Savi.
Asymptotic stability of the closedloop nonlinear system is
guaranteed by means of a Lyapunov function, which can clearly be
seen to be the analytical solution of the HamiltonJacobiBellman
equation thus guaranteeing both stability and optimality. The
formulated Theorem expresses explicitly the form of minimized
functional and gives the sufficient conditions that allow using the
linear feedback control for nonlinear system.
Nomenclature
A = bounded matrix (nxn)
B = constant matrix (nxm)
s
b = damping force, N/m/s
l
s
b = damping force linear, N/m/s
nl
s
b = damping force nonlinear, N/m/s
y
s
b = damping force symmetric, N/m/s
g = vector (nx1)
G = matrix (nxn)
s
k = suspension stiffness, N/m
l
s
k = suspension stiffness linear, N/m
nl
s
k = suspension stiffness nonlinear, N/m
t
k = tyre stiffness, N/m
s
m = sprung mass, kg
u
m = unsprung mass, kg
P = Riccati equation (nxn)
Q = matrix (nxn)
V = Lyapunov function
t = time, s
x = displacement of the flexible beam, m
x& = velocity of the flexible beam, m/s
x& & = acceleration of the flexible beam, m/s
2
y = state vector
c
z = vertical displacement of the sprung mass , m
Marat Rafikov et al
280/ Vol. XXX, No. 4, OcotoberDecember 2008 ABCM
c
z& = vertical velocity of the sprung mass , m/s
c
z& & = vertical acceleration of the sprung mass , m/s
2
w
z = vertical displacement of the unsprung mass , m
w
z& = vertical velocity of the unsprung mass , m/s
w
z& & = vertical acceleration of the unsprung mass , m/s
2
r
z = disturbance caused by road irregularities, m
Greek Symbols
α = stiffness parameter, N/m
ζ = viscous damping coefficient, N/m/s
γ = amplitude external input, m
ω = angular frequency, rad/s
Γ = neighborhood
Subscripts
s sprung
u unsprung
t tyre
Superscripts
l linear
nl nonlinear
y symmetric
Linear Design for Nonlinear System
We consider the nonlinear controlled system
Bu y g y t A y + + = ) ( ) ( & ,
0
) 0 ( y y = (1)
where
n
y ℜ ∈ is a state vector,
n n
t A
×
ℜ ∈ ) ( is a bounded matrix,
which elements are time dependent,
m n
B
×
ℜ ∈ is a constant matrix,
m
u ℜ ∈ is a control vector, and
n
y g ℜ ∈ ) ( is a vector, which
elements are continuous nonlinear functions, . 0 ) 0 ( = g
We remark that the choice of A(t) is not unique, and this
influences the performance of the resultant controller.
Assume that:
y y G y g ) ( ) ( = (2)
where
n n
y G
×
ℜ ∈ ) ( is a bounded matrix, which elements depend on
y . If we assumed (2), the dynamic system (1) has the following
form:
Bu y y G y t A y + + = ) ( ) ( & (3)
Next, we present an important result, concerning a control law
that guarantees stability for a nonlinear system and minimizes a
nonquadratic performance funcional.
Theorem 1. If there exist matrices Q(t) and R(t), positive definite,
being Q symmetric, such that the matrix:
) ( ) ( ) ( ) ( ) ( ) , (
~
y G t P t P y G t Q t y Q
T
− − = (4)
is positive definite for the bounded matrix G , then the linear
feedback control:
y t P B R u
T
) (
1 −
− = (5)
is optimal, in order to transfer the nonlinear system (3) from an
initial to final state:
0 ) ( =
f
t y (6)
minimizing the functional:
dt u R u y Q y J
T T
t
f
)
~
(
0
+ =
∫
(7)
where the symmetric matrix ) (t P is evaluated through the solution of
the matrix Ricatti differential equation:
0
1
= + − + +
−
Q P B PBR P A PA P
T T
&
(8)
satisfying the final condition:
0 ) ( =
f
t P (9)
In addition, with the feedback control (5), there exists a
neighborhood Γ ⊂ Γ
0
,
n
ℜ ⊂ Γ , of the origin such that if
0 0
Γ ∈ y ,
the solution , 0 , 0 ) ( > = t t y of the controlled system (3) is locally
asymptotically stable, and . ) 0 (
0 0 min
y P y J
T
=
Finally, if
n
ℜ = Γ then the solution , 0 , 0 ) ( > = t t y of the
controlled system (3) is globally asymptotically stable.
Proof. Lets consider the linear feedback control (5) with matrix P
determined by equation (8) which transfers the nonlinear system (3)
from an initial to a final state (6), minimizing the functional (7)
where the matrix Q
~
need to be determined.
According to the Dynamic Programming rules one knows that if
the minimum of functional (7) exists and if V is a smooth function of
the initial conditions, then it satisfies the HamiltonJacobiBellman
equation (Bryson and Ho, 1975):
0
~
min = 
¹

\

+ + u R u y Q y
dt
dV
T T
u
(10)
Considering a function:
y t P y V
T
) ( = (11)
where ) (t P is a symmetric matrix positive definite and it satisfies the
differential Riccati equation (8).
Note that the derivative of the function V, evaluated in the
optimal trajectory with control given by (5) is:
y P y y t P y y t P y V
T T T
&
&
&
&
+ + = ) ( ) ( − + = ) ( ) ( [ y G y t A y
T T T T
+ + −
−
y t P y y t P B R B t P y
T T T T
) ( ) ( ] ) ( ) (
1
&
] ) ( ) ( ) ( [ ) (
1
y t P B R B y y G y t A t P y
T T −
− + +
Then, substituting V
&
in the HamiltonJacobiBellman equation
(10) one obtains:
P y G P B PBR PA P A P y
T T T T
) ( [
1
+ − + +
−
&
0 ]
~
) ( = + + y Q y PG
An Optimal Linear Control Design for Nonlinear Systems
J. of the Braz. Soc. of Mech. Sci. & Eng. Copyright © ©© © 2008 by ABCM OctoberDecember 2008, Vol. XXX, No. 4 / 281
Then:
) ( ) ( ) ( ) ( ) (
~
y G t P t P y G t Q Q
T
− − = .
Note, that for positive definite matrices Q
~
and R, the derivative
of the function (11) is given by u R u y Q
~
y V
T T
− − =
&
, and, it is negative
definite. Then, the function (11) is Lyapunov function, and the
controlled system (3) is locally asymptotically stable. Integrating the
derivative of the Lyapunov function (11) given by
u R u y Q y V
T T
− − =
~
&
along the optimal trajectory, we obtain
. ) 0 (
0 0 min
y P y J
T
=
Finally, if
n
ℜ = Γ , global asymptotic stability follows as a direct
consequence of the radial unboundedness condition for the Lyapunov
function (11) ∞ → ) ( y V as ∞ → y .
We remark that according to the optimal control theory of linear
systems with quadratic functional (Anderson and Moor, 1990) the
solution of the nonlinear differential Riccati equation (8) is positive
definite and symmetric matrix 0 > P for all 0 > R e 0 ≥ Q given,
one can conclude the Theorem proof.
If the time interval is infinite and A, B, Q and R are matrix with
constant elements, the positive defined matrix P is the solution of the
nonlinear, matrix algebraic Riccati equation
0
1
= + − +
−
Q P B PBR P A PA
T T
. (12)
Linear Design for Duffing Oscilator
We will apply the proposed method to the Duffing oscillator that
is one of the paradigms of nonlinear dynamics.
The Duffing oscillator with the control law ) (t U is described by
the following nonlinear differential equation:
) ( cos 2
3
t U t x x x x + + − − − = ω γ ζ α & & & (13)
where α is the stiffness parameter, 0 > ζ is the viscous damping
coefficient, γ and ω are the amplitude and frequency of the
external input, respectively.
For 0 < α , the Duffing oscillator without control can be
interpreted as a model of a periodically forced steel beam which is
deflected toward the two magnets, as shown in Fig. 1.
Figure 1. Mechanical interpretation of Duffing oscillator.
Let the desired trajectory be a function ) (
~
t x . Then the desired
regimen is described by the following equation:
u t x x x x
~
cos
~
2
~ ~ ~ 3
+ + − − − = ω γ ζ α
& & &
(14)
where u
~
is a control function which maintains the Duffing oscillator
in the desired trajectory. If the function ) (
~
t x is a solution of equation
(13) without the control term then 0
~
= u .
Subtracting (14) from (13) and defining:
(
¸
(
¸
−
−
=
x x
x x
y
&
&
~
~
(15)
we will obtain the following system:
u x x y y y y
y y
+ + + − − − =
=
3 3
1 2 1 2
2 1
~
)
~
( 2ζ α &
&
(16)
where u U u
~
− = is feedback control.
So the equation (3) in this case has the following form:
u
y
y
x x x y x y
y
y
y
y
(
¸
(
¸
+
(
¸
(
¸
(
¸
(
¸
− + − + −
+
+
(
¸
(
¸
(
¸
(
¸
− −
=
(
¸
(
¸
1
0
0
0
~ ~
)
~
( )
~
(
0
2
1 0
2
1
2
1
2
1
2
1
2
1
ζ α &
&
(17)
Let the desired trajectory be a periodic orbit:
t t x sin cos 2
~
+ = (18)
and the parameters 1 , 125 . 0 , 1 = = − = γ ζ α and 1 = ω .
These are exactly the same desired trajectory and parameters as
reported by Sinha et al. (2000). For these parameters, the system (14)
without control possesses a chaotic attractor, shown in Fig. 2.
Figure 2. Phase portrait of uncontrolled Duffing oscillator.
Choosing
(
¸
(
¸
=
1 0
0 1
Q and [ ] 1 = R , then one obtains
(
¸
(
¸
=
1771 . 2 4142 . 2
4142 . 2 6825 . 3
P by solving the Riccati equation (12) using the
LQR function in MATLAB
®
. By evaluating the Q
~
, one has:
Marat Rafikov et al
282/ Vol. XXX, No. 4, OcotoberDecember 2008 ABCM
= − − = ) ( ) (
~
y PG P y G Q Q
T
−
(
¸
(
¸
(
¸
(
¸
−
−
22 12
12 11 1
0 0
)
~
, ( 0
p p
p p x y
Q
ϕ
(
(
¸
(
¸
−
(
¸
(
¸
−
0 )
~
, (
0 0
1 22 12
12 11
x y p p
p p
ϕ
(
¸
(
¸
+ =
0
2
)
~
, (
22
22 12
1
p
p p
x y Q ϕ
where
2
1
2
1 1
~ ~
)
~
( )
~
( )
~
, ( x x x y x y x y + + + + = ϕ .
Note that the function )
~
, (
1
x y ϕ has its minimum value
2
min
~
4
3
x = ϕ in x y
~
2
3
1
− = because
1 1
~
3 2 x y + = ′ ϕ 0 2 > = ′ ′ ϕ , then
2
1
~
4
3
)
~
, ( x x y ≥ ϕ and, admitting 1
~
3 ≥ ≥ x , one can evaluate Q
~
as:
=
(
¸
(
¸
+ ≥
(
¸
(
¸
+ ≥
0
2
4
3
0
2
~
4
3 ~
22
22 12
22
22 12 2
p
p p
Q
p
p p
x Q Q
(
¸
(
¸
1 632825 . 1
632825 . 1 6213 . 4
positive definite.
Finally, we can conclude that the optimal function u has the
following form:
2 1
1771 . 2 4142 . 2 y y u − − = . (19)
The phase portrait behavior of the controlled Duffing oscillator is
presented in Fig. 3.
Figure 3. Phase portrait of controlled Duffing oscillator.
Linear Design for Nonlinear Automotive Active
Suspension System
Consider a quarter vehicle model that captures the fundamental
features of the suspension design of the vehicle. A typical quartercar
suspension system is shown in Fig. 4.
Figure 4. Typical quartercar suspension system.
In this figure,
s
m represents the sprung mass, which corresponds
to 1/4 of the body mass, and the unsprung mass
u
m represents the
wheel at one corner. The parameters
t
k ;
s
k ;
s
b are the tyre stiffness,
the suspension stiffness, and the damping rate of the suspension,
respectively. The control signal u is generated by the actuator,
c
z
and
w
z denote the vertical displacement of the sprung mass and the
unsprung mass, respectively. The disturbance w is caused by road
irregularities.
Several methods to solve the active suspension problem have
been proposed. The most part of these research contributions are
based on linear timeinvariant suspension models for control design.
Thompson (1976) was the first to explore the use of optimal
control techniques to design an active law.
Active suspension design, using linear parameter varying control,
was considered by Gaspar et al. (2003).
The real physical systems always include nonlinear components
which must be taken into consideration. The suspension force
generated by the hydraulic actuator is inherently nonlinear; the
dynamic characteristics of suspension components, i.e., dampings
and springs, have nonlinear properties.
To resolve the active suspension problem for nonlinear system in
this work we will apply the linear feedback control design above
considered.
The equations of motion of the vehicle active suspension system
are (Gaspar et al., 2003):
( )
( )
u z z k
z z b z z b z z k z z k z m
u z z b z z b z z k z z k z m
r w t
c w
y
s c w
l
s c w
nl
s c w
l
s w u
c w
y
s c w
l
s c w
nl
s c w
l
s c s
+ − −
− − + − − − − − − =
− − − − + − + − =
) (
) ( ) (
) ( ) (
3
3
& & & & & &
& & & & & &
(20)
Here, parts of the nonlinear suspension damping
s
b are
l
s
b ,
nl
s
b
and
y
s
b .
The
l
s
b coefficient affects the damping force linearly, and
y
s
b
describes the symmetric behavior of the characteristics. Parts of the
nonlinear suspension stiffness
s
k are a linear coefficient
l
s
k and a
nonlinear one
nl
s
k .
Considering that the disturbance w, which is caused by road
irregularities, have a constant value, the desired trajectory is:
An Optimal Linear Control Design for Nonlinear Systems
J. of the Braz. Soc. of Mech. Sci. & Eng. Copyright © ©© © 2008 by ABCM OctoberDecember 2008, Vol. XXX, No. 4 / 283
0
~
,
~
, 0
~
,
~
= = = =
w w c c
z w z z w z
& &
(21)
Noting that the desired trajectory is a solution of the system (20)
without control and defining:
[ ]
T
w w c c
z w z z w z y & & − − = (22)
we will obtain the following equation in form of (1):
Bu y g y A y + + = ) ( & (23)
where:
(
(
(
(
(
(
(
¸
(
¸
−
+
−
− −
=
u
l
s
u
t
l
s
u
l
s
u
l
s
s
l
s
s
l
s
s
l
s
s
l
s
m
b
m
k k
m
b
m
k
m
b
m
k
m
b
m
k
A
) (
1 0 0 0
0 0 1 0
,
(
(
(
(
(
(
¸
(
¸
−
=
u
s
m
m
B
1
0
1
0
(
(
(
(
(
(
(
(
¸
(
¸
− + − −
− − −
=
2 4
3
1 3
2 4
3
1 3
) (
0
) (
0
) (
y y
m
b
y y
m
k
y y
m
b
y y
m
k
y g
u
y
s
u
nl
s
s
y
s
s
nl
s
. (24)
Considering:
y
g g g g
g g g g
y y G y g
(
(
(
(
¸
(
¸
= =
44 43 42 41
24 23 22 21
0 0 0 0
0 0 0 0
) ( ) (
where:
); , (
; ) (
4 2 44 42 24 22
2
1 3 43 41 23 21
y y
m
b
g g g g
y y
m
k
g g g g
s
y
s
s
nl
s
δ = = − = − =
− − = = − = − =
¹
´
¦
≠ − −
= −
=
0 if ) (
0 if 0
) , (
2 4 2 4
2 4
4 2
y y y y sign
y y
y y δ
we present the equation (23) in form (3).
The proposed linear feedback design procedure has been applied
to the quarter car suspension with the following values of the
parameters (Gaspar et al., 2003): kg m
s
290 = , kg m
u
40 = ,
m N k
l
s
/ 23500 = , m N k
nl
s
/ 2350000 = , m N k
t
/ 190000 = ,
s m N b
l
s
/ / 700 = e s m N b
y
s
/ / 400 = .
For these parameters the matrix A has the following form:
(
(
(
(
(
¸
(
¸
− −
− −
=
5 . 17 5 . 5337 5 . 17 5 . 587
1 0 0 0
414 . 2 034 . 81 414 . 2 035 . 81
0 0 1 0
A .
Choosing
(
(
(
(
(
¸
(
¸
=
100 0 0 0
0 100 0 0
0 0 100 0
0 0 0 100
Q , ] 00001 . 0 [ = R , then one
obtains:
(
(
(
(
(
¸
(
¸
−
− − −
−
−
=
0187 . 1 5918 . 2 2583 . 0 5828 . 1
5918 . 2 1816 . 6411 1348 . 110 2601 . 584
2583 . 0 1348 . 110 2626 . 10 0896 . 12
5828 . 1 2601 . 584 0896 . 12 4481 . 687
P
by solving the Riccati Eq. (12).
Finally, we can conclude that the optimal function u has the
following form:
4 3 2 1
7889 . 2457 7352 . 31494 7037 . 2892 5054 . 211 y y y y u − − + = (25)
For numerical simulations the disturbances of the initial
conditions m y 05 . 0
1
− = and m y 05 . 0
3
− = were considered. Fig. 5
and 6 show
1
y and
3
y variations without and with control (25),
respectively.
Figure 5. Comparison of the time traces of the variable of the controlled
system (23) and the same system without control.
Figure 6. Comparison of the time traces of the variable of the controlled
system (23) and the same system without control.
Marat Rafikov et al
284/ Vol. XXX, No. 4, OcotoberDecember 2008 ABCM
It is difficult to analyze the matrix Q
~
analytically in this case.
Fig. 7 shows the positive function y Q y t L
T
~
) ( = , calculated in
optimal trajectory.
Figure 7. Time trace of the positive function.
Conclusions
The optimal linear feedback control strategies for nonlinear
systems were proposed. Asymptotic stability of the closedloop
nonlinear system is guaranteed by means of a Lyapunov function
which can clearly be seen to be the solution of the HamiltonJacobi
Bellman equation thus guaranteeing both stability and optimality.
The formulated Theorem expresses explicitly the form of
minimized functional and gives the sufficient conditions that allow
using the linear feedback control for nonlinear system.
The numerical simulations for the Duffing oscillator and the
nonlinear automotive active suspension system are provided to show
the effectiveness of this method.
Finally we note that the proposed method can be applied to a
large class of nonlinear and chaotic systems.
Acknowledgements
The work was partially supported by the Brazilian Government
agencies CNPq and FAPESP.
References
Anderson, B.D.O. and Moor, J.B., 1990, “Optimal Control: Linear
Quadratic Methods”, PrenticeHall, NY.
Bardi, M. and CapuzzoDolcetta, I., 1997, “Optimal Control and
Viscosity Solutions of HamiltonJacobiBellman Equations”, Birkhäuser,
Boston.
Bernstein, D.S., 1993, “Nonquadratic cost and nonlinear feedback
control”, Int. J. Robust Nonlinear Contr. 3, pp. 211229.
Bryson, JR., A.E. and Ho, Y. C., 1975, “Applied optimal control:
Optimization, Estimation, and Control”. Hemisphere Publ. Corp., Washington
D.C..
Gaspar, P., Sazaszi, I. and Bokor, J., 2003, “Active suspension design
using linear parameter varying control”, International Journal of Vehicle
Autonomous Systems, Vol. 1, No. 2, pp. 206221.
Haddad, W.M., Chellaboina, V. S. and Fausz, J.L., 1998, “Robust
nonlinear feedback control for uncertain linear systems with nonquadratic
performance criteria”, Systems Control Lett., 33, pp. 327338.
Krasovskii, N. N., 1959, “On the theory of optimum control”, Appl. Math.
Mech., vol. 23, pp. 899919.
Kalman, R.E. and Bertram, J.E., 1960, “Control system analysis and
design via the second method of Lyapunov”, Journal of Basis Engineering,
vol. 82, pp. 371393.
Letov, A.M., 1961, “The analytical design of control systems”,
Automation and Remote Control, 22, pp. 363372.
Rafikov, M. and Balthazar, J. M., 2004, “On an optimal control design for
Rössler system”, Phys. Lett., A 333, pp. 241245.
Sinha, S.C., Henrichs, J.T. and Ravindra, B.A., 2000, “A general
approach in the design of active controllers for nonlinear systems exhibiting
chaos”, Int. J. Biff. Chaos, 10 (1), pp. 165178.
Thompson, A. G., 1976, “An active suspension with optimal linear state
feedback”, Vehicle System Dynamics, Vol. 5, pp. 187–203.
u ∈ ℜ m is a control vector. (1) where y ∈ ℜ n is a state vector. No. there exists a neighborhood Γ0 ⊂ Γ . N/m ζ = viscous damping coefficient. and J min = y0 P (0) y0 . such that the matrix: ~ Q ( y . t > 0. t > 0. minimizing the functional (7) ~ where the matrix Q need to be determined. rad/s = neighborhood (8) Subscripts s sprung u unsprung tyre t Superscripts l linear nl nonlinear y symmetric satisfying the final condition: P (t f ) = 0 (9) In addition. which elements are time dependent. the dynamic system (1) has the following form: & y = A(t ) y + G ( y ) y + Bu (11) (3) Next. being Q symmetric. concerning a control law that guarantees stability for a nonlinear system and minimizes a nonquadratic performance funcional. 1975): ~ dV min + yT Q y + u T R u = 0 u dt Considering a function: V = y T P (t ) y (10) (2) where G ( y ) ∈ ℜ n× n is a bounded matrix. If we assumed (2). m ~ J = ∫ (yT Q y + uT R u ) dt 0 tf (7) ω Γ where the symmetric matrix P (t ) is evaluated through the solution of the matrix Ricatti differential equation: & P + PA + AT P − PBR −1B T P + Q = 0 = angular frequency. of the controlled system (3) is globally asymptotically stable. OcotoberDecember 2008 ABCM . According to the Dynamic Programming rules one knows that if the minimum of functional (7) exists and if V is a smooth function of the initial conditions. then the linear feedback control: u = − R B P (t ) y −1 T (5) 280/ Vol. XXX. m (6) minimizing the functional: Greek Symbols α = stiffness parameter. positive definite. A(t ) ∈ ℜ n×n is a bounded matrix. Linear Design for Nonlinear System We consider the nonlinear controlled system & y = A(t ) y + g ( y ) + Bu . t ) = Q (t ) − G T ( y ) P (t ) − P (t )G ( y ) & − yT P (t ) B( R −1 )T BT ]P (t ) y + yT P (t ) y + + yT P(t )[ A(t ) y + G ( y ) y − B R −1BT P(t ) y ] & Then. We remark that the choice of A(t) is not unique. the solution y (t ) = 0. Lets consider the linear feedback control (5) with matrix P determined by equation (8) which transfers the nonlinear system (3) from an initial to a final state (6).Marat Rafikov et al & zc &&c z zw & zw &&w z zr = vertical velocity of the sprung mass . m 2 is optimal. If there exist matrices Q(t) and R(t). and this influences the performance of the resultant controller. 4. in order to transfer the nonlinear system (3) from an initial to final state: y (t f ) = 0 = vertical displacement of the unsprung mass . which elements are continuous nonlinear functions. Assume that: g ( y) = G( y) y Proof. y (0) = y0 Finally. m/s = vertical velocity of the unsprung mass . and g ( y ) ∈ ℜ n is a vector. with the feedback control (5). of the controlled system (3) is locally T asymptotically stable. which elements depend on y . substituting V in the HamiltonJacobiBellman equation (10) one obtains: ~ & yT [ P + AT P + PA − PBR −1BT P + GT ( y ) P + PG ( y ) + Q ] y = 0 (4) is positive definite for the bounded matrix G . Γ ⊂ ℜ n . m/s2 = disturbance caused by road irregularities. we present an important result. g (0) = 0. N/m/s γ = amplitude external input. where P (t ) is a symmetric matrix positive definite and it satisfies the differential Riccati equation (8). B ∈ ℜ n× m is a constant matrix. evaluated in the optimal trajectory with control given by (5) is: & & & & V = y T P (t ) y + y T P (t ) y + y T P y = [ yT AT (t ) + yT GT ( y ) − Theorem 1. of the origin such that if y0 ∈ Γ0 . then it satisfies the HamiltonJacobiBellman equation (Bryson and Ho. m/s = vertical acceleration of the sprung mass . Note that the derivative of the function V. if Γ = ℜ n then the solution y (t ) = 0. m/s = vertical acceleration of the unsprung mass .
Figure 1. γ = 1 and ω = 1 . Subtracting (14) from (13) and defining: x x − ~ y= ~ & & x − x we will obtain the following system: (15) Finally. ~ ~ ~ && = −α ~ − ~ 3 − 2ζ ~ + γ cos ω t + u & x x x x (14) ~ is a control function which maintains the Duffing oscillator where u in the desired trajectory. B. respectively.An Optimal Linear Control Design for Nonlinear Systems Then: ~ Q = Q (t ) − G T ( y ) P (t ) − P (t )G ( y ) .6825 2. 1 0 and R = [1] . Vol. of the Braz. Q and R are matrix with constant elements. ζ = 0. Mechanical interpretation of Duffing oscillator. For α < 0 . Then the desired x regimen is described by the following equation: of the function (11) is given by V& = − yT Q y − u T R u . the derivative Let the desired trajectory be a function ~ (t ) . No. and. if Γ = ℜ n . of Mech. ζ > 0 is the viscous damping coefficient. 4 / 281 . as shown in Fig. Sci. Integrating the derivative of the Lyapunov function (11) given by ~ & V = − y T Q y − u T R u along the optimal trajectory. We remark that according to the optimal control theory of linear systems with quadratic functional (Anderson and Moor. that for positive definite matrices Q and R. where α is the stiffness parameter. Copyright © 2008 by ABCM OctoberDecember 2008. T −1 T (12) 0 0 y1 0 + ~ ) 2 − ( y + ~ ) ~ − ~ 2 0 y + 1 u 1 x x x − ( y1 + x 2 Let the desired trajectory be a periodic orbit: ~ = 2 cos t + sin t x (17) Linear Design for Duffing Oscilator We will apply the proposed method to the Duffing oscillator that is one of the paradigms of nonlinear dynamics. Soc. 1. If the time interval is infinite and A. shown in Fig. one can conclude the Theorem proof. matrix algebraic Riccati equation & y1 = y2 & y2 = −α y1 − 2ζ y2 − ( y1 + ~ )3 + ~ 3 + u x x ~ where u = U − u is feedback control. 1990) the solution of the nonlinear differential Riccati equation (8) is positive definite and symmetric matrix P > 0 for all R > 0 e Q ≥ 0 given. it is negative definite. The Duffing oscillator with the control law U (t ) is described by the following nonlinear differential equation: (18) && = −α x − x3 − 2ζ x + γ cos ω t + U (t ) & x (13) and the parameters α = −1. then one obtains Q= 0 1 3. For these parameters. XXX. and the controlled system (3) is locally asymptotically stable. & Eng. global asymptotic stability follows as a direct consequence of the radial unboundedness condition for the Lyapunov function (11) V ( y ) → ∞ as y → ∞ . By evaluating the Q . These are exactly the same desired trajectory and parameters as reported by Sinha et al. So the equation (3) in this case has the following form: (16) & y1 0 & = y2 − α 1 y1 + − 2ζ y 2 PA + A P − PBR B P + Q = 0 .1771 ~ LQR function in MATLAB®. ~ Note. the positive defined matrix P is the solution of the nonlinear. Then. we obtain T J min = y0 P (0) y0 .4142 2. the Duffing oscillator without control can be interpreted as a model of a periodically forced steel beam which is deflected toward the two magnets. the system (14) without control possesses a chaotic attractor. Phase portrait of uncontrolled Duffing oscillator. the function (11) is Lyapunov function.4142 P= by solving the Riccati equation (12) using the 2. γ and ω are the amplitude and frequency of the external input. one has: Choosing J.125. Figure 2. (2000). If the function ~ (t ) is a solution of equation x ~ (13) without the control term then u = 0 . 2.
using linear parameter varying control. The most part of these research contributions are based on linear timeinvariant suspension models for control design. i.1771 y2 . Thompson (1976) was the first to explore the use of optimal control techniques to design an active law. l The bs coefficient affects the damping force linearly. bs are the tyre stiffness. Phase portrait of controlled Duffing oscillator. respectively.632825 positive definite. OcotoberDecember 2008 ABCM . The real physical systems always include nonlinear components which must be taken into consideration.632825 Finally. x x x x x Note that the function ϕ ( y1. admitting 3 ≥ ~ ≥ 1 ... 4. which corresponds to 1/4 of the body mass. The equations of motion of the vehicle active suspension system are (Gaspar et al. Typical quartercar suspension system. bs and bsy . and the unsprung mass mu represents the wheel at one corner. (2003). which is caused by road irregularities. ~ ) ≥ ~ 2 and.Marat Rafikov et al x 0 −ϕ ( y1. dampings and springs. was considered by Gaspar et al. Figure 3. A typical quartercar suspension system is shown in Fig. respectively. The parameters kt . Parts of the l nonlinear suspension stiffness k s are a linear coefficient k s and a nl nonlinear one k s . (19) In this figure. XXX. l nl l & & & & mu &&w = −ks (zw − zc ) − ks ( zw − zc )3 − bs ( z w − zc ) + bsy z w − zc − z (20) − kt ( zw − zr ) + u l nl Here. To resolve the active suspension problem for nonlinear system in this work we will apply the linear feedback control design above considered. have nonlinear properties. one can evaluate Q as: x 4 ϕ min = 3~ x 4 2 ~ 3 2 p Q ≥ Q + ~ 2 12 x 4 p22 p22 3 2 p12 ≥Q+ 0 4 p22 p22 = 0 Figure 4. and the damping rate of the suspension. The disturbance w is caused by road irregularities. the dynamic characteristics of suspension components. the suspension stiffness. x p22 1 x 2 p ) 12 p22 p22 0 where ϕ ( y1. 4. k s . zc and z w denote the vertical displacement of the sprung mass and the unsprung mass. Linear Design for Nonlinear Automotive Active Suspension System Consider a quarter vehicle model that captures the fundamental features of the suspension design of the vehicle. 3. 4. Several methods to solve the active suspension problem have been proposed. ~ ) 0 = Q + ϕ ( y1. 1 1. Considering that the disturbance w. parts of the nonlinear suspension damping bs are bs .6213 1. and bsy describes the symmetric behavior of the characteristics. 2003): l nl l & & & & ms &&c = k s (zw − zc ) + k s ( zw − zc )3 + bs ( zw − zc ) − bsy zw − zc − u z The phase portrait behavior of the controlled Duffing oscillator is presented in Fig. Active suspension design. have a constant value. the desired trajectory is: 282/ Vol. then x x 2 ~ 3 x x ϕ ( y1. ~ ) = ( y1 + ~ ) 2 + ( y1 + ~ ) ~ + ~ 2 .e. ~ ) p11 ~ Q = Q − GT ( y ) P − PG ( y ) = Q − 0 0 p12 p12 − p22 p − 11 p12 0 0 p12 ~ − ϕ ( y . ~ ) has its minimum value x 3 in y1 = − ~ because ϕ ′ = 2 y1 + 3~1 ϕ ′′ = 2 > 0 . we can conclude that the optimal function u has the following form: u = −2. ms represents the sprung mass. The suspension force generated by the hydraulic actuator is inherently nonlinear. No.4142 y1 − 2. The control signal u is generated by the actuator.
2583 P= − 584. Comparison of the time traces of the variable of the controlled system (23) and the same system without control.5918 0. ~ = 0.0896 10. Finally. Comparison of the time traces of the variable of the controlled system (23) and the same system without control. J. & Eng.5054 y1 + 2892.7889 y4 (25) (24) For numerical simulations the disturbances of the initial conditions y1 = −0.5 0 0 100 0 0 100 0 0 . l k s = 23500 N / m . Copyright © 2008 by ABCM OctoberDecember 2008. 5 and 6 show y1 and y3 variations without and with control (25).. ms Figure 5.034 2. ~ = w. (12).414 A= .4481 12. l bs nl k s = 2350000 N / m . kt = 190000 N / m .7352 y3 − 2457.0896 −584.B= 1 0 l 1 b − s mu mu 12. s g ( y) = 0 nl by ks − ( y 3 − y1 ) 3 + s y 4 − y 2 m mu u Considering: 0 g g ( y ) = G ( y ) y = 21 0 g 41 where: 0 g 22 0 g 42 0 g 23 0 g 43 0 g 24 y 0 g 44 g 21 = − g 23 = − g 41 = g 43 = − g 22 = − g 24 = − g 42 = g 44 = δ ( y2 .1348 6411.414 81. For these parameters the matrix A has the following form: Figure 6.5 − 5337.0187 1. 2003): ms = 290 kg . respectively. No. XXX.00001] . Sci.5828 687. Vol.05 m and y3 = −0. bsy δ ( y 2 . ~ = 0 & & zc zc zw zw (21) Noting that the desired trajectory is a solution of the system (20) without control and defining: 1 0 0 0 − 81. 0 0 0 1 17. 4 / 283 .1348 0.035 − 2. y4 ) = k snl ( y3 − y1 ) 2 .1816 − 2. Fig.5918 1.5 587.05 m were considered. mu = 40 kg . of the Braz. y 4 ). ms 0 if y4 − y2 = 0 sign ( y4 − y2 ) if y4 − y2 ≠ 0 we present the equation (23) in form (3).2601 1.An Optimal Linear Control Design for Nonlinear Systems ~ = w. 0 nl y b ks ( y 3 − y1 ) 3 − s y 4 − y 2 m ms .2583 − 2. we can conclude that the optimal function u has the following form: u = 211.5828 by solving the Riccati Eq.5 − 17. of Mech.2601 − 110. then one obtains: where: 0 l ks − ms A= 0 kl s mu 1 bl − s ms 0 l bs mu 0 l ks ms 0 l ( k s + kt ) − mu 0 0 l bs 1 − m ms s . Q= 0 0 100 0 0 0 100 0 & y = zc − w zc [ & zw − w zw ] T (22) we will obtain the following equation in form of (1): & y = A y + g ( y ) + Bu Choosing (23) R = [0. Soc.2626 − 110. The proposed linear feedback design procedure has been applied to the quarter car suspension with the following values of the parameters (Gaspar et al.7037 y2 − 31494. = 700 N / m / s e bsy = 400 N / m / s .
Int. 284/ Vol. Krasovskii. Mech. Chellaboina. NY. Math. 363372. “On an optimal control design for Rössler system”.C. S. 1990.A. Hemisphere Publ.. R. “Applied optimal control: Optimization. Corp. Bryson. V. I.E. and Control”. J. JR. Systems Control Lett. 2. 1993. Automation and Remote Control. pp.. 33. Bernstein. A. Appl. Time trace of the positive function. “Optimal Control and Viscosity Solutions of HamiltonJacobiBellman Equations”. Estimation. No.L. pp. References Anderson. Washington D. 206221. 1998. Vehicle System Dynamics. “Active suspension design using linear parameter varying control”. C.. pp. 1959. vol.. D.O.. J. pp.. 2000.. Sazaszi. and Bertram.. The numerical simulations for the Duffing oscillator and the nonlinear automotive active suspension system are provided to show the effectiveness of this method. XXX. J. 211229. OcotoberDecember 2008 ABCM .. and Ravindra. M. 165178. N. Gaspar. Letov. “Control system analysis and design via the second method of Lyapunov”.T. Vol.M.B. P. Y. Int. Birkhäuser. and Fausz. A 333. ~ Fig. pp. Bardi. 22. A. 1975. Acknowledgements The work was partially supported by the Brazilian Government agencies CNPq and FAPESP. 7 shows the positive function L(t ) = yT Q y . “Nonquadratic cost and nonlinear feedback control”. M.. Biff. 1976.E. Finally we note that the proposed method can be applied to a large class of nonlinear and chaotic systems. S. and Bokor. “A general approach in the design of active controllers for nonlinear systems exhibiting chaos”.S. 10 (1). G. Chaos. 23. pp. No. Sinha. Rafikov. Figure 7. 3. B. pp. and Moor. calculated in optimal trajectory. J. A. “Robust nonlinear feedback control for uncertain linear systems with nonquadratic performance criteria”. J. pp. 187–203.. Haddad. M. Lett. and Ho... “An active suspension with optimal linear state feedback”.. 1. 2003. J. “Optimal Control: Linear Quadratic Methods”. “The analytical design of control systems”.. Boston. 5. and CapuzzoDolcetta. “On the theory of optimum control”. 1997. vol.Marat Rafikov et al ~ It is difficult to analyze the matrix Q analytically in this case. I..E.. The formulated Theorem expresses explicitly the form of minimized functional and gives the sufficient conditions that allow using the linear feedback control for nonlinear system. N. 4.M. W.C. 2004. 82. Henrichs. Thompson. PrenticeHall. and Balthazar. International Journal of Vehicle Autonomous Systems. 241245. Kalman. Journal of Basis Engineering. Phys.. pp. Vol. 1960. 371393..D. J. Asymptotic stability of the closedloop nonlinear system is guaranteed by means of a Lyapunov function which can clearly be seen to be the solution of the HamiltonJacobiBellman equation thus guaranteeing both stability and optimality. Robust Nonlinear Contr. Conclusions The optimal linear feedback control strategies for nonlinear systems were proposed. 327338. B.. 899919. J. 1961..