4 22 11
BULLETIN OF KYOTO INSTITUTE OF TECHNOLOGY4Nov 2010
On a Synthesis of Multivariable PID Controllers
for a Class of Nonlinear Systems
Nagato O HSE and Yoshitaka M ATSUDA
Department of Mechanical and System Engineering
Kyoto Institute of Technology
Matsugasaki,
2010 Sakyo-ku, Kyoto
3 30 2010 7 606-8585
22
ohse@kit.jp
Abstract
This paper is concerned with a synthesis of multivariable proportional-integral-
derivative (PID) controllers for a class of nonlinear systems. First, mathematical
model of plant with sector-bounded nonlinearity and multivariable PID controller is
given. Secondly, the closed-loop system is described by a descriptor form. A suf-
cient condition for checking the -performance of descriptor systems with sector-
bounded nonlinearity is introduced. Based on the condition, the problem of synthe-
sizing multivariable PID controllers is formulated as a bilinear matrix inequality one.
The problem is solved by an iterative linear matrix inequality algorithm to synthesize
a PID controller so as to expand the region of sector and minimize -gain. Finally,
the synthesis method is applied to the synthesis of a PID controller for a ball-on-
wheel system to verify the effectiveness.
Key Words: multivariable PID controller; sector-bounded nonlinearity; descriptor
system; matrix inequality; -performacne.
1
33
1. Introduction
It goes without saying that most of dynamical behavior of industrial apparatus is de-
scribed by nonlinear mathematical models. Over the past few decades, the concept of
sector-bounded nonlinearity has been used in order to analize and synthesize a class of
nonlinear systems1) . In recent years, methods of analyzing and synthesizing the systems
with sector-bounded nonlinearity via linear matrix inequality (LMI) have been vigorously
studied, and applied to various nonlinear systems such as systems with input saturation2) .
The convex optimization problem subject to LMIs can be solved efciently by the interior-
point method3) .
The proportional-integral-derivative (PID) control scheme has been widely used in
various industrial control systems, and most of conventional PID controller synthesis
methods are based on the relatively simple representation of plant dynamics such as the
characterization by a single-input/single-output model. These are, however, not necessar-
ily sufcient for plants with multi-input/multi-output.
Recently, multivariable PID controller design method in time domain4),5),6),7) and a pre-
assigned structure-constrained PID controller design method8) in frequency domain have
been proposed. However, these methods have dealt with only design for linear plant.
In this paper, a method of synthesizing multivariable PID controllers for systems with
sector-bounded nonlinearity so as to minimize the -gain and expand the region of sector
is presented. Then, this paper extends the analysis condition proposed by F. Wu and B.
Lu2) .
The paper is organized as follows. In Section 2, mathematical model of nonlinear plant
with multivariable PID controller is given. The analysis condition is obtained in order to
synthesize multivariable PID controllers for systems with sector-bounded nonlinearity in
Section 3. Section 4 is devoted to the formulation of our problem as an optimization
one, which is subject to a bilinear matrix inequality (BMI). In Section 5, multivariable
PID controller is rst obtained for a linearized plant, and then, the controller for nonlinear
2
34
system is determined so as to expand the region of the sector and minimize the -gain by
an iterative LMI algorithm. In Section 6, the synthesis method is utilized for synthesizing
a PID controller for a ball-on-wheel system to illustrate the effectiveness.
The principal symbols used in this paper are listed below:
R set of all real #-vectors
R
set of all real " #-matrices
# #-identity matrix
transpose of a vector or matrix
is positive (semi-)denite
is positive (semi-)denite
diagonal matrix whose elements are
" trace of matrix
norm of matrix dened by !# 2 !
- norm of vector -' dened by -' -' '
2. Problem Statement
Consider a time-invariant nonlinear plant described by
+ ' + ' (' )' *'
)$'
+ '
(1)
,'
+ '
-'
+ ' (' *'
and
)' 5$
) ' (2)
where + ' R , (' R , ,' R and -' R
%
# are the plant state,
control input, measured output and controlled output, respectively. *' R is the
3
35
exogenous disturbance such that * , and , , , ,
,
,
,
and are constant matrices of appropriate dimensions. The matrices and
are
assumed to be of full column rank and full row rank, respectively, i.e., and
%.
The vector-valued function 5 : R R is a nonlinear function which is assumed to
belong to
) )
)$ )
)$
% 5 ) 5$ (3)
where )$ R ,
,
, and
.
) ) )
)$ )
The relation between the nonlinear function 5 and the sector $
)$
for given
,
and is shown in Fig. 1. Throughout this paper, the slope
matrix
is xed for simplicity.
Fig. 1. Relation between nonlinear function and sector.
The control input (' is generated by the following multivariable PID controller :
(' ,' ,3 3 ,'
(4)
where , and are the PID controller gain matrices of appropriate dimensions.
4
36
In this paper, we develop a method of synthesizing PID controller (4) so as to minimize
/ and "
, where / is an index in the performance condition -
/* , and
the minimization of "
stands for the enlargement of the sector.
3. Analysis of a Class of Descriptor Systems with Sector-bounded
Nonlinearities
In order to synthesize PID controller (4) for systems with sector-bounded nonlinearity
in %, we introduce the following proposition obtained by extending the result by F. Wu
and B. Lu2) .
Proposition: Assume that a scalar / , diagonal matrices
and
and coef-
cient matrices
and in the system
+'
+' )' *'
)$'
+' (5)
-'
+' *'
are given. If there exist a nonsingular matrix and positive-denite diagonal matrix
such that
(6)
(7)
/
where
, then, for all 5 %, the system
(5) with ) 5$
) satises -
/* provided that + .
5
37
Proof: By the Schur complement formula9) , we see that (7) is equivalent to
(8)
/
Multiplying (8) by + ) * on the left and by its transpose on the right and using (6),
we have
+ + )
)$ )
)$ / * * - -
(9)
'
For all 5 %, (9) implies
+ + / * * - -
(10)
'
Integrating both sides of (10) from ' to ' ', we have
- 3 -3 3 / * 3 *3 3
(11)
+ ' +' + +
If + , then from (11) we have
- 3 -3 3
/ * 3 *3 3
and therefore -
/* . (Q.E.D.)
4. Problem Formulation via Matrix Inequality
In this section, applying Proposition , we formulate our synthesis problem by matrix
inequality.
Consider a controller with $-dimensional vector process + ' R :
+ ' ,' +
(12)
(' ,' + ' ,'
6
38
where R is a preassigned constant matrix which is chosen appropriately ac-
cording to the structure of the PID controller to be synthesized. Setting ,
and yields the PID controller gains in (4). Furthermore, in-
troducing a vector +' dened by
+' + ' + ' + ' (13)
we describe the system which consists of (1) and (12) in the descriptor form:
+'
+' )' *'
)$'
+' (14)
-'
+' *'
where
Proposition proves that if there exist a nonsingular matrix , a positive-denite
diagonal matrix , a diagonal matrix
(
), a matrix and a positive
7
39
scalar / satisfying
(15)
' /
'
'
'
'
'
' '
(16)
/
where
'
'
'
'
then there exists a controller (12) (or (4)) which achieves the specication that the closed-
loop system satises the condition -
/* for + .
Thus, our problem of determining the controller gain matrix is formulated as fol-
lows:
nd , , ,
and / so as to
minimize / and "
subject to (15) and (16).
This is a BMI problem, and it is difcult to convert it into a sort of convex optimization
problems such as LMI problem.
8
40
5. Synthesis of PID Controller
In this section, noting that xing either or
in (16) yields LMI in
the other pair of matrices, we propose an iterative LMI algorithm to solve our problem.
The algorithm does not have a proof for global convergence.
In order to solve the problem of nding , and / so as to minimize / subject to
(15) and (16) with xed and
, we need to calculate and
in advance.
First, we obtain by solving the problem of synthesizing a PID controller for a
)
linearized plant, where the linearized plant is obtained by the approximation 5$ )$
(or
). Then, the resultant linearized closed-loop system is described as
follows:
+'
$
+' *'
(17)
-'
+' *'
where $ and
.
The controller gain in the linearized system (17) is denoted by in the sequel.
Given a matrix and a positive scalar /$ , the system (17) is admissible and /$
holds, where & is the transfer function matrix from * to -, if and only if there exists
a nonsingular matrix such that
(18)
& /$
$
$
$/ (19)
hold10) .
9
41
Consequently, the problem of synthesizing PID controller under
criterion, i.e.,
/$ is formulated as that of nding matrices , and a positive scalar /$
satisfying the conditions (18) and (19). Note that (19) is BMI in and .
Noting that xing either or in (19) yields the LMI condition in the other matrix,
we minimize /$ locally by the iterative LMI algorithm for and . Here, since the
algorithm requires an initial value of PID controller gain, rst, the initial value
is set
by such that the matrix $
is stable.
Procedure for Obtaining
Step 1: Set
, nd and /$ so as to minimize /$ subject to
and & /$
in (18) and (19) (
and /$
/$ ).
Step 2: With the obtained and /$ , nd subject to & / in (19).
Step 3: Again, with in Step 2, nd and /$ so as to minimize /$ subject to
and & /$ in (18) and (19) (
and /$
/$ ).
/ /$ , where is the prescribed tolerance, stop. Otherwise, let
Step 4: If $
and go back to Step 2.
Using obtained by the above procedure, we state an iterative LMI algorithm for
synthesis of PID controller (12) for a nonlinear plant (1) with (2) as follows:
An Iterative LMI Algorithm
Step 1: Choose a scalar positive constant and matrices
and
, and then set
and
.
Step 2: Find and / so as to minimize / subject to
and
'( , , /; ,
) in (15) and (16). Set
,
, /
/.
10
42
Step 3: Find and
so as to minimize tr
subject to ' /
in (16).
Step 4: If tr
0, then stop the algorithm. Otherwise, set
,
,
and go to Step 2.
6. A Numerical Example
In this section, we apply the approach proposed in this paper to the problem12) of
synthesizing a dynamic controller to balance a ball on the periphery of a wheel as shown
in Fig. 2, where 1 is the angle between the center of the ball and the vertical axis, 1
is the wheel angular position, and 3 and * are the control torque exerted on the wheel
and the disturbance added to the wheel, respectively. is the inertia of the wheel, "
is
the mass of the ball, and %
, % are the radii of the ball and the wheel, respectively. is
the gravitational acceleration. The physical parameters are listed in Table 1, in which the
numerical values are the same as in M. Ho and J. Lu12) .
Table 1. Physical parameters of ball-on-wheel system.
%
radius of the ball [m]
% radius of the wheel [m]
inertia of the wheel
[kgm ]
"
mass of the ball
[kg]
motor armature resistance
[(]
motor constant [Nm/A]
gravitational acceleration
[N/m ]
In this paper, we assume that the coefcient of friction is sufciently large and there-
fore the ball rolls on the wheel without slipping. Then, the equations of motion of the
11
43
Fig. 2. Ball-on-wheel system.
system can be written as follows:
*
"
% % %
1 ' "
% 1* ' 3 ' *'
(20)
%
% 1* ' % 1* ' ! 1 '
A voltage signal (' is generated according to the desired control law and it is supplied
to an amplier which drives a permanent magnet DC motor to control the wheel. The
relation between the control torque 3 ' and the control voltage (' is given by
3 ' ('
1 ' (21)
where is the motor armature resistance and
is the motor constant.
Here, note that to maintain the ball on the wheel the centripetal force must be larger
than the centrifugal force:
! 1 %
% 1 (22)
We dene the state variables as
+ ' 1 ' 1 ' 1 ' 1 '
12
44
and set
,' -' 1 ' 1 '
where the sensor gains are and . Then, from (20) and
(21), we see that the ball-on-wheel system is described by
) 4$
) ! )$ (23)
and (1) with the following coefcient matrices:
)
) )
) ) )
) )
. "
% %
%
) %
.
) "
% . ) %
.
) %
%
. ) "
% %
% .
) %
%
. ) % .
) %
% .
The relation between the function 4$
) and the sector is illustrated in Fig. 3. This gure
shows that $
)
implies 4 %.
The approximation 4$ )$ (or ! 1
) 1 ) yields a linearized closed-loop system
described by (17). Then, setting
13
45
Fig. 3. Relation between 4 and sector.
and using the procedure mentioned in the previous section, after iterations in the itera-
tive LMI algorithm, we obtain
/$
We choose a scalar 0 and an initial value of
as follows:
0
After
iterations in the iterative LMI algorithm given in Section 5, we obtain
/
Figures 4 and 5 show the value of performance index / and the slope
for each step,
respectively.
14
46
-gain
Iteration Number
Fig. 4. -gain /.
Slope
Iteration Number
Fig. 5. The slope
.
15
47
Disturbance [Nm]
Time [s]
Fig. 6. Disturbance *.
Disturbance *' added to the wheel is shown in Fig. 6. Behavior of 1 ', 1 ' and
3 ' is depicted in Figs. 7 to 9. We see from Fig. 10 that the constraint (22) is satised.
7. Conclusion
A method of synthesizing multivariable PID controllers for systems with sector-bounded
nonlinearity has been proposed. In order to expand the region of the sector and minimize
the -gain, an iterative LMI algorithm has been presented. Although the algorithm does
not have a proof for global convergence, it has a practical use. The authors have applied
the algorithm to design of the PID controller of a ball-on-wheel system, and conrmed
the effectiveness numerically.
Acknowledgment
The authors would like to appreciate Mr. Masayuki Maeda, graduate student of Kyoto
Institrute of Technology for his assistance in simulation work.
16
48
Angle
[degree]
Time [s]
Fig. 7. The angle 1 .
Angle
[degree]
Time [s]
Fig. 8. The angle 1 .
17
49
Control torque [Nm]
Time [s]
Fig. 9. The torque 3 .
Time [s]
Fig. 10. The constraint ! 1 %
% 1 in (22).
18
50
References
1) H.K. Khalil, Nonlinear Systems, Prentice Hall, NJ, 1996.
2) F. Wu and B. Lu, Anti-windup Control Design for Exponentially Unstable LTI Sys-
tems with Actuator Saturation, Systems & Control Letters, 52, 305-322, 2004.
3) P. Gahinet, A. Nemirovski, A.J. Laub and M. Chilali, LMI Control ToolBox, The
MathWorks, 1995.
4) N. Ohse, Y. Matsuda, and K. Tanaka, On a Synthesis of Multivariable PID Controllers
for a Class of Linear Systems, Proceedings of the 32nd Annual Conference of the
IEEE Industrial Electronics Society, Paris, France, 549-554, November, 2006.
5) F. Zheng, Q.-G. Wang, and T, H. Lee, On the Design of Multivariable PID Controller
via LMI Approach, Automatica, 38, 517-526, 2002.
6) C. Lin, Q.-G. Wang, and T. H. Lee, An improvement on multivariable PID controller
design via iterative LMI approach, Automatica, 40, 519525, 2004.
7) M. Mattei, Robust multivariable PID control for linear parameter varying systems,
Automatica, 37, 19972003, 2001.
8) M. Saeki, Fixed structure PID controller design for standard
control problem,
Automatica, 42, 93-100, 2006.
9) R. E. Skelton, T. Iwasaki and K. M. Grigoriadis, A Unied Algebraic Approach to
Linear Control Design, Taylor & Francis Ltd., London, 1998.
10) I. Masubuchi, Y. Kamitane, A. Ohara, and N. Suda,
Control for Descriptor Sys-
tems: A Matrix Inequalities Approach, Automatica, 33-4, 669673, 1997.
11) Y. Matsuda and N. Ohse, Synthesis of Dynamic Controllers for a Class of Nonlinear
Systems: An Application to a Ball-on-Wheel System, Proceedings of the 2006 IEEE
19
51
International Conference on Control Applications, Munich, Germany, 1061-1066,
October, 2006.
12) M. Ho and J. Lu,
PID Controller Design for Lure Systems and its Application
to a Ball and Wheel Apparatus, International Journal of Control, 78-1, 53-64, 2005.
13) K. Zhou, J. Doyle and K. Glover, Robust and Optimal Control, Prentice Hall, NJ,
1996.
20
52
PID
(PID)
PID
-
PID
(BMI)
- PID
ball-on-wheel PID
: PID ; ; ;
; -.
1
53
54
