Вы находитесь на странице: 1из 13

Asian Journal of Control, Vol. 8, No. 2, pp.

161-173, June 2006

161

Brief Paper

STATE-SPACE DIGITAL PID CONTROLLER DESIGN FOR


MULTIVARIABLE ANALOG SYSTEMS WITH
MULTIPLE TIME DELAYS
Jennifer M. Madsen, Leang-San Shieh, and Shu-Mei Guo
ABSTRACT
This paper presents a discrete-time state-space methodology for optimal
design of digital PID controllers for multivariable analog systems with multiple time delays. The multiple time-delayed multivariable analog systems
are formulated in a state-space generic form so that the exact discrete-time
state-space model can be constructed. Then, the optimal digital PID controller is designed via a state-feedback and state-feedforward LQR approach.
The developed PID controller can be applied to a general time-delayed multivariable analog system represented by a semi-proper or strictly proper
transfer function matrix. Illustrative examples are given to compare the performance of the proposed approach with alternative techniques.
KeyWords: State-space PID controller, digital design, LQR, multivariable
system, multiple time delay.

I. INTRODUCTION
Control of time-delay systems is an area of current research interest [1-9] due to the theoretical challenge, as
well as the wide range of practical applications for such
controllers. Delays are encountered in most real-world engineering systems, such as the turbojet engine, microwave
oscillator, transmission lines, and chemical industrial processes. Depending on the severity of the time delay, this
may be a main source of oscillation and possibly instability
in the system. An added complexity can be multiple time
delays in a coupled multi-input multi-output (MIMO) system.
Recently, a number of methods for designing Proportional-Integral (PI) and Proportional-Integral-Derivative
(PID) controllers for MIMO systems have been proposed
Manuscript received March 31, 2005; revised December 6,
2006; accepted February 22, 2006.
J.M. Madsen and L.-S. Shieh are with the Department of
Electrical and Computer Engineering, University of Houston,
Houston, Texas, U.S.A. (e-mail: lshieh@uh.edu).
S.-M. Guo is with the Department of Computer Science and
Information Engineering, National Cheng Kung University,
Tainan, Taiwan.
This work was supported by the U.S. Army Research Office
(DAAD 19-02-1-0321), NASA-JSC (NNJ04HF32G), and the
National Science Council of China (NSC 94-2213-E-006-068).

[8-13]. These methods address a number of the complexities such as loop interaction, however the selection of PID
controller design parameters is still commonly accomplished through empirical rules instead of analytical development. Also, few of the methods currently under investigation address the discretization of MIMO analog PID
controllers.
Continuing developments in computer technology including high-speed microprocessors and interface hardware
have made digital controllers the norm for controlling robots, motors, spacecraft, and many other systems. These
digital components are preferred over their analog counterparts because of their reduced cost, and increased flexibility, reliability and compactness, especially for the size
and complexity of a MIMO controller. There are currently
three approaches for designing a digital controller to be
applied to an analog plant. The first method is called the
direct sampled-data approach [14,15]. This process directly
develops a digital controller for the given analog plant.
Methods for this design process are not well-established
and are still being developed. Another possibility for digital
control design is called digital redesign [16-18]. In this
method an analog controller is developed to meet the stability and performance requirements for the system. Then
the analog controller is transformed into a digital controller
through any number of digital redesign schemes.
In this paper, for control of analog systems with multiple time delays, we apply the direct digital design method

162

Asain Journal of Control, Vol. 8, No. 2, June 2006

[18,19]. In this case, the analog system is first transformed


into an equivalent digital model. Next, a digital controller
can be designed for the digitized plant. For implementation,
the digital controller is then applied to the original analog
system. This method has a number of advantages, including
the fact that most parameter identification techniques are
developed in the discrete-time domain, and thus a MIMO
digital controller can be developed directly based on the
identified discrete-time plant model. Also, direct digital
design is often less computationally intensive and more
straightforward to implement, allowing for reconfiguration
and re-tuning based on software changes. A possible
drawback to this design technique is that it only considers
the system behavior at discrete sampling instants and gives
no control of intersample behavior in the system. For a
short sampling period, the discretized model of the original
time-delayed MIMO analog plant may become a high dimensional one, and thus increase the computation load. For
a long sampling period, the intersample behavior of the
hybrid system should be post- checked for possible divergences. However, careful selection of the sampling rate can
mitigate these risks and provide the desired system performance.
The method of this paper uses a state-space PID
controller design technique because of some important
advantages over the frequency-domain PID controller design schemes. One important consideration is that most
frequency-domain PID design methods have been developed for single-input single-output (SISO) processes which
can be approximated by the first-order plus dead time
(FOPDT) or second-order plus dead time (SOPDT) models.
The state-space analog PID controller design for multivariable system described by a strictly proper transfer function matrix with and without time delays was recently proposed by Zhang et al. [8-10], in which the small time delay
terms were first approximated by Pade approximation
method. Then, based on the obtained overall approximated
analog model, an analog PID controller was designed via
linear quadratic regulator (LQR) approach. When the
computational time delay is sufficiently small, a prediction-based digital redesign method [8,9] was utilized to
convert the analog PID controller to a digital PID one. In
addition, an observer of the approximated analog model
was introduced to reject the output disturbance. However,
when the delay time is sufficient large, the Pade approximation method leads to an overall high-order
non-minimum phase approximate model. As a result, a
de-tuning process is still required [20-22] when the digitally redesigned PID controller is applied to the original
multiple time-delayed multivariable system. Furthermore,
the prediction-based digital redesign method is not effective to discretize a long time-delayed analog controller.
In this paper, the proposed state-space digital PID
controller design for a multivariable system described by a
semi-proper or strictly proper transfer function matrix can
overcome the afore-mentioned disadvantages because it has

no specific requirements for a reduced-order model. Thus


the original MIMO analog system described by a semiproper or strictly proper transfer function matrix with multiple time delays is converted into an exact discrete-time
model via the proposed state-space generic structure. This
method can also be applied to plants regardless of system
stability, number of inputs and outputs, degree of system,
minimum-phase properties, length of dead time, and system coupling properties. In this paper it will be shown that
the MIMO optimal PID parameters can be determined by
tuning the weighting matrices in the LQR performance
indices, and that the closed-loop stability is guaranteed
during this tuning process. Furthermore, a method for transient improvement and disturbance rejection is introduced.
The rest of the paper is organized as follows. Section 2
details the proposed method for developing the digital
state-space representation of a linear multivariable system
with multiple time delays. A digital state-space PID controller is then designed in Section 3 to be used in series
with the plant. A state-feedback/feedforward LQR design
method is applied for tuning of the MIMO control parameters, and to guarantee closed-loop system stability. Section
4 presents two numerical examples highlighting the performance of the proposed approach, and Section 5 provides
concluding remarks.

II. STATE-SPACE REPRESENTATION OF A


MULTIPLE TIME-DELAYED SYSTEM
We assume that a transfer function matrix, T(s), exists
for a linear multivariable system with multiple time delays,
output Y(s), and input U(s), such that

Y (s) = T ( s) U ( s) .

(1)

For simplicity we can show that for a first-order model of a


system with m inputs and p outputs, this strictly proper
transfer function matrix may take the form

b11e L11s

s a11
b e L21s
21
T ( s ) = s a21

bp1e Lp1s

s a p1

b12 e L12 s
s a12
b22 e L22 s
s a22
#
"

b1m e L1m s

s a1m
b e L2 m s

" 2m
s a2 m ,

%
#

L s
bpm e pm

"
s a pm
"

(2)

where aij, bij, Lij R, and Lij 0 for i = 1, 2, , p, and j = 1,


2, , m. Here, the exponential terms represent the time
delay in each subsystem.
For each element, Tij(s), of the transfer function matrix
we can write

Tij ( s ) : sxij ( s ) = aij xij ( s ) + bij e

Lij s

u j (s) ,

(3)

J.M. Madsen et al.: State-Space Digital PID Controller Design for Multivariable Analog Systems

and the system output is given by

yi ( s ) =

quirement, the determinant of the controllability matrix, P,


must not be zero as shown by the following:

xij ( s) .

(4)

j =1

For an example system with two inputs and two outputs, we introduce the state-space generic form [1],

x11 (t ) a11


x12 (t ) = 0
x (t ) 0
21
x22 (t ) 0

a12

a21

0 x11 (t )

0 x12 (t )
0 x21 (t )

a22 x22 (t )

b11u1 (t L11 )

b12 u2 (t L12 )
+
,
b u (t L )
21 1
21

b22 u2 (t L22 )

a12

a21

b11e L11s

0
+
L21s
b21e
0

det P = det [ B

AB ]

b11 0
0 b
12
= det
b21 0
0 b
22

a11b11
0
a21b21
0

0
a12 b12

0
a22 b22

= b11 b12 b21 b22 (a21 a11 ) (a12 a22 ) .

(5)

(11)

y (t ) = C x(t ) ,

where x(t) Rn, u(t Td) Rm, y(t) Rp, and A, B, and C
are constant matrices of appropriate dimensions.
The zero-order hold sampling of the system (11)
yields a discrete-time model of this system with a sampling
period, T:

0 x11 (t )

0 x12 (t )
0 x21 (t )

a22 x22 (t )

L12 s
b12 e
u1 (t )

.
0 u2 (t )
b22 e L22 s

k +

x(k + ) = e A x(k ) + k

(6)

(7)

In a condensed notation, these equations are

x (t ) = A x(t ) + B ( L) u (t ) ,

(8)

y (t ) = C x(t )

e A( k + ) B u ( ) d ,
(12)

The system output is written as

x11 (t )

y
(
t
)
1
1
0
0
1
x12 (t )
=
.

y2 (t ) 0 0 1 1 x21 (t )
x22 (t )

(10)

Therefore, our system is controllable if a11 a21, a12 a22,


and b11b12b21b22 0.
Again considering a MIMO system with m inputs, p
outputs, n states, and an input time delay, Td. The state
equations can be given by the following:

x (t ) = A x(t ) + B u (t d )

or an alternate mixed generic form,

x11 (t ) a11


x12 (t ) = 0
x (t ) 0
21
x22 (t ) 0

163

where u(t) is a piecewise-constant vector represented by u(t)


= u(kT) for kT t < (k + 1) T, k = 0, 1, 2, For our system, the time delay represented by Td can be given by Td =
(d 1) T + , where d > 0 is an integer and 0 < < T. Thus,
the first term represents the number of full sampling periods that the signal is delayed, and the second term represents the fraction of a sampling period in the delay. Another variable we will use in deriving the digital state equation is the ratio between and the sample time T, r = /T,
where 0 < r < 1.
Taking into account the input delay, the integral in
(12) may be divided into two parts and evaluated as follows
[2]:
kT +

x(kT + T ) = e AT x(kT ) + kT

e A( kT +T ) Bu (kT dT ) d

kT +T

When L = 0, the mixed generic input vector, B(L), is


denoted as B. For the diagonal system matrix A, and the
generic structure of B, the controllability matrix can be
expressed as
2

P = ( B, AB, A B, " , A

m p 1

B) .

(9)

Because of the special form of the A and B matrices, we


can determine that the 2 2 example system is controllable
if the rank of P = [B AB] is equal to 4. To satisfy this re-

+ kT + e A( kT +T ) Bu (kT (d 1)T ) d
= Gx(kT ) + H11u (kT dT ) + H10 u (kT (d 1)T ) ,
(13)
where G = e , H11 = [G G ] A B, H10 = [G I] A1B,
and G r = eA. It is noted that the matrix-valued function,
AT

(1)

[eXT I] X1 = [G I] X1, shall be represented as


when X is a singular matrix.

(1r)

T
i =1 i !

( XT )

i 1

Asain Journal of Control, Vol. 8, No. 2, June 2006

164

In order to express the system in traditional state-space


form [2], we define an augmented state vector for each
system consisting of the original state, x(kT), and the delayed input vectors, u(kTdT), u(kT(d1)T), , u(kTT).
Using this method, for a system with a time delay less than
one sampling period and d = 1, we can write the following:

x(kT + T ) G
u (kT ) = 0

H11 x(kT ) H10


u (kT ) ,
+
0 u (kT T ) I

x(kT )
y (kT ) = [C 0]
.
u (kT T )

(14)

H1e

y1 (kT )

y2 (kT ) = C
# 1e

yq (kT )

For longer time delays where d > 1. the state-space representation is given by

x (kT + T )
u (kT (d 1)T )

u (kT (d 2)T )

=
#

u (kT T )

u (kT )

G
0

0
#

0
0

H11
0
0
#
0
0

H10
I
0
#
0
0

"
"
"
%
"
"

0
0
0
#
0
0

0
0

0
#

I
0

x(kT )

u (kT dT )

u (kT (d 1)T )

u (kT 2T )
u (kT T )

0
0

0
+ u (kT ) ,
#

0
I

x(kT )

u (kT dT )

u (kT (d 1)T )
y (kT ) = [C 0 0 " 0 0]
. (15)
#

u (kT 2T )
u (kT T )

We can combine systems with different time delays into


one augmented system. In general, the state equations for q
systems block combined are given by

x1e (kT + T ) G1e


x2e (kT + T ) =


xqe (kT + T )

G2e

Gqe

x1e (kT )

x2e (kT )
#

xqe (kT )

H 2e

H qe

u1 (kT )

u2 (kT ) ,
#

uq (kT )

x1e (kT )

x2 e (kT )
.
" Cqe
#

xqe (kT )

C2e

(16)

III. DESIGN OF A DIGITAL PID


CONTROLLER USING LQR APPROACH
For control of our MIMO digitized plant, we can use a
PID controller in series with the plant and close the loop
with unity output feedback as shown in the block diagram
of Fig. 1.
Let us assume that the load disturbance, d(kT) = 0, and
the minimal realization of the plant with semi-proper
transfer function matrix, Tp(z), can be written in the form

x p (kT + T ) = G p x p (kT ) + H p u p (kT ) , x p (0) = 0 ,


y p (kT ) = C p x p (kT ) + D p u p (kT ) ,
n1

(17)

m1

p1

where xp(kT) R , up(kT) R , yp(kT) R , and Gp,


Hp, Cp, and Dp are constant matrices of appropriate dimensions. In general, for a real plant, the forward input matrix
Dp in (17) is a null matrix. The state-space realization of
the plant Tp(z) can be the form shown in (16). When the
load disturbance occurs at the output of the plant, and the
direct link from input to output of the plant is permissible,
the matrix Dp can be utilized as a forward constant controller to improve the transient response and to reject the constant disturbance during the initial phase of the response
following a change in the reference signal or disturbance.
We also assume that the state-space representation of
the PID controller, Tc(z), can be given by

xc (kT + T ) = Gc xc (kT ) + H c uc (kT ) , xc (0) = 0 ,


yc (kT ) = Cc xc (kT ) + Dc uc (kT ) ,

(18)

where xc(kT) R n2 , uc(kT) R p1 , yc(kT) R m1 , r(kT)


R p1 , and Gc, Hc, Cc, and Dc are constant matrices of appropriate dimensions. The assignment of the pre-designed
d(kT)
r(kT)

Er

uc (kT)

Tc(z)

PID Controller

up(kT)

Tp(z)
Plant

Fig. 1. Discrete-time cascaded closed-loop system.

yp(kT)

J.M. Madsen et al.: State-Space Digital PID Controller Design for Multivariable Analog Systems

u p (kT ) = u p (kT ) Dc uc (kT )

PID controller, Tc(z), is a design freedom. For simplicity,


Tc(z) can be chosen as a diagonal matrix, in which each
diagonal element is a strictly proper rational function with
the desired poles of the PID controller. Then, any available
minimal realization technique [19] can be applied to determine the parameter matrices Gc, Hc, and Cc of the state
equations in (18). The selection of the forward gain matrix
Dc is another design freedom. For simplicity, it can be
chosen as an identity matrix or the static decoupler for the
system.
As shown by the block diagram in Fig. 1, we can write
the input to the controller block as

= u p (kT ) Dc ( y p (kT ) + Er r (kT ))


= u p (kT ) Dc ( C p x p (kT ) D p u p (kT ) + Er r (kT ))
= u p (kT ) + Dc C p x p (kT ) + Dc D p u p (kT ) Dc Er r (kT ) .
(23)
Solving for up(kT) yields

u p (kT ) = K 0 u p (kT ) + K 0 Dc C p x p (kT )


K 0 Dc Er r (kT ) ,

= C p x p (kT ) D p u p (kT ) + Er r (kT ) , (19)


where Er is a constant matrix of appropriate dimension.
The augmented system equations of (17) and (18) can be
given by

ya (kT ) = y p (kT ) = Ca xa (kT ) + D p u p (kT ) ,

xa (kT + T ) = G a xa (kT ) + H a u p (kT ) + E a Er r (kT ) , (25)


where

G p + H p K 0 Dc C p
G a =
H c C p H c D p K 0 Dc C p

(20)

where

0
,
Gc

0
Ea = , Ca = C p
Hc

Hp
Ha =
,
H c Dp
0 .

u p (kT ) = K a xa (kT ) = K1 x p (kT ) K 2 xc (kT ) , (27)


where Ka = [K1 K2]. Using this result in (24) allows the
desired original plant input, up(kT), to be given by

u p (kT ) = K 0 K1 x p (kT ) K 0 K 2 xc (kT )

u p (kT ) = K1 x p (kT ) K 2 xc (kT ) Dc uc (kT ) , (22)

+ K 0 Dc C p x p (kT ) K 0 Dc Er r (kT )

where the control gains K1 and K2 are to be determined.


The control law up(kT) in (22), which involves another
input uc(kT), is not suitable for LQR design. To eliminate
the input uc(kT), it is possible to define a modified plant
input, u p (kT ) , for design as follows:

= K 1 x p (kT ) K 2 xc (kT ) + E r Er r (kT )


= K a xa (kT ) + E r Er r (kT ) ,

Dp

Dc
Er

uc(kT)

(26)

Now the modified plant input in (25) can be designed


and written in terms of the gains, Ka, in the following
manner:

(21)

Examining the detailed block diagram in Fig. 2, it can


be shown that for the system in (20),

r(kT)

0
,
Gc

H p K0
H p K 0 Dc
H a =
, E a =
.
H c D p K 0
H c + H c D p K 0 Dc

x (kT )
xa (kT ) = p
,
xc (kT )
Gp
Ga =
H cC p

(24)

where K0 = (Im1 Dc Dp)1. Note that when Dp = 0p1m1 or Dc


= 0m1p1, K0 = Im1.
Substituting (24) into the first equation in (20) gives
the augmented system state equation

uc (kT ) = y p (kT ) + Er r (kT )

xa (kT + T ) = Ga xa (kT ) + H a u p (kT ) + Ea Er r (kT ) ,

165

xc(kT+T)=Gcxc(kT)+Hcuc(kT)

x c (kT)

K2

(28)

d(kT)

Plant

- up(kT)
- -

xp (kT+T)=Gp xp (kT)+Hp up (kT)


yp (kT)=Cp xp (kT)

PID Controller

K1
Fig. 2. Discrete-time closed-loop system.

x p(kT)

y p(kT)

Asain Journal of Control, Vol. 8, No. 2, June 2006

166

where K1 = K 0 (K1 Dc C p ), K 2 = K 0 K 2 , K a = [ K1 K 2 ] ,
and E r = K 0 Dc . Finally, the augmented state equations
for the designed closed-loop system in (25) can be written
in the form

xa (kT + T ) = GCL xa (kT ) + E a Er r (kT ) ,

x(kT + T ) = G x(kT ) + H u (kT ) ,


y (kT ) = C x (kT ) + D u (kT ) ,

ya (kT ) = y p (kT ) = Ca xa (kT ) + D p u p (kT )


= (Ca D p K a ) xa (kT ) + D p E r Er r (kT ) ,

(29)

= G a H a K a

G p + H p K 0 ( Dc C p K1 )
H p K0 K2

=
.
(
)

H
C
D
K
D
C
D
K
K
G
p 0 c p
p 0 1
c + H c Dp K0 K2
c p
(30)

has the transfer function given by

u p (kT ) = K a xa (kT ) ,

(31)

to guarantee that the closed loop poles of the system lie


inside the unit circle. The following quadratic performance
index, J, is minimized for the system:

J = xaT (kT )Qxa (kT ) + u pT (kT ) Ru p (kT ) , (32)


k =0

where the weighting matrices Q = QT 0 and R = RT > 0.


Thus, the optimal control gain, Ka, can be determined as

K a = R + H aT PH a

H aT PG a ,

(33)

using P = PT > 0, the solution to the following discrete-time


Riccati equation [18,19]:
T

P = G aT PG a + Q H aT PG a R + H aT PH a H aT PG a .

(34)
In order to ensure that at the steady-state the output
yp(kT) tracks the input r(kT) in (29), the forward gain Er for
the control law in (28) can be calculated. At the steadystate, xa(kT+T) = xa(kT), and the first equation in (29) can
be solved as

xa (kT ) = ( I ( n1 + n2 ) G a + H a K a ) 1 E a Er r (kT ) . (35)


Using this result in the output equation of (28) along with
the requirement that ya(kT) = r(kT) at the steady-state, Er is
given by
1

Er = D p E r + (Ca D p K a )( I ( n1 + n2 ) G a + H a K a ) 1 E a .

(36)
Dc and Dp are two additional gain matrices that may be
tuned to give the desired closed-loop transient response and

(38)

At the transient, z = , and from (38) we have y(z) = Du(z)


Thus, it can be shown that at the transient the closed-loop
system in Fig. 2 reduces to
1

y p ( z ) = D p Dc I p1 D p Dc r ( z )

We can use the LQR method to find an optimal control law for the system in (25) as

(37)

y ( z ) = [ D + C ( zI G ) 1 H ] u ( z ) .

where

GCL

disturbance rejection for the system during the initial phase


of the response following a change in the reference signal
or disturbance. It can be observed that a system with state
equations of the form

(39)

+ I p1 D p Dc d ( z ) ,

when Er = Ip1. For a multivariable system, dc and dp can be


defined as dc = || Dc || and dp = || Dp || . This gives

y p ( z) =

d p dc
1 d p dc

r ( z) +

1
d ( z) .
1 d p dc

(40)

At the transient, we desire y p ( z ) 1 r ( z ) + 0 d ( z ) for


input tracking and disturbance rejection. From (40) we see
that by selecting

d c >> 1 ,

d p >> 1 ,

(41)

the desired result is achieved.


When the direct link from input to output of the plant
is permissible, we can select dc and dp as in (41) for input
tracking and disturbance rejection. For the case where the
direct link from input to output of the plant is not allowed,
but the plant has a semi-proper transfer function matrix,
then dc >> 1 can be chosen. To achieve input tracking and
disturbance rejection when the input-output link is not
permissible and the plant has a strictly proper transfer function matrix, an observer can be designed for the plant in
(17) with Dp = 0, using the equivalent disturbance
method proposed in [10]. Defining a dual system with Gdual
= GpT, Hdual = CpT, Cdual = HpT, and Qdual = CdualT(W)Cdual,
gives a solution to the discrete-time Riccati equation,

Pdual = GdualT Pdual Gdual + Qdual


T

H dualT Pdual Gdual Rdual + H dualT Pdual H dual H dualT Pdual Gdual ,
(42)
where the weighting matrices W and Rdual can be tuned to
achieve the desired closed-loop transient tracking and disturbance rejection. The observer gain is calculated as

J oT = ( Rdual + H dualT Pdual H dual ) 1 H dualT Pdual Gdual . (43)


Now, using x p (kT ) to represent the estimate of the real

J.M. Madsen et al.: State-Space Digital PID Controller Design for Multivariable Analog Systems

plant states xp(kT), we have the discrete-time observer state


equation as

0.766 e s

0
+
0.606 e 7 s

x p (kT + T ) = G p x p (kT ) + H p u p (kT )


J o C p x p (kT ) y p (kT ) ,

(44)

y p (kT ) = C p x p (kT ) + d (kT ) .

IV. SIMULATION EXAMPLES


Example 1. Consider an example using a slightly modified
version of the Wood and Berry empirical model of a distillation column that is used to separate a methanol-water
mixture [1],

18.9 e3s

21 s + 1 u1 ( s ) d1 ( s )

+
,
19.4 e3s u2 ( s ) d 2 ( s )

14.4 s + 1
(46)

where the outputs, y1(s) and y2(s), are the overhead and
bottom compositions of methanol, respectively, and the
inputs, u1(s) and u2(s), are the reflux flow rate and steam
flow rate to the boiler. Output disturbances d1(s) and d2(s)
are also included. The transfer function from u1(s) to y2(s)
has been adjusted to set up unstable dynamics for the
transfer function with the largest delay.
To track a specific ramp input and reject constant disturbances, Fliess et al. [1] developed an original generic
model as in (8) and designed a generalized PI controller as

y + (s)
y + ( s)
,
K II
s
s2

x11 (t )

y1 (t ) 1 1 0 0 x12 (t )
=
.

y2 (t ) 0 0 1 1 x21 (t )
x22 (t )

(45)

It should be noted that according to the Internal Model


Principle (IMP) [10,19], at the steady state, the PID controller is capable to carry out input tracking and disturbance
rejection for constant inputs and disturbances.

u (s) = K p x + ( s) K I

(48)

with the output,

with the output described by

12.8 e s

y1 ( s) 16.7 s + 1

=
7 s
y2 ( s) 6.6 e
10.9 s 1

0.900 e 3s u1 (t )
u2 (t ) ,
0

1.347 e 3s

167

(47)

where Kp, KI and KII are control gain matrices, and x+(s)
and y+(s) are predicted state and output, respectively. Implementing this analog controller, with its predicted states
and outputs, may be a complex process.
Transforming the system in (46) with zero output disturbance into its state-space representation as described in
Section 2 gives the following:

0
0
0 x11 (t )
x11 (t ) 0.060
x (t ) 0
0.048
0
0 x12 (t )
12 =

0
0.092
0 x21 (t )
x21 (t ) 0
x (t ) 0
0
0
0.069 x22 (t )
22

(49)

This system is expressed as four subsystems, each with a


different input time delay. Subsystem 1 has delay Td1 = 1,
Subsystem 2 has delay Td2 = 3, the unstable Subsystem 3
has delay Td3 = 7, and Subsystem 4 has delay Td4 = 3. For
this example, we will assume that the sample time, T, is 2.
For Subsystem 1, we have Td1 = (d11)T + 1 = 1, and
we can select d1 = 1 and 1 = 1 to satisfy this equation. The
ratio, r1, is thus given by r1 =

1 1
= = 0.5 .
T 2

Now, Subsystem 1 has a discrete-time state equation


as given in (13),

x11 (kT + T ) = G1 x11 (kT ) + H111u1 (kT T ) + H110 u1 (kT ) .


(50)
The augmented state-space representation is written as

x11 (kT + T ) G1

=
x12 (kT + T ) 0

H111 x11 (kT ) H110

+
u1 (kT ) ,
0 x12 (kT ) 1
(51)

or

x1e (kT + T ) = G1e x1e (kT ) + H1e u1 (kT ) ,

(52)

where G1 = 0.887, H111 = 0.701, and H110 = 0.744.


Using a similar method, we can select d2 = 2, 2 = 1,
and r2 = 0.5 for Subsystem 2 to find the following discretetime state equation:

x21 (kT + T ) = G2 x21 (kT ) + H 211u2 (kT 2T ) + H 210 u2 (kT T ) .


(53)
Again, the augmented state-space representation is

x21 (kT + T ) G2


x22 (kT + T ) = 0
x (kT + T ) 0
23

H 211
0
0

H 210 x21 (kT )

1 x22 (kT )
0 x23 (kT )

0

+ 0 u2 (kT ) ,
1

(54)

or

x2e (kT + T ) = G2e x2 e (kT ) + H 2e u2 (kT ) ,

(55)

Asain Journal of Control, Vol. 8, No. 2, June 2006

168

where G2 = 0.909, H211 = 0.838, and H210 = 0.879.

H1e
0
+
H 3e
0

With d3 = 4, 3 = 1, and r3 = 0.5, Subsystem 3 has a


discrete-time state equation given by

x31 (kT + T ) = G3 x31 (kT ) + H 311u1 (kT 4T )


+ H 310 u1 (kT 3T ) ,

(56)

H 311
0
0
0
0

H 310
1
0
0
0

0
0
1
0
0

x p (kT + T ) = G p x p (kT ) + H p u p (kT ) ,

0 x31 (kT )
0 x32 (kT )

0 x33 (kT )
1 x34 (kT )

0 x35 (kT )

0
0

+ 0 u1 (kT ) .
0

1

y p (kT ) = C p x p (kT ) + D p u p (kT ) ,

1 0 1 0 0 0 0 0 0 0 0 0 0
Cp =
,
0 0 0 0 0 1 0 0 0 0 1 0 0
(64)
(57)

and Dp is to be determined.
For a controller for our two-input two-output system,
let us assume a MIMO filtered PID controller transfer
function matrix described by

Gc ( z ) = G0 ( z ) + Dc ,

x3e (kT + T ) = G3e x3e (kT ) + H 3e u1 (kT ) ,

(58)

x41 (kT + T ) = G4 x41 (kT ) + H 411 u2 (kT 2T )


+ H 410 u2 (kT T ) .

(59)

As before, the augmented state-space representation is

H 411
0
0

H 410 x41 (kT )

1 x42 (kT )
0 x43 (kT )

0

+ 0 u2 (kT ) ,
1

1
G0 ( z ) = diag
.
( z 1) ( z )

(67)

with

(61)

where G4 = 0.870, H411 = 1.214, and H410 = 1.302.


The four subsystems can be combined into one
discrete-time state equation given by

0
0
G3e
0

In this example, for simplicity we will pre-assign the filter


factor as = 0.5, select Dc as the static decoupler for the
system, and set Dp = 02 in (17) for a closed-loop response
controlled solely by the PID controller in (65). In this case,
it is easy to show that the state-space representation of the
filtered PID controller is given by

yc (kT ) = Cc xc (kT ) + Dc uc (kT ) ,

or

0
G2e
0
0

(66)

xc (kT + T ) = Gc xc (kT ) + H c uc (kT ), xc (0) = 0 ,

(60)

x4e (kT + T ) = G4e x4e (kT ) + H 4e u2 (kT ) ,

(65)

where G0(z) is of the form

where G3 = 1.201, H311 = 0.695, and H310 = 0.634.


Finally, we select d4 = 2, 4 = 1, and r4 = 0.5 for Subsystem 4 to get the following discrete-time state equation:

x1e (kT + T ) G1e


x (kT + T ) 0
2e
=
x3e (kT + T ) 0
x (kT + T ) 0
4e

(63)

where

In compact form, this is

x41 (kT + T ) G4


x42 (kT + T ) = 0
x (kT + T ) 0
43

(62)

Written in compact form, we now have plant state equations of the form

and an augmented state equation,

x31 (kT + T ) G3
x (kT + T ) 0
32

x33 (kT + T ) = 0
x (kT + T ) 0
34

x35 (kT + T ) 0

0
H 2 e u1 (kT )

.
0 u2 (kT )
H 4 e

0 x1e (kT )
0 x2 e (kT )

0 x3e (kT )
G4e x4 e (kT )

1
0

0.5 1.5
Gc =
0
0

0
0
1

0
0
0
0
,
0
1

0.5 1.5

1 0 0
,
0
0 1 1

Cc =

1
Hc =
0

0
0
,
0

0.153
.

0.054
0.104

0.157

Dc =

(68)
Now using a modified plant input as in (23),

u p (kT ) = u p (kT ) + Dc uc (kT ) ,

(69)

the augmented system state equations become

xa (kT + T ) = G a xa (kT ) + H a u p (kT ) + E a Er r (kT ) ,

J.M. Madsen et al.: State-Space Digital PID Controller Design for Multivariable Analog Systems

ya (kT ) = y p (kT ) = Ca xa (kT ) ,

169

(70)

as described in (25) and (26).


The LQR method described in Section 3 is used to determine the optimal control law for the system, where the
performance index weighting parameters take the values of
Q = 0.01*I17 and R = I2. Now, solving the Riccati equation
in (34), the optimal control gain, Ka, is calculated as

K a = R + H aT PH a

H aT P G a = [ K1 K 2 ] ,

(71)

where

K1 =
0.539 0.438 0.606 0.573 1.181 3.310 1.920 3.319 2.683 2.126 0.398 0.568 1.186

,
0.410 0.262 0.440 0.333 0.632 0.267 0.253 0.515 0.552 0.543 0.365 0.272 0.424

(72)
and

0.051 0.100 0.052 0.104


K2 =
.
0.057 0.116 0.050 0.102

(73)
Fig. 3. Closed-loop response.

The desired control law is given in (22) by

u p (kT ) = K1 x p (kT ) K 2 xc (kT ) Dc uc (kT ) . (74)

The ramp-input response of the controlled unstable


analog system in (46) via the predicted PI controller (47)
from [1] is shown in Fig. 4.

Thus, the optimally designed closed-loop system in (29)


has become

xa (kT + T ) = (G a H a K a ) xa (kT ) + E a Er r (kT ) ,


ya (kT ) = Ca xa (kT ) ,

(75)

with poles at
0.134 0.495i 0.044 0.201i 0.000 0.000 0.000
0.442

.
0.400
0.442 0.760 0.907
0.530 0.361i 0.631 0.289i

(76)
The corresponding digital PID controller has the form of

Gc ( z ) = Dc + K 2 ( zI 4 Gc ) H c
1

0.097 z 0.005( z 1)
0.104 z

0.049
0.065 + z 1 z 0.5

1
=
.
z
z

z
z

0.118
0.004(
1)
0.105
0.006(
1)
0.068 +

+
0.215 +
+

z 1
z 0.5
z 1
z 0.5

(77)
The gain, Er can be chosen to tune the final value of
the system response. For this example, using the final value
theorem as in (36) gives Er = I2.
Figure 3 shows a numerical simulation of the closedloop system using the designed digital controller with the
original time-delayed plant in (46). A ramp transfer from y1
= 0 at t = 10 to y1 = 2 at t = 50 is desired while keeping y2
= 0. A load perturbation of amplitude 0.5 occurs on y1 at t =
62, and a second load perturbation of the same amplitude
occurs on y2 at t = 77. The above-mentioned performance
and disturbance conditions were specified in [1].

Fig. 4. Closed-loop response from [1].

Since the design objectives and controller structures of the


two methods are different, it is not appropriate to directly
compare these methods. Nevertheless, from Fig. 3 and Fig.
4 we observe that the results from this papers method exhibit a smaller overshoot in both inputs and outputs but
slower ramp-input tracking than those in [1]. It is important
to note that the proposed digital PID controller can be more
easily implemented than the predicted analog PI controller
in (47).

Asain Journal of Control, Vol. 8, No. 2, June 2006

170

Example 2. This example uses a model of an industrialscale polymerization reactor [10],

22.89 e 0.2 s

y1 ( s) 4.572 s + 1
=


0.2 s
y2 ( s) 4.689 e
2.174 s + 1

11.64 e0.4 s

1.807 s + 1 u1 ( s)

5.80 e 0.4 s u2 ( s )

1.801 s + 1

4.243 e 0.4 s

3.445 s + 1
+
d (s) .
0.601 e 0.4 s

1.982 s + 1

(78)

For this system, the outputs y1(s) and y2(s) are measurements representing the reactor conditions, the inputs u1(s)
and u2(s) are the set-points of two feed-flow loops, and the
disturbance d(s) represents the purge flow rate of the reactor.
Transforming this system with zero output disturbance
into its state-space representation as described in Section 2
gives the following:

0
0
0
x11 (t ) 0.219

0
0
0.553
x12 (t ) = 0
x (t ) 0
0
0.460
0
21

0
0
0.505
x22 (t ) 0
5.007 e

+
0.2 s
2.157 e

0
0.2 s

x11 (t )

x12 (t )
x (t )
21
x22 (t )

0
6.442 e

0.4 s

0
3.220 e 0.4 s

u1 (t )
u (t ) ,
2

(79)

0
H 2 e u1 (kT )

.
0 u2 (kT )
H 4 e

(81)

with

0.947 0.969
G1e =
,
0
0

0.249 0
H1e =
,
0
1

0.871 0.877 0.627


0 0

0
0
1
, H 2e = 0 0 ,
G2e =

0
0
0 1
0
0.891 0.403
,
G3e =
0
0

0.107 0
,
H 3e =
0
1

0.870 0.439 0.313


0
1 ,
G4e = 0

0
0
0

H 4e

0 0
= 0 0 .

0 1
(82)

Written in compact form, we have plant state equations of


the form

x p (kT + T ) = G p x p (kT ) + H p u p (kT ) ,


y p (kT ) = C p x p (kT ) + D p u p (kT ) ,

(83)

where

1 0 1 0 0 0 0 0 0 0
Cp =
,
0 0 0 0 0 1 0 1 0 0

(84)

and Dp is to be determined.
The initial PID controller for this system is designed,
using the Internal Model Principle (IMP) [10,19], as

Gc ( z ) = G0 ( z ) + Dc ,

with the output,

(85)

where G0(z) is of the form

x11 (t )

y1 (t ) 1 1 0 0 x12 (t )
.

y2 (t ) 0 0 1 1 x21 (t )
x22 (t )

(80)

This system is expressed as four subsystems, each with its


own input time delay. For this example, we will assume
that the sample time, T, is 0.25. The four subsystems can be
combined into one discrete-time state equation given by

x1e (kT + T ) G1e


x2e (kT + T ) = 0
x3e (kT + T ) 0


x4e (kT + T ) 0

H1e
0
+
H 3e
0

G2e

G3e

0
0

G4e

x1e (kT )

x2 e (kT )
x3e (kT )

x4 e (kT )

( z 1)( z )
1
G0 ( z ) =

,
1

( z 1)( z 2 )
0

(86)

with the i values selected from the poles of the disturT

bance dynamics in (78) as 1 = e 3.445 = 0.930 , and


T

2 = e1.982 = 0.882 . For simplicity, Dc is selected as an


identity matrix, Dc = I2. In this case, it can be shown that
the state-space representation of the filtered PID controller
is given by

xc (kT + T ) = Gc xc (kT ) + H c uc (kT ), xc (0) = 0 ,


yc (kT ) = Cc xc (kT ) + Dc uc (kT ) ,

(87)

J.M. Madsen et al.: State-Space Digital PID Controller Design for Multivariable Analog Systems

with

0
0
1.0 0
1
0 1.0
0
0
0
, Hc =
Gc =
0 0.930
0
0
1
0

0
0
0
0.882

0
1
,
0
1

0
0
14.286
14.286
,
Cc =
8.438
0
8.438
0
1 0
Dc =
.
0 1

(88)

Using the augmented state equation in (25),

xa (kT + T ) = G a xa (kT ) + H a u p (kT ) + E a Er r (kT ) ,


(89)
selecting the optimal LQR parameters in (32) as Q = I14 and
R = I2, and experimentally determining a value of Dp =
20*I2 in order to negate the effects of the output disturbance, the optimal control gain in (31), Ka = [K1 K2], can be
calculated as

K1 =
0.038 0.002 0.045 0.006 0.003 0.004 0.002 0.001 0.000 0.000

,
0.002 0.002 0.011 0.011 0.020 0.044 0.002 0.041 0.000 0.004

(90)
and

0.501 0.000 0.193 0.000


K2 =
.
0.000 0.501 0.000 0.183

(91)

The desired control law is given in (22) by

u p (kT ) = K1 x p (kT ) K 2 xc (kT ) Dc uc (kT ) . (92)


Thus, the optimally designed closed-loop system in (29)
has become

The gain, Er is chosen to tune the final value of the system


response as in (36). For this example, Er = I2.
In [10], an analog PID controller was developed for
this same system. Figure 5 compares the unit-step response
of the original time-delayed system using this papers designed digital controller applying Dp without a state observer, the designed digital controller with Dp = 0 and the
observer described below, the analog controller with a state
observer in [10], and the analog controller without a state
observer in [10]. The sampling period in these results was
T = 0.25, and unit set-point changes were introduced in
y1(t) at t = 1, and y2(t) at t = 20. In addition, unit disturbances were introduced at t = 40. These simulation results
show that the analog controller developed in [10] requires a
high-order state observer for disturbance rejection. The
design method of this paper developed the digital controller
based on the exact state, and makes use of a feedforward
plant gain, Dp, to accomplish even better transient tracking
and disturbance rejection without the need for a state observer.
When the exact state is not available, the exact
discrete-time state can be constructed from the input and
output data of the plant via the ideal state reconstructor
method [23,24]. When the state is not available and the
disturbance rejection for the plant with a strictly proper
transfer function matrix is required, a multi-objective digital observer can be constructed, in which the observation
error at the transient state is purposely retained and utilized
in load disturbance compensation. For this example, values
of the optimal LQR parameters in (42) were selected as W
= 100*I2 and Rdual = I2. This gives the optimal observer
gain,

J oT =
0.651 0.000 0.701 0.000 0.000 0.260 0.000 0.350 0.000 0.000

,
1.274 0.000 1.618 0.000 0.000 0.508 0.000 0.809 0.000 0.000

(96)

xa (kT + T ) = (G a H a K a ) xa (kT ) + E a Er r (kT ) ,


ya (kT ) = (Ca D p K a ) xa (kT ) + D p E r Er r (kT ) ,
(93)
with poles at

0.000, 0.000, 0.000, 0.126 0.061i, 0.244 0.005i,

.
0.843 0.065i, 0.870, 0.916, 0.918, 0.951, 0.153
(94)
The corresponding digital PID controller has the form of

Gc ( z ) = Dc + K 2 ( zI 4 Gc ) H c
1

0.500 z 2.762( z 1)

0
3.262 + z 1 z 0.930

.
=
0.500 z 1.545( z 1) (95)

0
2.045
+

z 1
z 0.882

171

[10]
[10]

Fig. 5. Closed-loop response, T = 0.25.

Asain Journal of Control, Vol. 8, No. 2, June 2006

172

The same parameters for the digital PID controller were


selected as in (88), but because for this case Dp = 02, and
the plant is represented by a strictly proper transfer function matrix, a smaller value of Dc was selected as

0.031 0.062
Dc =
.
0.025 0.122

(97)

As the static decoupler for the system, this value was selected to reduce the overshot at the transient and increase
the decoupling of the system. As a result of this change, the
digital PID controller must be retuned as in (33).
It is important to note that with a sampling time of
T = 0.25, the analog controller in [10] transformed to a
digital controller via the Bilinear Tustin Transform method
[18] makes the system unstable, but the digitally designed
controller proposed in this paper still performs extremely
well. Figure 6 reduces the sample time to T = 0.15, and
compares the response of the original time-delayed system
using this papers designed digital controller applying Dp
without a state observer, the designed digital controller
with Dp = 0 and the observer described above, the analog
controller with a state observer in [10], and the controller
and observer in [10] transformed to digital via the Bilinear
Tustin Transform method.
From Figures 5 and 6, it is observed that the proposed
digital controller gives quicker tracking and much better
disturbance rejection than the analog controller developed
in [10] due to the introduced plant feedforward gain, Dp.
When the direct link from input to output of the plant is not
permissible, a multi-objective observer is constructed for
disturbance rejection. Figure 6 shows that this closed- loop
response still gives slightly quicker tracking and better
disturbance rejection than the analog controller developed
in [10] and transformed to digital via the Bilinear Tustin
Transform method.
Another notable feature of this control design is that
due to the gain Dp, the control input to the plant, up(kT), is
significantly less than the analog inputs in the analog system of [10]. This result for T = 0.25 is shown in Fig. 7.

[10]
[10]

Fig. 6. Closed-loop response, T = 0.15.

[10]
[10]

Fig. 7. Control inputs, T = 0.25.

For a stable or unstable multivariable high-order system with significant multiple time delays, a bisection
searching method is suggested to find a suitable sampling
period for digital design. In this manner, a reasonable
tradeoff between the closed-loop response and the computational complexity caused by a high-dimensional generic
state-space model can be achieved.

V. CONCLUSION
In this paper, a new methodology for discrete-time
state-space optimal design of digital PID controllers was
developed for a multivariable system represented by a
semi-proper or strictly proper transfer function matrix with
multiple time delays. The advantages of the proposed control approach compared to existing methods are as follows:
(1) an exact discrete-time state-space model of the system
can be constructed so that direct digital design of the controller may be accomplished; (2) digital implementation of
the controller reduces hardware complexity and increases
flexibility of the system; (3) the MIMO PID control parameters are systematically tuned using LQR such that the
closed-loop stability is guaranteed; (4) this method can be
applied to plants regardless of system stability, dimension
and degree of system, minimum-phase properties, length of
dead time, and system coupling properties; (5) a method to
improve transient response and accomplish disturbance
rejection is proposed. A disadvantage of the proposed approach is that the dimensions of the augmented state-space
description become significantly large when the sampling
period T is extremely small compared to the time delay Td.
As a result, computational difficulties with the proposed
method might occur for these cases. The simulation results
demonstrate that this controller design method provides
good performance with input tracking and disturbance rejection for systems with significant multiple time delays.

J.M. Madsen et al.: State-Space Digital PID Controller Design for Multivariable Analog Systems

REFERENCES
1. Fliess, M., R. Marquez, and H. Mounier, An Extension
of Predictive Control, PID Regulators and Smith Predictors to Some Linear Delay Systems, Int. J. Contr.,
Vol. 75, No. 10, pp. 728-743 (2002).
2. Astrom, K.J., and B. Wittenmark, Computer Controlled
Systems Theory and Design, Prentice-Hall, New Jersey, pp. 38-42 (1999).
3. Silva, G.J., A. Datta, and S.P. Bhattacharyya, New Results on the Synthesis of PID Controllers, IEEE Trans.
Automat. Contr., Vol. 47, No. 2, pp. 241-252 (2002).
4. Fridman, E., L. Fridman, and E. Shustin, Steady Modes
in Relay Control Systems with Time Delay and Periodic
Disturbances, ASME J. Dyn. Sys. Meas. Contr., Vol.
122, pp. 400-407 (2000).

173

16. Yackel, R.A., B.C. Kuo, and G. Singh, Digital Redesign of Continuous Systems by Matching of States at
Multiple Sampling Periods, Automatica, Vol. 10, No. 1,
pp. 105-111 (1974).
17. Shieh, L.S., B.B. Decrocq, and J.L. Zhang, Optimal
Digital Redesign of Cascaded Analogue Controllers,
Optim. Contr. Appl. Meth., Vol. 12, pp. 205-219 (1991).
18. Kuo, B.C., Digital Control Systems, Holt, Rinehart and
Winston, New York, pp. 321-338, 404-405, 601-645
(1980).
19. Goodwin, G.C., S.F. Graebe, and M.E. Salgado, Control
System Design, Prentice-Hall, N.J., pp. 162-169, 267268, 685-688 (2001).
20. Wang, Q.G., C.C. Hang, and X.P. Yang, Single- Loop
Controller Design via IMC Principles, Automatica, Vol.
37, No. 12, pp. 2041-2048 (2001).

5. Normey-Rico, J.E. and E.F. Camacho, Robust Tuning


of Dead-Time Compensators for Processes with an Integrator and Long Dead-Time, IEEE Trans. Automat.
Contr., Vol. 44, pp. 1597-1603 (1999).

21. Wang, Q.G., H. Ru, and X.G. Huang, An Effective


Frequency Domain Approach to Tuning Non-PID Controllers for High Performance, ISA Trans., Vol. 41, pp.
37-49 (2002).

6. Sun, Y.J., J.G. Hsieh, and H.C. Yang, On the Stability


of Uncertain Systems with Multiple Time-Varying Delays, IEEE Trans. Automat. Contr., Vol. 42, pp.
101-105 (1997).

22. Zhang, Y., Q.G. Wang, and K.J. Astrom, Dominant


Pole Placement for Multi-Loop Control Systems,
Automatica, Vol. 38, No. 7, pp. 1213-1220 (2002).

7. Yan, J.J., J.S.H. Tsai, and I.E. Sheen, Stability Analysis


of Large-Scale Time-Delay Systems via Evolutionary
Programming Algorithm, ASME J. Dyn. Sys. Meas.
Contr., Vol. 123, pp. 293-296 (2001).
8. Zhang, Y., L.S. Shieh, C.R. Lu, and S.M. Guo, Digital
PID Controller Design for Multivariable Analogue Systems with Computational Input-Delay, IMA J. Math.
Contr. Info., Vol. 21, pp. 433-456 (2004).
9. Zhang, Y., L.S. Shieh, C.M. Akujuobi, and W. Ali,
Digital PID Controller Design for Delayed Multivariable Systems, Asian J. Contr., Vol. 6, No. 4, pp.
483-495 (2004).
10. Zhang, Y., L.S. Shieh, and A.C. Dunn, PID Controller
Design for Disturbed Multivariable Systems, IEE
Proc.-Contr. Theory Appl., Vol. 151, No. 5 (2004).
11. Astrom, K.J., K.H. Johansson, and Q.G. Wang, Design
of Decoupled PI Controllers for Two-by-Two Systems,
IEE Proc. Contr. Theory Appl., Vol. 149, pp. 74-81
(2002).
12. Chen, D. and D.E. Seborg, Multiloop PI/PID Controller
Design Based on Gershgorin Bands, IEE Proc. Contr.
Theory Appl., Vol. 149, pp. 68-73 (2002).
13. Zheng, F., Q.G. Wang, and C.C. Hang, On the Design
of Multivariable PID Controllers via LMI Approach,
Automatica, Vol. 38, No. 3, pp. 517-526 (2002).
14. Chen, T. and B. Francis, Optimal Sampled-Data Control
Systems, Springer-Verlag, London, pp. 221-239 (1995).
15. Rafee, N., T. Chen, and O.P. Malik, A Technique for
Optimal Digital Redesign of Analog Controllers, IEEE
Trans. Contr. Syst. Technol., Vol. CST-5, pp. 89-99
(1997).

23. Polites, M.E., Ideal State Reconstructor for Deterministic Digital Control Systems, Int. J. Contr., Vol. 49,
pp. 2001-2011 (1989).
24. Shieh, L.S., G.C. Chen, and J.S.H. Tsai, Hybrid Suboptimal Control of Multi-Rate Multi-Loop Sampled-Data
Systems, Int. J. of Sys. Sci., Vol. 23, No. 6, pp. 839-854
(1992).

Вам также может понравиться