Proc. Natl. Sci. Counc.

ROC(A)
Vol. 23, No. 6, 1999. pp. 782-788

(Scientific Note)

State-Space Model Conversion of a System with State

Delay
C HUEN -M ING CHEN AND K UANG-H WA WANG

Department of Electrical Engineering

Chung-Cheng Institute of Technology
Taoyuan, Taiwan, R.O.C.

(Received January 12, 1999; Accepted June 11, 1999)

ABSTRACT

A new approach to model conversion of a state-delay system is introduced in this paper. The discrete-
time model is especially derived from the continuous-time system, that it has a wider range for the sampling
period in real implementation. A simple method is also proposed to determine the unknown state-delay
time from the available discrete-time model, and the continuous-time model is established through some
important results demonstrated in this paper.

Key Word: state delay, state-space model, model conversion

I. Introduction finite spectrum assignment method to solve the design

The time delay phenomenon in control paths or
state variables is unavoidable in many physical systems.
Time delays are common in real systems, for example,
chemical processes and biological systems, and the
design of controllers for such systems depends criti-
cally on knowledge of the delays. It is well known
that control system behavior is more sensitive to time
delay than to other linear system parameters. In fact,
a closed loop control system may be unstable or may
exhibit unacceptable transient response characteristics
if the time delay used in the system model for controller
design does not coincide with the actual process time
delay. Ignorance of the computation delay during
analysis and design of digital control systems may lead
to unpredictable and unsatisfactory system performance.
Generally speaking, there are two types of time
delay: input delay and state delay. Input delay is caused
by the transmission of a control signal over a long
distance, or the delay is sometimes built into a system
deliberately for control purposes (Jury, 1964). State
delay is due to transmission or transport delay among
interacting elements in a dynamic system (Tsien, 1954;
Amstrong and Tripp, 1981). The input delay system
has been developed completely and has been reported
by Amstrong and Tripp (1981), Astrom and Wittenmark
(1990), and Chen and Chang (1998). As for the state
delay system, Manitius and Olbort (1979) proposed a
problem. Alekal (1969), Chyung and Lee (1966), Eller
et al. (1969), Muller (1971), and Ross (1971) studied
the optimal control theory for the state delay system.
However, most methods were developed for continu-
ous-time systems rather than discrete-time systems and
seem to be impractical, not only the implementation
difficulty, but also because developed control laws or
design approaches which involve too much analytical
mathematics.
Based on the developed discrete-time model for
a type of state delay, Amstrong and Tripp (1981) used
the discrete-time control theory to design a digital
controller that can be easily realized by means of
microprocessors. However, the developed method is
restricted in that the sampling period must be shorter
than the state delay time. This is not convenient for
real implementation. Recently, Chen and Chang (1998)
proposed an improved step to avoid this limitation, but
the general formation can not be obtained via such a
method, also it can not simplify the digital implementa-
tion. In this paper, a general counterpart of the discrete-
time model is derived, where the sampling period can
be longer than the state delay time, and where the
general form of the microprocessor application is more
effective and convenient in real implementation.
Applying these discrete-time techniques to state delay
systems is one of the main goals of this paper. Another
main purpose of this paper is to determine the unknown

− 782 −
 State-Space Model Conversion with Delay

state delay from the derived discrete-time model. Based x(kT)= 1 [x(kT+ τ )+x(kT− τ )]; (5a)
2
on the geometric-series method (Shieh et al., 1980) and
the principal nth-root of a matrix (Shieh et al., 1986), thus,
the state-delay model conversion can be obtained
in state-space form, which will be useful in future x(kT−τ )=2x(kT)−x(kT+ τ). (5b)
applications. Since the proposed method is derived
based on the state-space model, it can be easily ex- From Eq. (5b), the main problem appears to be x(kT+τ),
tended for the MIMO (multi-input and multi-output) so the solution x(kT− τ ) can be evaluated as sampled
system. data. The simple derivation of x(kT+ τ) is given in the
following.
II. Continuous-to-Discrete Model Let us consider the state transition equation in Eq.
Conversion (2a) by using a time substitution to obtain

A linear time-invariant system with state delay is x(kT+T+ τ )

described by
kT + T + τ
=e A(T+τ)x(kT)+ e A(kT+T+τ−λ)A1 x( λ −τ )dλ
x (t)=Ax(t)+A 1x(t− τ)+Bu(t), (1) kT

kT + T + τ
where A1∈Rn×n is the system matrix for the state delay. + e A(kT+T+τ−λ )Bu(kT)d λ . (6)
kT
A∈R n×n and B∈R n×p are the system matrix and input
matrix, respectively. The discrete-time description of Multiplying Eq. (2a) by e A τ to yields
Eq. (1) with Z.O.H. (zero-order hold) can be obtained
as e Atx(kT+T)

x(kT+T)=Gx(kT)+Fx(t− τ )+Hu(kT), (2a) kT + T

=e A(T+τ)x(kT)+ e A(kT+T+τ− λ)A 1x( λ − τ)d λ
kT
where T is the sampling period, and
kT + T
AT
+ e A(kT+T+ τ−λ)Bu(kT)d λ . (7)
G=e , (2b) kT

kT + T Subtracting Eq. (7) from Eq. (6) yields

F= e A(kT+T−λ)A1 dλ =[G−I n ]A−1A 1, (2c)
kT
x(kT+T+ τ )−e Aτx(kT+T)
kT + T
H= e A(kT+T−λ )Bd λ=[G−I n]A −1B. (2d) kT + T + τ
kT = e A(kT+T+τ−λ )A 1x( λ − τ)d λ
kT + T
It is seen that the main problem of Eq. (2a) depends
on whether x(t− τ ) in kT~kT+T can be represented as kT + T + τ
+ e A(kT+T+τ−λ )Bu(kT)d λ . (8)
piecewise-constant data. Utilizing the mean-value kT + T
method, x(t− τ ) in kT~kT+T becomes
Simplifying the integral in Eq. (8), we can obtain
x(t− τ )≈ 1 [x(kT+T− τ )+x(kT− τ )], (3)
2 x(kT+T+ τ)−G 1x(kT+T)
and by Eq. (3), we can obtain Eq. (2a) as
= 1 P 1[x(kT+T+ τ)+x(kT+T)]+Γ 1 u(kT), (9a)
2
x(kT+T)=Gx(xT)+ 1 F[x(kT+T− τ )+x(kT− τ)]
2 where
+Hu(kT). (4)
G 1=e At, (9b)
1. Case 1: τ <T kT + T + τ
P1= e A(kT+T+ τ− λ )A 1 d λ =[e Aτ−I n]A −1A 1 , (9c)
kT + T
It seems to be too difficult to calculate x(kT+T−
τ) and x(kT−τ). We can evaluate x(kT) using the mean kT + T + τ
value of x(kT+ τ ) and x(kT− τ ): Γ 1= e A(kT+T+τ−λ )Bdλ =[e Aτ−I n]A −1B, (9d)
kT + T

− 783 −
 C.M. Chen and K.H. Wang

and I n denotes an n×n identity matrix. By rearranging where

Eq. (9), we can obtain
S 1=(I n+ 1 P 1) −1, (14b)
x(kT+T+ τ )=(G 1+ 1 P 1)x(kT+T) 2
2
S 2=(2I n−G 1− 1 P1 ), (14c)
+ 1 P 1x(kT+T− τ )+Γ 1 u(kT). (10) 2
2 1
Q=(I n + FS 1S 2) −1. (14d)
If we move backwards one sampling period T, Eq. (10) 2
becomes 2. Case 2: τ >T
x(kT+ τ)=(G 1+ 1 P 1)x(kT) May other researchers have proposed versions of
2
1 the state-delay model that are suitable for τ>T. Amstrong
+ P 1x(kT− τ )+Γ 1u(kT−T). (11) and Tripp (1981) used the trapezoidal integration to
2
derived a general discrete-time model as
Using Eq. (11), the above Eq. (5b) can be obtained as
x(kT+T)=Gx(kT)+ T e ATA 1x(kT−mT)
x(kT− τ )=2x(kT)=(G 1+ 1 P 1)x(kT)− 1 P 1x(kT−τ ) 2
2 2
T
+ A 1x(kT+T−mT)+Hu(kT). (15)
−Γ 1u(kT−T), (12a) 2
Chen and Chang (1998) used the block pulse function
and x(kT− τ ) is to derive a general discrete-time model as

x(kT− τ)=(I n + 1 P 1 ) −1(2I n−G 1 − 1 P 1 )x(kT) x(kT+T)=Gx(kT)+ 1 F[x(kT+T−mT)+x(kT−mT)]

2 2 2
−(I n+ 1 P 1) −1Γ 1 u(kT−T). (12b) +Hu(kT). (16)
2
If we move forward one sampling period T, Eq. (12b) Comparing Eqs. (4), (15), and (16), we find that they
becomes all assume that τ≈mT, and that their sampled-data models
are all the same. It is hard to determine which model
x(kT+T−τ )=(I n+ 1 P 1) −1(2I n−G 1 − 1 P 1)x(kT+T) is more accurate (Chen and Chang, 1998).
2 2
−(I n + 1 P 1 ) −1Γ 1u(kT). (13)
2 III. Discrete-to-Continuous Conver-
sion
Using Eq. (12b) and Eq. (13), the discrete-time model
of Eq. (4) becomes Assume that the state-delay τ =mT+ τ ' (m is an
integer, m≥1, τ'<T). To evaluate the integral term in
x(kT+T) Eq. (2c), we can split kT~kT+T into kT~kT+ τ ' and
kT+ τ '~kT+T, respectively. Then we have
=Gx(kT)+ 1 F[x(kT+T− τ)+x(kT− τ)]+Hu(kT)
2
x(kT+T)=Gx(kT)+F1 x(kT−T−mT)
=e ATx(kT)+ 1 F[(I n + 1 P 1 ) −1(2I n −G 1− 1 P 1)x(kT+T)
2 2 2
+F 0x(kT−mT)+Hu(kT), (17a)
−(I n+ 1 P 1) −1Γ 1u(kT)+(I n + 1 P 1 ) −1(2I n −G 1− 1 P 1)
2 2 2 where
• x(kT)−(I n + 1 P 1 ) Γ 1u(kT−T)]+Hu(KT),
−1
kT + τ '
2 F1= e A(kT+T− λ)A 1d λ =[G−G (1− γ)]A −1 A 1, (17b)
kT
and the final result becomes
kT + T
x(kT+T) F 0= e A(kT+T− λ)A1 dλ =[G (1− γ)−I n]A −1A 1, (17c)
kT + τ '

=Q(G+ 1 FS 1S 2 )x(kT)+Q(H− 1 FS 1 Γ 1)u(kT)

2 2 γ = τ' . (17d)
T
1
− QFS 1 Γ 1 u(kT−T),
2
(14a) Let (1− γ )≡q/l be a fractional number, where G (1− γ)=

− 784 −
 State-Space Model Conversion with Delay

l
( G )q represents the q's power of the principal lth-root =[I n− 1 (1− γ )TA]−1(1− γ )TAA −1A1
2
(Shieh et al., 1986) of the matrix G. If A in Eq. (17)
is singular, the computations of F1 and F0 can be carried =[I n− 1 (1−γ )TA] −1(1− γ )TA1 . (23a)
2
out using the geometric-series method (Shieh et al.,
1980). It is noted that F can be rewritten as By means of some mathematical operations, Eq. (23a)
becomes
F=F 1+F 0. (18)
F0=(1−γ)T[A1+ 1 AF0]=(T−τ')[A1+ 1 AF 0]≡(T−τ')J,
2 2
Let us consider the discrete-time system (G, F 1 , F 0) (23b)
as shown in Eq. (17a). We wish to find the associated
A, A1, B and τ in the continuous-time system as shown where J=A1 1 AF 0 . By comparing the associated ele-
2
in Eq. (1). The relationship between the matrices A ments of F 0 and J in Eq. (23b), the state-delay time
and G is τ ' can be evaluated as

A= 1 ln G. n p f ij
T
(19)
τ ' = T – n ×1 p Σ Σ
i =1 j =1 J
, (24)
ij
From Eq. (2c) and Eq. (18), we can obtain
where f ij and J ij represent the elements of the ith-row
A1 =A(G−I n) −1 F=A(G−I n ) −1(F1 +F 0). (20) and jth-column of F 0 and J, respectively. From the
viewpoint of estimation structure, F1 and F0 are related
From Eq. (2d), we can obtain to x(kT−T−mT) and x(kT−mT), respectively. Once m
is identified, the delay time is found to be τ =mT+ τ '.
B=A(G−I n) −1H. (21) Using Eqs. (19)-(21) and Eq. (24), we can obtain the
continuous-time model in Eq. (1) from the discrete-
The determination of the state-delay time τ can be time model in Eq. (17a).
described as follows. Let us consider the following
infinite series: IV. Examples
G 1−γ 1. Example 1

=e (1−γ)AT Let us consider a wind tunnel system (Amstrong

and Tripp, 1981), where the control input is a unit step
=I n+(1− γ )TA+ 1 [(1− γ )TA] 2+ 1 [(1− γ )TA] 3+... function. The state-delay system is described by
2! 3!
∞
=I n+(1− γ )TA+ 1 [(1− γ )TA] 2+ Σ 1 [(1− γ )TA] j x (t)=Ax(t)+A 1x(t− τ )+Bu(t)
2! j = 3 j!

∞ with
≈I n +(1− γ )TA+ 1 [(1− γ )TA] 2+ Σ j1– 1 [(1− γ )TA] j
2! j =3 2
– 0.1
=I n+{I n+ 1 [(1− γ )TA] x(0) = 8.55 ,
2
0
∞
+Σ 1 [(1− γ )TA] j−1}(1− γ )TA
j =3 2j – 1 where

=I n+{I n− 1 (1− γ )TA} −1(1− γ )TA – 0.509 0 0

2
A= 0 0 1 ,
for 1 (1− γ )T< 1 (22) 0 – 36 – 9.6
2 A

Substituting Eq. (22) into Eq. (17c) yields

0 – 0.006 0
F0 =[G (1−γ)−I n]A −1A 1 A1 = 0 0 0 ,
0 0 0
=[I n+{I n − 1 (1− γ )TA} −1 (1− γ )TA−I n]A −1A 1
2

− 785 −
 C.M. Chen and K.H. Wang

0
B= 0 ,
36

and the time delay τ=0.06 [sec]<T=0.2 sec. Thus, the

corresponding discrete-time model of Eq. (14) is given
by

x(kT+T)

=Q(G+ 1 FS 1 S 2 )x(kT)+Q(H− 1 FS 1 Γ 1)u(kT)

2 2
− 1 QFS 1 Γ 1 u(kT−T),
2
where

0.9032 0 0
G=e AT= 0 0.6245 0.0701 ,
0 – 2.5247 – 0.0488

0 – 0.0011 0
F=[G−I 3]A −1A1 = 0 0 0 ,
0 0 0

1.0000 0.0002 0
S 1=(I 3+ 1 P 1) −1= 0 1.0000 0 , Fig. 1. State trajectories of both the continuous-time model and
2 0 0 1.0000 discrete-time model for τ =0.06 and T=0.2.

S 2=[2I 3−G 1− 1 P 1]
2 0.9032 – 0.0010 0.0000
x(kT + T) = 0 0.6245 0.0701 x(kT)
1.0301 0.0002 0 0 – 2.5247 – 0.0488
= 0 1.0534 – 0.0446 ,
0 1.6069 1.4819
– 0.0001
+ 0.3755 u(kT)
2.5247
1.0000 – 0.0006 0.0000
Q=[I 3 − 1 FS 1 S 2] −1= 0 1.0000 0 ,
2
0 0 1.0000 0.3047 × 10– 4
+ 0 u(kT – T) .
0
0
H=[G−I 3]A −1B= 0.3755 ,
2.5247 The state responses of both the continuous-time (using
the Runge-Kutta fourth-order algorithm) and discrete-
time models are shown in Fig. 1.
0
Γ 1=[G 1 −I 3 ]A −1B= 0.0534 ;
1.6069 2. Example 2

thus, Consider a wind tunnel continuous-time state-

− 786 −
 State-Space Model Conversion with Delay

delay model given as obtained by Eqs. (19)-(21) as

– 0.509 0 0 – 0.5090 0 0
x(t) = 0 0 1 x(t) A = 1 ln G = 0 0.0000 1.0000 ,
T
0 – 36 – 9.6 0 – 36.0000 – 9.6000

A 1 = A(G – I n )– 1F = A(G – I n )– 1(F 1 + F 0)

0 – 0.006 0 0
+ 0 0 0 x(t – 0.06) + 0 u(t) .
0 0 0 36 0 – 0.0060 0
= 0 0 0 ,
By applying continuous-to-discrete model conversion 0 0 0
to Eq. (17a), we can obtain the desired equivalent
discrete-time model with a sampling period of T=0.2
0
sec as B = A(G – I n )– 1H = 0.0000 .
36.0000
x(kT+T)=Gx(kT)+F 1x(kT−T−mT)
Also, by Eqs. (23b) and (24), we can obtain
+F 0 x(kT−mT)+Hu(kT),

where 0 – 0.0058 0
1
J = A 1 + AF 0 = 0 0 0 ,
2
0.9032 0 0 0 0 0
G= 0 0.6245 0.0701 ,
0 – 2.5247 – 0.0488 and the state-delay time τ=0.0600. the obtained results
are quite satisfactory.

0 – 0.0011 0 V. Conclusions
F= 0 0 0 ,
The general model of a state-delay system for the
0 0 0 case where the delay time is shorter than sampling time
has been developed in this paper. In practical applica-
tions, a very short sampling period interval will lead
to an increase in the computation time and will create
0 – 0.3302 × 10– 3 0
a serious dimensionality problem. Based on the model
F1 = 0 0 0 , derived in this paper, we can utilize a longer sampling
0 0 0 time for sampled-data models. Furthermore, it is more
convenient to apply the discrete-time control theory in
order to design a practical delay system. On the other
hand, discrete-to-continuous conversion can be applied
0 – 0.8108 × 10– 3 0 to system identification, and the delay time can be
F0 = 0 0 0 , effectively evaluated to determine its corresponding
state-space model in the continuous-time domain. By
0 0 0
employing the application methods presented in this
paper, a more useful and practical control design can
be obtained.
0
H = 0.3755 .
2.5247 Acknowledgment

This work was supported in part by the National Science

In the meantime, according to G, F, F1, F0, the evaluate Council of the Republic of china under grant NSC 87-2218-E-014-
continuous-time system matrices A , A 1 and B can be 004.

− 787 −
 C.M. Chen and K.H. Wang

References Control, 10, 678-678.

Jury, E. I. (1964) Theory and Application of Z-Transform Method.
Wiely, New York, NY, U.S.A.
Alekal, Y. (1969) Synthesis of Feedback Controllers for Systems
Manitius, A. and A. Olbrot (1979) Finite spectrum assignment
with Time-delay. Ph.D. Dissertation. University of Minnesota,
problem for systems with delays. IEEE Transactions on Auto-
Minneapolis, MN, U.S.A.
matic Control, 13, 541-553.
Amstrong, F. S. and J. S. Tripp (1981) An Application of Multi-
Muller, C. E. (1971) Optimal Feedback Control of Hereditary
variable Design Techniques to the Control of the National
Processes. Ph.D. Dissertation. University of Minnesota,
Transonic Facility. NASA Tech Paper, NASA, Washionton,
Minneapolis, MN, U.S.A.
D.C., U.S.A.
Ross, D. W. (1971) Controller design for time lag systems via a
Astrom, K. J. and B. Wittenmark (1990) Computer-Controller
quadratic criterion. IEEE Transactions on Automatic Control,
Systems-Theory and Design, pp. 40-43. Prentice Hall, Englewood
9, 664-672.
Cliffs, NJ, U.S.A.
Shieh, L. S., H. Wang, and R. E. Yates (1980) Discrete-continuous
Chen, C. M. and K. E. Chang (1998) The sampled-data model of
model conversion. Appl. Math. Modeling, 4, 449-455.
a system with both input and state lag. Journal of Control System
Shieh, L. S., J. S. H. Tasi, and S. R. Lian (1986) Terming continuous-
and Technology, 6, 51-58.
time state equations from discrete-time state equations via the
Chyung, D. H. and E. B. Lee (1966) Linear optimal control problems
principal qth root method. IEEE Trans. Automatic Control, AC-
with delay. SIAM Journal on Control Optimization, 36, 540-
31, 454-457.
575.
Tsien, H. S. (1954) Linear Systems with Time Lag. McGraw-Hill,
Eller, D. H., J. K. Aggarwal, and H. T. Banks (1969) Optimal control
New York, NY, U.S.A.
of linear time-delay systems. IEEE Transactions on Automatic

 !"#$%&'()*

 !"#$

 !"#$%!&"'



 !"#$%&'()*+,-./0123 45!"#$6789:67-.;<=+,12>?@6

 !"#$%&'()*+,-.'/01
23456+,-789:;<=>?@ABCD+EFGHIJ

 !"#$%&'()*+,-.-$/012
3456789:
$12;< !=

− 788 −