ARTICLE IN PRESS

Control Engineering Practice 13 (2005) 519531

Robust decentralized parameter identication for two-input

two-output process from closed-loop step responses
Shao-Yuan Lia,*, Wen-Jian Caib, Hua Meia, Qiang Xiongb
a

Department of Automation, Institute of Automation, Shanghai Jiao Tong University, 1954 Hua-shan Rd, Shanghai 200030, PR China
b
School of EEE, Nanyang Technological University, Singapore. 639798, Singapore
Received 14 April 2003; accepted 26 April 2004
Available online 19 June 2004

Abstract
In this paper, a novel parameter identication method for closed-loop two-input two-output (TITO) processes from step-test is
proposed. Through sequential step change of set points, the coupled closed-loop TITO system is decoupled equivalently into four
independent single open-loop processes with same input signal acting on the four transfer functions. Consequently, existing
identication methods for single-loop process can be extended to TITO systems and the parameters of rst- or second-order plus
dead-time models for each transfer function can be directly obtained by using the linear regression equations derived for the
decoupled identication system. The proposed method is simple for engineering application and robust in the presence of large
amounts of measurement noise. Simulation examples are given to show both effectiveness and practicality of the identication
method for a wide range of multivariable processes.
r 2004 Elsevier Ltd. All rights reserved.
Keywords: Identication; Two-input two-output process; Step test; Decoupled identication system; Least squares methods

1. Introduction
In multivariable process control, most schemes such as inverse Nyquist array or characteristic locus methods Astrom
and Hagglund, 1984, 1995 require a full model of the process in the form of a transfer-function matrix or a frequencyresponse matrix over the entire working frequency range. In many cases, such a model is not available and physical
modeling may require a prohibitive engineering effort. Therefore, practical and effective estimation of the full process
models becomes appealing and has been an active research area of control engineering for a few decades. A
considerable number of identication methods and their application to other engineering elds including advanced
control strategy, optimization and signal processing have been reported in the literature (Loh & Vasnani, 1994; Poulin,
Pomerleau, Desbiens, & Hodouin, 1996; Wang & Cai, 2003; Zhu & Butoyi, 2002).
There are two ways to identify a multivariable process for control application, i.e. open-loop and closed-loop ones.
In any case, an excitation of the process is needed to extract useful information on process dynamics. For open-loop
transient response experiments, step or pulse excitation signals are commonly injected at the process inputs, and the
response is measured (Choi, Lee, Jung, & Lee, et al., 2000; Young, 1970). The main advantage of the step test is that
the testing procedure is simple and requires little prior knowledge. However, it is quite sensitive to non-linearity in the
system (Luyben, 1991). For closed-loop identication, a majority of existing techniques are in the frequency domain
while the frequency range of interest for such applications is usually from zero up to the process critical frequency
(Loh, Hang, Quek, & Vasnani, 1993; Shen, Wu & Yu, 1996; Wang & Shao, 1999; Wang & Cai, 2003). Since the closedloop testing causes less perturbation to the process, it is preferred to open-loop one in process control practice.

*Corresponding author. Tel./fax.: +86-21-62932114.

E-mail address: [email protected] (S.-Y. Li).
0967-0661/$ - see front matter r 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2004.04.017

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

520

As the step test is the simplest and dominant in process control applications, there has been strong research interest
in using such a test to determine the dynamics of unknown processes. Bi, Cai, Lee, and Wang, et al., 1999 proposed a
simple yet robust identication method to obtain a rst-order plus dead-time model for a linear monotonic process
from a step test of open-loop control systems. The identication method was later applied to design auto-tuning PID
controllers for heating, ventilation and air-conditioning (HVAC) systems (Bi, Cai, Wang, & Hang, et al., 2000),
experimental results have demonstrated the effectiveness of the technique. Wang and Cluett 1994 presented an
identication algorithm for processes operating in closed-loop, the algorithm involves tting two Laguerre models
directly to the control signal and the process output signal generated by a step change in the set-point. Wang, Guo and
Zhang (2001); Wang and Zhang (2001) proposed some robust identication methods for linear time-delay processes
from step responses in both time domain and frequency domain. These results was also applied to PID controller autotuning for multivariable processes Wang, Huang and Guo (2000).
In this paper, an engineering oriented identication technique for multivariable process is proposed, which extend
SISO identication method in Bi, Cai, Lee, and Wang, et al., (1999) to the multivariable systems. The proposed
method only requires step response data of closed-loop process, and no prior knowledge of the process dynamics and
of the controller dynamics is needed. Through sequential step change of set points, the coupled closed-loop TITO
system is decoupled equivalently into four individual single open-loop processes with same input signal acting on the
four transfer functions. Consequently, existing identication methods for single-loop process can be extended to TITO
systems and the parameters of rst- or second-order plus dead-time models for each transfer function can be directly
obtained by using the linear regression equations derived for the decoupled identication system. The proposed
method is simple for engineering application and robust in the presence of large amounts of measurement noise.
Various typical multivariable processes have been employed to illustrate the effectiveness of the method. It offers a
good engineering tool for control engineers in retuning an existing multivariable control system and designing
advanced controllers for multivariable processes.

2. Formulation of decentralized TITO Identication systems

Consider a TITO process under decentralized control as shown in Fig. 1, where ri ; ei ; ui and yi ; i; j 1; 2 are set
points, errors, controllers and process outputs, Ki ; zi and Gij controllers, noises and process transfer functions,
respectively. In general, K1 and K2 could be any type of controllers that make the closed loop system stable. To simplify
our derivation, the notation ri ; ei ; yi i; j 1; 2; are used in both s and t domain. The fundamental relationship between
error signals and transfer function outputs for the system are expressed as:
y1 G11 K1 e1 G12 K2 e2 ;
y2 G21 K1 e1 G22 K2 e2 ;
e 1 r 1  y1 ;
e 2 r 2  y2 :

r1 +

e
1

1
K1

u1

G11

y1
+

Controller 1

G21
G12
Controller 2

r2 +

e2

K2

u2

G22

Fig. 1. Closed-loop TITO control system

y2

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

521

Fig. 2. The settling time ts parameter identication.

Assume that the process initially rests at a steady state with initial set points, errors and output variables as
r01 ; r02 ; y01 ; y02 ; e01 and e02 ; respectively, with
e01 r01  y01 ;
e02 r02  y02 :

To identify the process parameters, the test involves the following two steps:
1. Make a step change in r1 from r01 to r11 ; with r2 kept unchanged, record the error signals for the two loops, until the
rst steady state is reached at t9t1 T > ts (see Fig. 2), ts is the maximum settling time of all transfer functions.
The incremental equation from initial state to the rst state becomes:
Dy11 G11 K1 De11 G12 K2 De12 ;

3a

Dy12 G21 K1 De11 G22 K2 De12 ;

3b

where
De11 r11  r01  y11  y01 Dr1  Dy11 ;

3c

De12 y12  y02 Dy12 :

3d

2. Make a step change in r2 from r02 to r12 ; while keeping r1 r11 as before, record the error signals for the two loops,
until the second steady state is reached at t9t2 T > ts : Again, the incremental equation from the rst state to the
second state can be written as
Dy21 G11 K1 De21 G12 K2 De22 ;

4a

Dy22 G21 K1 De21 G22 K2 De22 ;

4b

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

522

where
De21 y21  y11 Dy21 ;

4c

De22 r22  r02  y22  y12 Dr2  Dy22 ;

4d

Combine (3a), (3b), (4a) and (4b) into matrix form

Y AX ;

5a

where
2

Dy11

K1 De11

6 Dy1 7
6 0
6
7
6
Y 96 22 7; A96
4 Dy1 5
4 K1 De21
Dy22

K2 De12

0
K2 De22

K1 De11
0

K2 De12 7
7
7
0 5

K1 De21

K2 De22

G11

6G 7
6 12 7
and X 96
7:
4 G21 5

5b

G22

Because of detA K1 K2 2 De11 De22  De21 De12 2 ; thus the matrix A is nonsingular if
De11 De22  De21 De12 a0:
From (3ad) and (4ad), it can be proved that
 


K1 0
1 0
G11
De11 De22  De21 De12 det

0 1
0 K2 G21

G12
G22

detI KG:

For a BIBO stable system, detI KGc0 8s in Laplace domain. Therefore, A is usually nonsingular and X can
be solved by
X A1 Y :

Substituting (3cd), (4cd) and (5b) into (6) with Duji Ki Deji ; i; j 1; 2; we obtain
G11 s

Dy11  Du22  Dy21  Du12 y11 s

;
9
us
Du11 Du22  Du21 Du12

G12 s

Dy21  Du11  Dy11  Du21 y21 s

;
9
us
Du11 Du22  Du21 Du12

G21 s

Dy12  Du22  Dy22  Du12 y12 s

;
9
us
Du11 Du22  Du21 Du12

G22 s

Dy22  Du11  Dy12  Du21 y22 s

:
9
us
Du11 Du22  Du21 Du12

In terms of parameter identication, the coupled closed-loop TITO system has been decoupled into four
individual single open-loop systems with same input signal acting on the four transfer functions.
Consequently, the problem of identication of coupled closed-loop TITO system is transformed into the
identication of multi-single open-loop problem. The relation between the original system input/output and the
decentralized identication system are given as
 The system input u in Laplace domain for all of the four loops is
us Du11 sDu22 s  Du21 sDu12 s

and that in time domain is

ut Du11 tDu22 t  Du21 tDu12 t:

 Corresponding to the input u given in (8), the output signals yij both in Laplace domain and in time domain
respectively for the equivalent identication system are
1.
y s Dy1 s  Du2 s  Dy2 s  Du1 s;
11

y11 t L1 Dy11 s  Du22 s  Dy21 s  Du12 s

Dy11 tDu22 t  Dy21 tDu12 t:

10a

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

2.

523

y21 s Dy12 s  Du22 s  Dy22 s  Du12 s;

y21 t L1 Dy12 s  Du22 s  Dy22 s  Du12 s

Dy12 tDu22 t  Dy22 tDu12 t:

3.

10b

y12 s Dy21 s  Du11 s  Dy11 s  Du21 s;

y12 t L1 Dy21 s  Du11 s  Dy11 s  Du21 s

Dy21 tDu11 t  Dy11 tDu21 t:

4.

10c

y22 s Dy22 s  Du11 s  Dy12 s  Du21 s;

y22 t L1 Dy22 s  Du11 s  Dy12 s  Du21 s

Dy22 tDu11 t  Dy12 tDu21 t:

In such above formulas, the operator

10d

 means a convolution operation.

Remark 1. During the testing, step change for each set point is performed sequentially as shown in Fig. 2, where ts is
defined as the maximum time required for all errors to settle at steady state under the two testing.
Remark 2. By using incremental signals, zero initial conditions for the equivalent decentralized identification system are
guaranteed, that is ut yij t 0; tp0, for i; j 1; 2, regardless the value of initial steady state errors.

3. Least squares method

For the decentralized identication systems, the forward transfer function can be represented by
Gij s

bij1 sn1 bij2 sn2 ? bijn Lij s

e
;
aij1 sn1 aij2 sn2 ?aijn

i; j 1; 2:

sn

11

Following Remark 2, initial conditions for the equivalent decentralized identication systems are zeros. Therefore,
Eq. (11) can be written equivalently in differential equation form as
n1
yn
t ? aijn yij t bij1 un1 t  Lij bij2 un2 t  Lij ? bijn ut  Lij :
ij t aij1 yij

Dene
Z m

f t9

0;t

Z tZ

Z t2
?
mf t1 dt1 ?dtm ;
0
0
0
|
{z
}

12

tm

tXmaxt1 ; t2 :

For an integer mX1; Eq. (12) can be solved by integrating n times:

Z 1
Z 2
Z n1
Z
yij t aij1
yij t aij2
yij t ? aijn1
yij t aijn
bij1

0;t

0;t

ut  Lij bij2

0;t

0;t

ut  Lij ? bijn1

0;t

yij t

0;t

n1

ut  Lij bijn

0;t

ut  Lij :

13

0;t

Using rst-order Taylor expansion for the unknown time delays, i.e., eLij s 61  Lij s; Eq. (13) becomes
Z 1
Z 2
Z n1
Z n
yij t aij1
yij t aij2
yij t ? aijn1
yij t aijn
yij t
bij1

0;t

0;t

ut bij2

0;t

ut ? bijn1

0;t

 Lij bij1 ut  Lij bij2  ?  Lij bijn1

0;t

n1

ut bijn

0;t

n2

ut  Lij bijn
0;t

0;t

ut
0;t

n1

ut:
0;t

14

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

524

In process control applications, the rst- or second- order plus time delay models are very popular since they can be
used to represent both monotonic, oscillatory and non-minimum phase processes. The solutions for n=1 and 2 under
the framework of Eq. (14) are given as follows:
1. n 1; Eq. (14) becomes
Z t
Z t
yij t aij1
yij t dt  Lij bij1
ut dt bij1 ut
0

15

which can be written into the compact form

g1ij t f1ij t
T  y1ij ;

16

where
g1ij t yij t;
h R
i
R
f 1ij t
T  0t yij t dt  0t ut dt ut ;


y1ij aij1 Lij bij1 bij1 :

17

2. n 2; Eq. (14) becomes

Z t
Z tZ t
Z t
Z tZ t
yij t dt  aij2
yij t1 dt1 dt bij1
ut dt bij2
uij t1 dt1 dt
yij t  aij1
0
0
0
0
0
0
Z t
ut dt:
 Lij bij1 ut  Lij bij2

18

It is again expressed as
g2ij t f2ij t
T y2ij ;

19

where
g2ij t yij t;
 Z t
Z tZ t
Z tZ t
Z t
yij t dt 
yt1 dt1 dt
ut dt
ut1 dt1 dt  ut
f 2ij t
T 
0
0
0
0
0
0


y2ij aij1 aij2 bij1 bij2 Lij bij1 Lij bij2 :


ut dt ;

20

Eq. (15) and (18) can be solved by the least squares methods for each transfer functions i; j 1; 2; and k 1; 2 to
form the regression form
Gij Cij Yij ;

21

where Gij gkij t1 ; gkij t2 ; ?

for Yij are

1
* ij CT Cij
Y
CTij Gij :
ij

gkij tN
T ; Cij fkij t1 ;

fkij t2 ;

? fkij tN
T : Its least squares estimation

* ij are found from Eq. (22), aij1 ; bij1 and Lij can be recovered from
Once Y
3
2
3 2
y1ij1
aij1
7
6
7 6
1
7
4 bij1 5 6
4 yij3 5
Lij
y1 =y1
ij2

22

23

ij3

for n 1; and aij1 ; aij2 ; bij1 ; bij2 and Lij from

3
2
3
2
y2ij1
aij1
7
6
7
6 a 7 6 y2
ij2
7
6 ij2 7 6
7
7 6
6
7
6 bij1 7 6 y2
ij3
7
7 6
6
7
7 6
6
4 bij2 5 6 y2ij4 7
5
4
Lij
2
2
yij5 =yij3
for n 2; respectively.

24

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

525

Remark 3. To integrate the equivalent input ut, the first-order Taylor series expansion is adopted for the time delay i.e.
eLij s 61  Lij s: Even though it showed very accurate results in our simulation, this approximation may not be sufficient in
some cases. More accurate Lij can be obtained by either using higher-order Taylor series expansion or frequency domain
identification approach.
Remark 4. In the noise-free testing environment, the least squares method (22) yields very good results. Now suppose that
there exist white noises z1 and z2 as shown in Fig. 1, then the equivalent input and outputs are rewritten as
# zu ;
ut ut
y11 t y# 11 t zy11 ;
y12 t y# 12 t zy12 ;
y21 t y# 21 t zy21 ;
y22 t y# 22 t zy22 :
# and y# ij ; i; j 1; 2; are input/output signals; zu ; zyij are colored noises, respectively, which are functions of z1
Here, ut
* ij will deviate from Yij : A common method is to pre-treat ut and
and z2 : Due to the existence of these colored noises, Y
yij t through some low-pass lters. In the proposed method, we suggest integrating (14) in the both sides, which can
eliminate stochastic noise effectively.

4. Simulation example
The Wood and Berry binary distillation column plant (Wang, Guo & Zhang 2001c) is a typical TITO process with
strong interaction and signicant time delays, which transfer function matrix is given as
2
3
12:8es 18:9e3s
6 1 16:7s
1 21s 7
7:
Gs 6
4 6:6e7s
19:4e3s 5
1 10:9s 1 14:4s
For the decentralized closed-loop system with KP1 0:5271; KI1 0:0763; KD1 0:45; and KP2 0:1064; KI2
0:018 KD2 0:02; the step-test error signals without noise and the differentials of the equivalent signals u and yij
(i; j 1; 2) are shown in Figs. 3 and 4, respectively.
In practice, the real measurement of the process output under closed-loop test are inevitably corrupted by
measurement noise, which leads to the corruption of the constructed test signals. To show the effectiveness of the
proposed approach under measurement noise conditions, we dene the noise-to-signal ratio as
meanabsnoise
NSR
:
25
meanabssignal
In order to show the effectiveness of the proposed approach by simulations, a criterion in frequency domain is
dened as
(
)

#
Gjo


Gjo
  100% ;
E max 
26

oA0;oc

Gjo
#
where Gjo and Gjo
are the actual and estimated process frequency response, respectively, and +Gjoc p:
The identied FOPDT models and parameters under difference noise level are given in Table 1. Fig. 5 shows the
Nyquist curves of identied models under different NSR.

5. Application to multistage gasliquid absorption column system

Consider a multistage gasliquid absorption column shown in schematic form in Fig. 6 Giovani, Serkan and Oscar,
2003. The column features n trays where the transfer of a component of interest from the gas stream into the liquid
stream is accomplished in each tray as the rising gas stream bubbles through a liquid layer that ows horizontally
across the tray. The variable x and y; respectively, denote the composition of the species of interest in the liquid and gas
phases. The manipulated variables are the compositions xf and yf of the feed streams, and the measured process
variables are the compositions xp and yp of the product streams.

ARTICLE IN PRESS
526

S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

Fig. 3. System step tests without noises.

Fig. 4. Equivalent signals for the Wood & Berry model.

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

527

Table 1
Identication results under different noise level for example 1
Noise level (%)
0

#
Gs

E(%)
G# 11 s

12:7929 1:0044s
e
16:6193s 1

G# 12 s

18:90 3:01s
e
21:001s 1

G# 21 s

6:5998
e6:9852s
10:9205s 1

19:3942 3:006s
e
G# 22 s
14:4339s 1
10

G# 11 s

12:7782 1:1848s
e
16:6079s 1

18:7041 2:8636s
e
G# 12 s
20:6871s 1
G# 21 s

6:5809
e7:1670s
10:7352s 1

19:2724 2:9083s
e
G# 22 s
14:2769s 1
20

G# 11 s

12:6057 1:4046s
e
16:3221s 1

18:2795 2:7695s
e
G# 12 s
19:9567s 1
G# 21 s

6:5443
e7:3482s
10:4986s 1

19:0350 2:8450s
e
G# 22 s
13:9227s 1

0.3573

0.0255

0.2749

0.6187

4.2731

2.4054

2.6477

2.5753

9.4847

8.1978

6.6869

7.6368

A complete process model has been developed in the literature (Bequette, 1998) under the assumptions that (i) only
the component of interest is transferred from one phase to the other, (ii) the vapor stream leaving a stage is in
thermodynamic equilibrium with the liquid on that stage, and (iii) the liquid holdup M in each tray is very large
compared to the vapor holdup, which implies that the feed and product ow rates are at steady state. Furthermore, the
thermodynamic equilibrium in the ith stage is represented by the expression yi axi ; where a is an equilibrium
parameter, and mass balances for stages i 2 to n  1 yield the modeling equations
dxi
L
L Va
Va
xi xi1 ; i 1; 2; y; n  1:
xi1 
27
M
M
M
dt
In turn, the feed tray (i 1) and the product tray (i n) are, respectively, modeled according to the mass balances
dx1
L Va
Va
L
x1 x2  xf

28
M
M
M
dt

ARTICLE IN PRESS
528

S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

Fig. 5. Nyquist curves under different NSR.

Liquid
Feed

yp

Gas
Product

1
2

.
.
.
n-1
n

Liquid
Product

yf

Gas
Feed

Fig. 6. Schematic of the gas absorption column with n equilibrium stages.

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

529

and
dxn
L
L Va
V
xn y f :
xn1 
M
M
M
dt

29

Finally, the output equations are

xp xn ;

30

yp ax1 :

31

For the particular case of a ve-tray

space representation is
2
0:325
0:125
2
3
dx1 =dt
6
0:200
0:325
6 dx =dt 7 6
6 2
7 6
6
0
0:200
6
7
4 dx3 =dt 5 6
6
0
0
4
dx4 =dt
0
0


xp t
yp t

0:5

column, model (27)(31) reduces to the two-input/two-output, whose state0

0:125

0:325
0:200

0:125
0:325

0
2 3
x1
6 7
6 x2 7
1 6 7
6 x3 7;
7
0 6
6 7
4 x4 5
x5

0:200

32

x1

76 7 6
76 x2 7 6
76 7 6
76 x3 7 6
0
76 7 6
76 7 6
0:125 54 x4 5 4
0:325
x5
0

0:200

0
0

0
0

0:250

3
7
7

7 xf t
7
7 y t ;
7 f
5

Table 2
Identication results under different noise level for example 1
Noise level (%)
0

#
Gs

e
G# 11 s

G# 12 s

G# 21 s

G# 22 s

20

388:9s2

143s2

1
e5:13s
68:42s 2:508

12:92s 1
46:47s 2:079

57:19s2

12:92s 1
18:59s 0:8318

4028s2

1
e5:55s
715:4s 26:3

G# 11 s

1
e4:12s
451s2 71:79s 2:506

G# 12 s

13:92s 1
151:8s2 48:77s 2:079

G# 21 s

14:16s 1
60:87s2 19:75s 0:8314

G# 22 s

1
e5:14s
4148s2 772:9s 26:12

1.1291e-6

1.5949e-6

6.0621e-6

3.9266e-8

9.1318e-6

4.599e-6

2.7462e-5

5.8458e-7

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

530

Fig. 7. Nyquist curves of 5 trays gas absorption column under different NSR.

where the system matrices are obtained after substituting in Eqs. (27)(31). The equilibrium parameter a 0:5 and the
following steady-state values of process parameters: L=4/3 kg mol/min, V=5/3 kg mol/min, M=20/3 kg mol/min, the
steady-state input xf ;ss 0 and yf ;ss 0:1:
By selecting K1 1 and K2 1 for the closed-loop system, the obtained identication model under 0% NSR and
20% NSR respectively are given in Table 2, where a time domain identication error is dened as
e

N
1X
2
#
ykTs  ykT
s
;
N k1

32

#
where ykTs is the actual process output under a step change, while ykT
s is the response of the estimated process
under the same step change. Fig. 7 gives the Nyquist curves of those models under different NSR.

6. Conclusions
In this paper, a novel identication method based on step tests for multivariable process has been presented. By
decoupling the closed-loop multivariable system into independent single open-loop systems with same input signal
acting on the multiple transfer functions, most restrictions of existing multivariable process identication methods has
been relaxed. The proposed method requires no prior knowledge of the process and the rst-order or second-order plus
delay model elements of the transfer function matrices can be directly obtained. The main advantages of the proposed
methods are: (1) The testing procedure is straightforward, the computation is simple, and can be easily implemented;
(2) It is valid for closed-loop tests; and (3) It can identify both close-coupled and weak-coupled processes. Various
processes have been employed to demonstrate the effectiveness and practicability of the method. Simulation results
show that the proposed identication method is practical and accurate even in the noisy environment, especially in the
frequency range from zero to the frequency corresponding to p phase lag. It also showed that the time delays can be
accurately approximated by rst-order Taylor series, however, more accurate approximations are expected by
frequency domain identication approach which is currently under investigation. Even though the method is derived
for TITO systems, the extension to general MIMO system is straightforward.

ARTICLE IN PRESS
S.-Y. Li et al. / Control Engineering Practice 13 (2005) 519531

531

Acknowledgements
The authors would like to acknowledge the nancial support of the High Technology Research and Development
Program of China (Grant No.2002AA412130) and the Specialized Research Fund for the Doctoral Program of Higher
Education of China (Grant No.: 20020248028). The authors are grateful to the anonymous reviewers for their valuable
recommendations.

References
Astrom, K. J., & Hagglund, T. (1984). Automatic tuning of simple regulators with specications on phase and amplitude margins. Automatica, 20(5),
645651.
Astrom, K. J., & Hagglund, T. (1995). Pid controllers: Theory, design, and tuning (2nd ed). Research Triangle Park, NC: Instrument Society of
America.
Bequette, W. B. (1998). Process dynamics modeling analysis and simulation Vol. 42(5) (pp 1293-1303). Englewood Cliffs, NJ: Prentice-Hall.
Bi, Q., Cai, W. J., Lee, E. L., & Wang, Q. G., et al. (1999). Robust identication for rst-order plus dead-time model from step response. Control
Engineering Practice, 7, 7177.
Bi, Q., Cai, W. J., Wang, Q. G., & Hang, C. C., et al. (2000). Advanced controller auto-tuning and its application in HVAC systems. Control
Engineering Practice, 8(6), 633644.
Choi, J. Y., Lee, J., Jung, J. H., & Lee, M., et al. (2000). Sequential loop closing identication of multivariable process models. Computers and
Chemical Engineering, 24, 809814.
Giovani, C. N., Serkan, K., & Oscar, D. C. (2003). A polynomial perspective on the stability of multivariable predictive controllers. Computer &
Chemical Engineering, 27, 10971111.
Loh, A. P., Hang, C. C., Quek, C. K., & Vasnani, V. U. (1993). Autotuning of multiloop proportional-integral controllers using relay feedback.
Industrial Engineering and Chemical Research, 32, 11021107.
Loh, A. P., & Vasnani, V. U. (1994). Necessary conditions for limit cycles in multiloop relay systems. IEE Proceedings on D Control Theory and
Application, 141(3), 163168.
Luyben, W. L. (1991). Process modeling, simulation and control for chemical engineers. New York: McGraw-Hill.
Poulin, E., Pomerleau, A., Desbiens, A., & Hodouin, D. (1996). Development and evaluation of an auto-tuning and adaptive PID controller.
Automatica, 32(1), 7182.
Shen, S. H., Wu, J. S., & Yu, C. C. (1996). Use of biased-relay feedback for system identication. A.I.Ch.E. Journal, 42(4), 11741180.
Wang, L., & Cluett, W. R. (1994). System identication based on closed-loop step response data. IEE Proceedings on the Control Theory and
Application, 141(2), 107110.
Wang, Q. G., Huang, B., & Guo, X. (2000). Auto-tuning of TITO decoupling controllers from step tests. ISA Transactions, 39, 407418.
Wang, Q. G., & Zhang, Y. (2001). Robust identication of continuous systems with dead-time from step responses. Automatica, 37, 377390.
Wang, Q. G., Guo, X., & Zhang, Y. (2001). Direct identication of continuous time delay systems from step responses. Journal of Process Control,
11, 531542.
Wang, Y. G., & Shao, H. H. (1999). PID autotuner based on gain and phase margin specications. Industrial Engineering and Chemical Research,
38(8), 30073012.
Wang, Y. G., & Cai, W. J. (2003). Decentralized relay-based multivariable process identication in frequency domain. IEEE Transactions on
Automatic Control, 48(5), 873878.
Young, P. C. (1970). An instrumental variable method for real time identication of a noisy process. Automatica, 6, 271287.
Zhu, Y. C., & Butoyi, F. (2002). Case studies on closed-loop identication for MPC. Control Engineering Practice, 10, 403417.