You are on page 1of 13

Article

pubs.acs.org/IECR

Controller Design for MIMO Processes Based on Simple Decoupled


Equivalent Transfer Functions and Simplied Decoupler
C. Rajapandiyan and M. Chidambaram*
Department of Chemical Engineering, Indian Institute of Technology Madras, Chennai 600 036, India
ABSTRACT: A method for the independent design of proportional-integral/proportional-integral derivative (PI/PID)
controllers is proposed based on the equivalent transfer function (ETF) model of the individual loops and the simplied
decoupler matrix. It is shown that the conventional eective open-loop transfer function (EOTF, derived from the dynamic
relative gain array (DRGA)) is equivalent to the ETF (derived from the relative normalized gain array (RNGA) and relative
average residence time array (RARTA)). This relation is used to approximate the decoupled process models as ETF models. The
simplied decoupler is shown to decompose the multiloop systems into independent loops (multi-single loop systems) with the
ETFs as the resulting decoupled process model. The concept of the ETF (perfect control approximation) is validated by
introducing the decoupler. Based on the corresponding ETFs, the decentralized PI controllers are designed using the simplied
internal model control (SIMC) method. Three simulation examples of multi-input multi-output (MIMO) process models are
considered to demonstrate the simplicity and eectiveness of the proposed method. The performance of the proposed control
system is compared with the ideal, normalized, inverted decoupling, and centralized control systems.

1. INTRODUCTION
Most industrial processes are multi-input multi-output
(MIMO) systems. Controller design for MIMO processes is
dicult compared to that of the single-input single-output
(SISO) processes, due to the interactions between the input/
output variables. Since the control loops interact with each
other, the tuning of one loop cannot be carried out
independently. The MIMO process can be controlled by
decentralized (multiloop) controllers, or decoupled controllers,
or centralized controllers. In multiloop control, the MIMO
processes are treated as a collection of multi-single loops. The
controller is designed and implemented on each loop by
considering the loop interactions. Multiloop controllers have
been widely used due to their reasonable performances,
simplicity, and robustness. Many design methods are reported
in the literature such as detuning method,1 sequential loop
closing method,2 relay auto tuning method,3 and independent
design method.4 Decentralized controllers work well when the
interactions among the loops are modest. If the interactions are
signicant, centralized controllers are desirable. It is dicult to
design a controller for each loop independently.5 The number
of tuning parameters of proportional-integral derivative (PID)
controllers for centralized control system is 3n2, where n is the
number of inputs. The next approach5 is to design the
decoupler with a decentralized control system. This method
allows us to use SISO controller design methods and the
number of tuning parameters (for PID controller), in this case
is 3n.
There are three types of basic decoupling techniques
available,6 ideal, simplied, and inverted decoupling methods.
The ideal decoupling scheme needs to calculate the inverse of
the process transfer function matrix. It may result in too
complicated decoupling elements, and the ideal decoupling is
sensitive to modeling errors.7 The simplied decoupling
technique has a simple decoupler form, but the controller
cannot be designed directly from the decoupled process model
2012 American Chemical Society

without introducing the model reduction technique. Weischedel and McAvoy8 have compared the ideal and simplied
decoupling control techniques, and they reported that the
simplied decoupling is more robust than the ideal decoupling.
The inverted decoupling method has the advantages of both
the ideal and simplied decoupling techniques. The implementation of the higher dimension system could result in to
physical nonrealizability,7,9 and this technique is more sensitive
to modeling errors.7,8,10 Gagnon et al.7 and Wade11 discussed
the implementation issues in the inverted decoupling method.
Tavakoli et al.,5 Kumar et al.,12 and Maghade and Patre13
have used the simplied decoupler plus a decentralized
controller. First, the simplied decoupler matrix is introduced
in to the process model to decouple the multiloop process into
equivalent independent single loops. Second, the resultant
decoupled process model is approximated as a rst order plus
time delay (FOPTD) or second order plus time delay
(SOPTD) model using a suitable model reduction technique
(Maclaurin series,4 open-loop transient response plot, graphical
approach12 and frequency response tting13). Third, the
proportional-integral (PI)/PID controllers are designed independently based on the corresponding reduced decoupled
process model. Recently, several researchers have introduced
the concept of equivalent transfer functions/eective open-loop
transfer functions (ETFs/EOTFs)/eective open-loop process
(EOP) to take into account the loop interactions in the design
of multiloop control systems.14,15,1820 Vu and Lee4 have
proposed an independent design method for the design of
multiloop controllers. Based on the assumption of perfect
controllers (other loop), the EOTF is rst derived to
Received:
Revised:
Accepted:
Published:
12398

March 17, 2012


July 5, 2012
August 27, 2012
August 28, 2012
dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

equivalent to the ETF (derived from RNGA and RARTA).


This relation is used to approximate the decoupled process
model (by the introduction of simplied decoupler matrix) as
the ETF model and the SISO controller is designed based on
the corresponding ETF. The main advantage of this proposed
method compared to the conventional simplied decoupling
approach5,12,13 is not requiring the model reduction steps (to
reduce the decoupled process model). The formulation of
EOTFs for the higher dimension systems is dicult. Further,
the model reduction methods also pose problems for higher
dimension systems. On other hand, the derivation of ETFs is
easy for higher dimension systems. Cai et al.6 have calculated
the decoupler matrix by choosing the decoupled process model
parameters (g*) as

decompose multiloop control system in to a set of equivalent


independent single loops. By using the model reduction
technique (Maclaurin series), the EOTF model is reduced to
a FOPDT/SOPTD model. Based on the reduced EOTF model,
the decentralized controllers are designed independently.
Cai et al.6 have proposed a normalized decoupling control for
the TITO process. The ETF matrix is derived from the
concepts of relative gain array (RGA), relative normalized gain
array (RNGA), and relative average residence time array
(RARTA). The relationship between the ETF and process
transfer function matrices is derived. For deriving the decoupler
matrix, the ETF matrix is used instead of the inverse of process
transfer function. On the basis of the ETF matrix, the criteria
for determining stable, proper, and casual ideal diagonal
decouplers are established. In the performance of the
decoupled control system, it is found that the interaction is a
considerable one and the response is sluggish. Choosing of the
decoupled process model involves some ad-hoc procedures.
Huang et al.14 formulated EOPs without the prior knowledge of
controller dynamics of other loops and that the controllers are
independently designed based on the derived EOPs for the
corresponding equivalent single loops. Xiong et al.15 have
derived the eective transfer function in terms of eective
relative gain array for the multivariable system. This eective
transfer function model incorporates the interactions among
the loops. The decentralized control system is designed by
using the single loop design approaches based on the derived
eective transfer function. The eective transfer function
provides both gain and phase information required for the
design of the decentralized controllers.
He et al.16 have used the relative normalized gain array
(RNGA) concept to determine the interaction measurements
and loop pairing. This RNGA includes both the steady state
and the transient information of the transfer function matrix.
Naini et al.17 introduced eective relative energy array (EREA)
concept. EREA is utilized to propose a new loop pairing
method for the MIMO systems. Shen et al.18 obtained a full
controller matrix by the independent design approach.
Controllers for the selected loops are designed independently
based on the gain and phase margin (GPM) method. The
eective transfer function is obtained by utilizing the RNGA
and the relative average residence time array (RARTA)
concepts.
Hu et al.19 have proposed a decentralized controller design
method for the integrating MIMO systems. In this method,
based on the ETF, the decentralized controller is designed by
the simple internal model control (SIMC) method. Recently,
Kumar et al.20 have proposed two methods of designing
centralized control system for multi-input, multi-output
(MIMO) processes. Centralized PI controllers are designed
based on the direct synthesis approach. In the direct synthesis
method, the inverse of the process transfer function matrix is
approximated by the ETF matrix. In Kumar et al.'s paper,20 it is
shown that the centralized control system works better for the
systems having RGA less than 1. From the literature, it is
observed that the ETF matrix is mainly used to nd the inverse
of the process model6,18,20 and the perfect control approximation is used to incorporate the loop interactions of
multivariable systems in the form of ERGA/EOTF/ETF.
In the present work, we combine the simplied decoupler
approach5,12,13 with the ETF model approximations6,18,20 in
order to get the advantages of both the methods. The
conventional EOTF (derived from DRGA) is shown to be

T
D(s) = G1(s)G*(s) = G (s)G*(s)

whereas, in the present method, there is no need to assume the


expression for g*; instead, one can take the ETFs as g*. In the
proposed method, the method of designing decoupling control
system consists of three steps. First, the multiloop process is
decoupled in to independent loops (multi-single loop system)
by using the simplied decoupler matrix. Next, the resulting
decoupled process model is approximated using the concept of
ETF. Based on the corresponding ETFs, the decentralized PI/
PID controllers are designed using the simplied internal
model control (SIMC) method. Three simulation examples of
MIMO industrial process models are considered to show the
simplicity and eectiveness of the proposed method. The
objective of the present work is to show the performance
improvement and robustness of the combination of simplied
decoupler matrix and ETF based controller is better than the
existing methods (ideal and normalized decoupling, and
centralized controllers).

2. EFFECTIVE OPEN-LOOP TRANSFER FUNCTION


(EOTF)
Consider the two-input two-output (TITO) systems with the
decentralized control system as shown in Figure 1. G(s) and

Figure 1. Simplied decentralized control system of a TITO process.

Gc(s) are process transfer function matrix and decentralized


controller matrix of TITO systems respectively and are
represented as
g11 g12

G (s ) =
g21 g22
(1)
12399

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research


g
0
c11

Gc(s) =
0 gc22

Article
eff
g22
=

(2)

K pij eijs
(ijs + 1)

i = 1, 2

j = 1, 2
(3)

This input-output relation can be expressed as


Y (s ) = G (s )U (s )
y (s )
1

Y (s ) =
y (s)
2

(4)

u1(s)

U(s) =
u 2(s)

where Y(s) and U(s) are output and input vectors respectively.
The inputoutput relationship for the TITO system can be
written as21
(6)

y2 (s) = g21(s)u1 + g22(s)u 2(s)

(7)

u1

In the TITO system, when the second loop is closed, the


input from ui to yi has two transmission paths. The combination
of two transmission paths is considered as eective open-loop
dynamics. If the second feedback controller is in the automatic
mode, with yr2 = 0, then the overall closed-loop transfer
function between y1 and u1 is given by21
g g g
y1
= g11 12 21 c2
1 + gc2g22
u1

KNij =

kij

u1

= g11

ij

ij + ij

KN,11
KN =
KN,21

u2

= g22

(8)

g21g12(gc1g11)
g11(1 + gc1g11)

(12)

k11
k12

KN,12 11 + 11 12 + 12
=
KN,22 k 21
k 22

21 + 21 22 + 22

(13)

In these expressions, ij = ij + ij is the average residence


time and it represents the response speed of the controlled
variable yi to manipulated variable uj. The relative normalized
gain array is expressed using the normalized gain matrix, and
hence, the RNGA (denoted as ) can be obtained as

Similarly, for the second loop


y2

(11)

and normalized gain matrix is expressed as

g12g21(gc2g22)
g22(1 + gc2g22)

g11

kij

This can be written as


y1

g21g12

3. EQUIVALENT TRANSFER FUNCTION


Because most processes are open-loop stable and exhibit
nonoscillatory behavior for step inputs, the response is
simplied either by analytical or empirical methods to a
FOPDT model for designing the controllers. To describe the
dynamic properties of a transfer function, the normalized
gain6,16 (KNij) for a particular transfer function, gij(s), is dened
as

(5)

y1(s) = g11(s)u1 + g12(s)u 2(s)

= g22

eff
eff
Here, g11
and g22
are the eective open-loop transfer
4
functions (EOTF). These EOTFs are complicated transfer
function models, and it is dicult to directly use them for the
controller design. For the purpose of controller design,
resulting EOTFs are reduced to FOPTD models using
Maclaurin series.4 This method poses complications in higher
dimension systems, in the formulation of EOTFs and in the
model reduction step. In the present work, by using RGA,
RNGA, and RARTA concepts,6,16,1820 the expression for ETF
can be derived easily for higher dimension systems also.

The process transfer function models are expressed as FOPDT


models
gij(s) =

y2

= KN KNT

(9)

(14)

where

In MIMO systems, the open-loop dynamics between


controlled variable (yi) and manipulated variable (ui) not
only depend on the corresponding transfer function model (gii)
but also depend on the other processes and controllers in all
other loops. This implies that the tuning of one controller
cannot be done independently and it depends on the other
controllers. The complicated relations of eqs 8 and 9 can be
simplied by assuming two assumptions:4,19 First, the perfect
controller approximation for the other loop (the output attains
steady state with no transient) was used to simplify the eqs 8
and 9), that is,
gcigii
= 1 i = 1, 2
(1 + gcigii)


11
12

=
21 22

(15)

Relative average residence time6,16 (ij), which is dened as


the ratio of loop yiuj average residence times, when other
loops are closed and when other loops are open:
ij =

ij
ij

ij
ij

(16)

When the relative average residence times are calculated for


all the elements in the transfer function matrix, the results are in
array form, and it is called a relative average residence time
array6,16 (RARTA):

Second, ETFs have the same structure of the corresponding


open-loop model. By using the perfect controller approximation, eqs 8 and 9 can be approximated as follows:
g g
y
g11eff = 1 = g11 12 21
g22
u1
(10)

11 12
=

21 22
12400

(17)

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article


11
12
11 12
==
21 22 21 22

(18)

where is Hadamard division.


Using the denition of the relative average residence time, it
is possible to write6,16
ij =ijij
=ijij + ijij

(19)

=ij + ij

(20)

By using RGA, RNGA, and RARTA, it is possible that a


transfer function element of a MIMO process when other loops
are closed6,16 can be approximated by a transfer function
element having the same form as open-loop transfer function
element:
1

gij (s) = kij


eijs
ij s + 1
gij (s) =

Figure 2. Simplied decoupled control system of a TITO process.

(21)

kij

1
eijijs
ij ijijs + 1

(22)

3.1. Relationship between gii(s) and


The EOTF can
be expressed in terms of elements of dynamic relative gain array
(DRGA) as4
g
giieff = ii
ii
(23)
geff
ii .

/[
=

c, ij

c, ij

g* 0
11

G(s)D(s) =

*
0 g22

(28)

* 0
g11 g12 1 d12 g11

g
g

21 22 d 21 1 0 g *
22

(29)

d12(s) =

(yi /ui)d]all loops open

(yi /ui)d]all other loops closed except for loop y uj


i

d 21(s) =

gij(s)
gij (s)

DRGA is expressed as
(s ) = G (s )
g (s )
11
=
g (s )
21

G (s)
1/g (s) 1/g (s)
g12(s)
12

11

g21(s)
1/
g
(
s
)
1/
g
(
s
)

21
22

*=
g11
(24)

* =
g22

By comparing eqs 23 and 24, it is visible that


= gij (s)

(25)

g11(s)

(30)

g21(s)
g22(s)

(31)

y1
u1

y2
u2

= g11

= g22

g12g21
g22

(32)

g21g12
g11

(33)

By comparing eqs 10 and 11 to 32 and 33, the equations are


found to be equivalent with each other. Perfect controller
assumption is used for deriving eqs 10 and 11. The decoupler is
introduced for deriving eqs 32 and 33. The controllers designed
based on the EOTFs/ETFs give a better performance when
introducing the decoupler. g*11 and g*22 can be approximated
directly by ETF models. Both the formulations of EOTFs and
ETFs (to decompose individual loops) are based on the perfect
control approximations, and this assumption is fully validated
only by introducing the decouplers among the loops.
Wang et al.23 have suggested a method to obtain the
realizable decoupler, that is,

From eq 25, it is shown that the conventional EOTF (from


DRGA) is equivalent to the ETF (from RNGA).

4. DECOUPLING CONTROL DESIGN


Consider the two-input two-output systems with the decoupled
control as shown in Figure 2. The relationship between the
input vector and process output vector is given by
Y (s) = G(s)D(s)U (s)

g12(s)

In the presence of the decoupler, the TITO system behaves


like two independent loops for which the controllers can be
designed independently, and it is expressed as22

giieff

(27)

Decouplers can be designated as22

The dynamic relative gain array (DRGA) is dened as6


ij = [

y (s ) g g 1 d u ( s )
12
1
1 = 11 12

y (s) g21 g22 d 21 1 u (s)


2
2

(26)

For the two-input two-output (TITO) process22


12401

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

g (s )

12 ev(12 11)s
ev(22 21)s
g11(s)

D(s) =
g21(s) v( )s

v(11 12)s
21
22
e
e

g22(s)

2.2 1.3
K=
2.8 4.3
1.6254 0.6254
= K K T =

0.6254 1.6254
1.5537 0.5537
= KN KNT =
0.5537 1.5537

where

v( ) =

1,

if 0

0,

if < 0

8
7.3
Tar =

11.3 9.55
0.2750 0.1781
KN = K Tar =
0.2478 0.4503

23

In the Wang et al. method, the extra time delay is to be


incorporated to the decoupler matrix in the decoupler (in
nonrealizable cases). In such cases, the incorporated extra time
delay to the decoupler matrix will also change the
corresponding decoupled process model. Hence, the extra
time delay is to be added with the corresponding ETF model
and it will be used for controller design purposes.
In this present work, the controllers are designed based on
the corresponding ETFs of each individual loops using the
SIMC tuning method.24 This method is used here for its
simplicity and robustness. The SIMC method PI controller
forms as

1
gcii(s) = kc1 +

Is

0.9559 0.8853
==

0.8853 0.9559

By using the RGA and RNGA concepts, the ETF model


parameters are deduced by
1.3535 2.0786
K = K =

4.4769 2.6455
6.6910 6.1971
T = T =

8.4103 8.7940

0.9559 0.2656
L = L =

1.5935 0.3345

(34)

T = [ij ]

For the FOPTD SISO systems, the SIMC-PI24 settings are


given by
1

ii
kci(s) =

kii (ci + ii)

(35)

Ii = min(ii , 4(ci +ii))

(36)

L = [ij ]

The ETF model matrix is expressed as


1.3535 e0.9559s 2.0786 e0.2656s

(6.1971s + 1)
(6.6910s + 1)
G (s) =

1.5935s
2.6455 e0.3345s
4.4769 e
(8.4103s + 1)
(8.7940s + 1)

In the SIMC design method, the tuning parameter ci is


suggested to be taken as time delay ii for better performance.
In the present work, the single-input single-output (SISO)
SIMC tuning method is used for the demonstration purpose
only; other methods such as the ZieglerNichols tuning rule
can also be used.

(38)

The actual EOTF models are derived using eqs 10 and 11


and are expressed as
g11eff =

y1
u1

(7.7878s + 0.8465) e1.75s


2.2 es
+
(7s + 1)
(66.5s 2 + 16.5s + 1)
(39)

5. SIMULATION EXAMPLES

eff
g22
=

We consider three simulation examples to show the simplicity


and eectiveness of the proposed method.
5.1. Example 1. The transfer function matrix of VL column
given by Luyben1 is considered and it is expressed as
2.2 es
1.3 e0.3s

(7s + 1)
(7s + 1)
G (s ) =

1.8s
4.3 e0.35s
2.8 e
(9.5s + 1) (9.2s + 1)

K = [kij ]

y2
u2

4.3 e0.35s
1.65 e1.1s

(9.2s + 1)
(9.5s + 1)

(40)

It is found that the open-loop time response of the ETF


(derived from RNGA) model is found to adequately match the
actual EOTF (derived from DRGA).
Simplied decoupler matrix (by Wang et al.23 method) is
expressed as

1
0.5909

1.45
s
D(s) = (5.9907s + 0.6512) e
0.7s
e

(9.5s + 1)

(37)

The normalized gain matrix (KN), RGA (), RNGA (),


average residence time (Tar), and RARTA () are calculated as

(41)

The extra time delay is incorporated in to the decoupler


matrix.
12402

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

Table 1. Decoupling Control System Elements Using Dierent Techniques6 for VL Column
decoupling schemes
ideal decoupling

decoupled process model


s
* = 2.2 e
g11
(7s + 1)

* =
g22

inverted decoupling (Nx= e0.7s)

normalized decoupling

proposed method

4.3 e1.05s
(9.2s + 1)

*=
g11

2.2 es
(7s + 1)

* =
g22

4.3 e1.05s
(9.2s + 1)

*=
g11

2.0785 e0.9558s
(6.6910s + 1)

* =
g22

4.4769 e1.5935s
(8.7939s + 1)

*=
g11

1.3535 e0.9559s
(6.6910s + 1)

* =
g22

2.6455 e1.5935s
(8.7943s + 1)

decoupler matrix

SIMC controller param.

1
gc11 = 1.59091 +

7s

(89.87s + 9.46)
d11 =
(25.116s 2 + 59.112s + 5.82)
d 22 =

(89.87s + 9.46)e0.7s
(25.116s 2 + 59.112s + 5.82)

d12 =

(53.105s + 5.59)
(25.116s 2 + 59.112s + 5.82)

d 21 =

( 42.504s 2 + 52.052s + 6.16) e0.7s


(25.116s 2 + 59.112s + 5.82)

1
gc22 = 1.01881 +

8.4s

1
0.5909

D(s) = (5.9907s + 0.6512) e0.75s

(9.5s + 1)

1
gc11 = 1.59091 +

7s

(8.4103s + 1)
1.5357

(8.7939s + 1)

D(s) =

0.6903s
(6.1970s + 1) e
1.6923 e1.2590s

(6.6910s + 1)

gc11 =1.68411 +

6.6910s

gc22 =0.61631 +

8.7939s

1
0.5909

D(s) = (5.9907s + 0.6512) e1.45s


0.7s
e

(9.5s + 1)

gc11 = 2.58581 +

6.6910s

1
gc22 =1.01881 +

8.4s

1
gc22 =1.60661 +

8.276s

Figure 3. Closed-loop responses (sequential step changes in the set point) for the VL column.

* 0
g11 g12 1 d12 e0.7s g11

g21 g22 d 21 e0.7s 0 g *


22

*=
g11

* =
g22

2.6455 e1.0345s
(8.7940s + 1)

Using the tuning rules of eqs 35 and 36, the controller

The incorporated extra time delay in the decoupler matrix


will make the decoupled process model as
* = g
g11
11

1.3535 e0.9559s
(6.6910s + 1)

parameters are estimated based on the corresponding ETF


models. In the SIMC24 tuning formula, the closed-loop time

* = g e0.7s
g22
22

constant ci is taken as same as ii and the PI parameters are


given by

The decoupled process models are expressed as


12403

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

Table 2. Comparison of IAE and ISE Values for Dierent Decoupling Control Schemes for VL Column (Example 1)
ideal
performance indices
yr1
IAE

ISE

y1

normalized

inverted

y2

y1

y2

y1

y2

2.1830

0.7186

2.4370

2.2640

2.1790

0.0111

1.9330

0.0156

0.0049

2.4870

0.4360

3.6120

0.0020

2.4320

0.0027

2.1150

disturbance
yr1

0.9950
1.6910

1.438
0.0947

0.9573
1.6890

2.3840
0.7161

0.9950
1.6930

1.4380
1.54 105

0.9540
1.4480

1.4600
3.49 105

yr2

2.5 106

1.8840

0.0249

2.6930

5.71 107

1.7750

1.21 106

1.5370

disturbance

0.0621

0.2279

0.0591

0.4063

0.0621

0.2179

0.0531

0.1928

of input uncertainty plot, the proposed and normalized


decoupled control systems show the same stability in low and
high frequency regions. However, the ideal decoupling shows
the less stability in low frequency region and better stability in
high frequency region. Overall, the proposed method has better
robust stability than that of the normalized and ideal
decoupling control schemes.
5.2. Example 2. The transfer function matrix of industrialscale polymerization (ISP) reactor proposed by Chien et al.27 is
given by

(42)

For the VL distillation column system, the decoupler,


controller matrix, and decoupled process models for the
proposed method and for the three decoupling control
techniques are listed in Table 1.
Figure 3 shows the closed-loop responses of the VL
distillation column subject to the sequential unit step changes
at t = 0 and t = 50, respectively, and to show the disturbance
rejection performance, a step disturbance d = 0.5 at t = 100 is
introduced in both the loops. It can be seen that the interaction
is reduced and the set point response is improved in the
proposed decoupling control method compared to that of ideal,
inverted, and normalized decoupling control techniques. The
IAE and ISE values are listed in Table 2, and it is shown that
the IAE and ISE values are low for the proposed method
(designed based on ETFs). This indicates that response is fast
and less oscillations in both response and interactions. For
quantitative performance measurement, the sum of IAE and
ISE values are listed in Table 3. The overall response in terms

22.89 e0.2s 11.64 e0.4s

4.572s + 1 1.807s + 1
(s ) =

0.2s
5.8 e0.4s
4.689 e
2.174s + 1 1.801s + 1

performance indices
a

ideal

normalized

inverted

proposed

5.3935
3.6697

8.7490
5.1230

4.6241
3.4681

4.0663
2.9850

(43)

The normalized gain matrix (KN), RGA (), RNGA (),


average residence time (Tar), and RARTA () are calculated as
22.89 11.64
K=

4.689
5.8
0.7087 0.2913
= K K T =

0.2913 0.7087
0.5482 0.4518
= KN KNT =
0.4518 0.5482

Table 3. Performance Indices for VL Column (Example 1)

y2

yr2

0
2.58581 +

6.6910s

Gc(s) =

0
1.60661 +

8.276s

IAE
ISEa

proposed

y1

4.772 2.207
Tar =
2.374 2.201

Sum of main and interaction responses.

4.7967 5.2741
KN = K Tar =

1.9751 2.6352

of IAE and ISE values are better in the case of proposed


method. The inverted decoupling method is also found to be
sensitive to modeling errors,7,8 and it cannot be directly
implement to the higher dimension systems.7,9
Robustness. Among the various methods available for the
study of the robustness analysis in multivariable systems,25 the
method based on the inverse of maximum singular value is easy
to use and to compare the dierent control system
stabilities.20,26 The stability bound of the VL column is
shown in Figure 4. Figure 4 shows the frequency plot of
inverse of maximum singular value, indicating the stability
bounds of VL column. The region below the curve denotes the
stability region and the region above the curve denotes the
instability region. More area under the curve indicates high
stability of the system. At low frequency region, the proposed,
normalized, and ideal decoupling control systems have same
stability. As the frequency increases, the proposed method
shows more stability in the output uncertainty plot. In the case

0.7736 1.5508
==

1.5508 0.7736

By using the RGA and RNGA concepts, the ETF model


parameters are deduced by
32.3003 39.9535
K = K =

16.0947 8.1844
3.5368 2.8022
= T =

3.3713 1.3932
0.1547 0.6203
= L =

0.3102 0.3094

ETF model is expressed as


12404

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

Figure 4. Stability regions of output and input uncertainties for the VL column. (Left gure: for output uncertainty.)

Table 4. Decoupling Control System Elements Using Dierent Techniques for ISP Reactor
decoupling schemes
inverted decoupling29

proposed method

decoupled process model


0.4s

22.89 e
(4.572s + 1)
0.4s
* = 5.8 e
g22
(1.801s + 1)
*=
g11

*=
g11

32.3003 e0.3547s
(3.5368s + 1)

* =
g22

8.1844 e0.3094s
(1.3932s + 1)

32.3003 e0.1547s 39.9535 e0.6203s

(2.8022s + 1)
3.5368s + 1
G (s) =

0.3102s
8.1844 e0.3094s
16.0947 e
(3.3713s + 1)
1.3932s + 1

decoupler matrix

y1
u1

*=
g11

y2
u2

1
gc22 = 0.38811 +

1.801s

32.3003 e0.3547s
3.5368s + 1

* =
g22

1
gc11 = 0.15441 +

2.8376s

gc22 = 0.27511 +

1.3932s

8.1844 e0.3094s
1.3932s + 1

The controller parameters for the proposed and inverted


decoupling methods are estimated by using eqs 35 and 36. The
controller and decoupler elements of the proposed and inverted
decoupling methods are listed in Table 4.
The centralized control parameters for the ISP reactor
system are estimated by Xiong et al.28 method using the SIMC
tuning method.

(44)

(16.948s + 9.4103) e0.2s


22.89 e0.2s
+
(4.572s + 1)
(3.9284s 2 + 3.981s + 1)

1
1

0.35041 +
0.49931 +

1.6s
1.5416s

Gc(s) =

0.05861 + 1 0.38811 + 1

1.807s
1.801s

(45)
eff
g22
=

1
gc11 =0.24961 +

3.2s

(2.3249s + 0.5085)e0.2s

e0.2s
(1.807s + 1)

D(s) =

( 1.4560s 0.8084)

(2.174s + 1)

The actual EOTF models are derived using eqs 10 and 11


and are expressed as
g11eff =

SIMC controller param.

(2.3249s + 0.5085)
1

(1.807s + 1)

D(s) =

( 1.4560s 0.8084)

(2.174s + 1)

(10.9017s + 2.3844) e0.4s


5.8 e0.4s
+
(1.801s + 1)
(3.9284s 2 + 3.981s + 1)

(47)

(46)

Figure 5 shows the closed-loop step responses of the ISP


reactor subject to the sequential unit step changes at t = 0 and t
= 25, respectively. It can be seen that the interaction is reduced
and set point response is improved in the proposed decoupled
control system compared to the centralized control system
(Xiong et al.:28 controller tuned by the SIMC method).
Compared to centralized control system, the proposed

It is found that the open-loop time response of the ETF


(derived from RNGA) model adequately matches that of the
actual EOTF (derived from DRGA).
The decoupler matrix elements are estimated by Wang23 et
al. method. The decoupled process models are expressed as (g11
*
= g11 e0.2s),
12405

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

Figure 5. Closed-loop response (sequential step changes in the set point) for the ISP reactor system.

0.098

0.043
K=
0.012

0.013
139

172
Tar =
184

188

decoupling control system has lead-lag lter term with the


diagonal PI control elements. This lead-lag term improves the
performance better than the centralized control system. The
IAE values are listed in Table 5, and it is shown that the IAE
values are considerably reduced for the proposed decoupled
control system compared to that of inverted decoupling and
centralized control systems.
Table 5. Comparison of IAE Values for the Dierent Control
Schemes for ISP Reactor

0.036 0.014 0.017

0.092 0.011 0.012


0.016 0.102 0.033

0.015 0.029 0.108


176 190 185

146 189 191


185 134 172

190 169 146

IAE values
method used
proposed method

change in set point


yr1

y1

y2

IAEa

0.9674

0.0024

1.8948

centralized controller
inverted decoupler
a

yr2

0.0141

0.9109

yr1

0.7270

0.4039

yr2

0.5917

0.9864

yr1

1.0813

0.0001

yr2

0.0014

0.8776

= K K T
1.2207 0.2051 0.0053 0.0103

0.1947 1.2198 0.0136 0.0116

=
0.0106 0.0085 1.1095 0.0904

0.0154 0.0062 0.0907 1.1124

2.7090
1.9604

Sum of main and interaction responses.

KN = K Tar

5.3. Example 3. The transfer function matrix of the


temperature control of the four room process proposed by
Shen et al.30 is given by
0.098 e17s

122s + 1

25s
0.043 e
147s + 1
G(s) =
0.012 e31s
153s + 1

0.013 e32s

156s + 1

= 1.0 e03
0.7050

0.2500

0.0652

0.0691

0.036 e27s 0.014 e32s 0.017 e30s

149s + 1
158s + 1
155s + 1

0.092 e16s 0.011 e33s 0.012 e34s


130s + 1
156s + 1
157s + 1

34s
16s
0.016 e
0.102 e
0.033 e26s
151s + 1
118s + 1
146s + 1

0.015 e31s 0.029 e25s 0.108 e18s

159s + 1
144s + 1
128s + 1

0.2045 0.0737 0.0919

0.6301 0.0582 0.0628


0.0865 0.7612 0.1919

0.0789 0.1716 0.7397

= KN KN T
1.1389 0.1287 0.0036 0.0067

0.1238 1.1389 0.0078 0.0073


=
0.0057 0.0055 1.0710 0.0597

0.0094 0.0046 0.0596 1.0737

(48)

The normalized gain matrix (KN), average residence time (Tar),


RGA (), RNGA (), and RARTA () are calculated as
12406

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

Table 6. Decoupling Control System Elements Using Normalized30 and Proposed Decoupling Method for Example 3
decoupling schemes
proposed method

normalized30

0.9330

0.6361
==
0.5404

0.6083

decoupled process model

SIMC controller param.

15.8615s

*=
g11

0.0803 e
(113.8299s + 1)

gc11 = 44.68551 +

113.8299s

* =
g22

0.0754 e14.9383s
(121.3735s + 1)

gc22 = 53.87921 +

119.5064s

* =
g33

0.0919 e15.4444s
(113.9022s + 1)

gc33 = 40.12511 +

113.9022s

* =
g44

0.0971 e17.3739s
(123.5480s + 1)

gc44 = 36.61751 +

118.6090s

*=
g11

e21.8281s
(113.8299s + 1)

gc11 = 2.60861 +

113.8299s

* =
g22

e21.3160sd
(121.3735s + 1)

gc22 = 2.84701 +

121.3735s

* =
g33

e22.2075s
(113.9022s + 1)

gc33 = 2.56451 +

113.9022s

* =
g44

e23.125s
(118.609s + 1)

gc44 = 2.67131 +

123.5480s

113.8299

93.5129
= T =
82.6888

94.8909

0.6274 0.6821 0.6461

0.9336 0.5731 0.6269


0.6532 0.9653 0.6604

0.7460 0.6575 0.9652

93.4854 107.7763 100.1438

121.3735 89.4106 98.4299


98.6274 113.9022 96.4196

118.6090 94.6824 123.5480

By using the RGA and RNGA concepts, the ETF model


parameters are deduced by
15.8615

15.9036

=L=
16.7539

19.4648

0.0803 0.1755 2.6653 1.6468

0.2209 0.0754 0.8097 1.0363

K=K=
1.1315 1.8867 0.0919 0.3648

0.8420 2.4132 0.3197 0.0971

16.9403 21.8281 19.3827

14.9383 18.9138 21.3160


22.2075 15.4444 17.1760

23.1250 16.4379 17.3739

0.0803 e15.8615s 0.1755 e16.9403s


2.6653 e21.8281s
1.6468 e19.3827s

93.4854s + 1
107.7763s + 1
100.1438s + 1
113.8299s + 1

15.9036s
0.0754 e14.9383s 0.8097 e18.9138s
1.0363 e21.3160s
0.2209 e
93.5129s + 1
121.3735s + 1
89.4106s + 1
98.4299s + 1
G(s) =

1.1315 e16.7539s
1.8867 e22.2075s 0.0919 e15.4444s 0.3648 e17.1706s

98.6274s + 1
113.9022s + 1
96.4196s + 1
82.6888s + 1
0.8420 e19.4648s
2.4132 e23.1250s
0.3197 e16.4379s 0.0971 e17.3739s

94.8909s + 1
118.6090s + 1
94.6824s + 1
123.5480s + 1

(49)

The decoupler is expressed as

( 44.8163 0.3673)e10s ( 17.4338 0.1429)e15s ( 21.1633 0.1735)e13s

1
149s + 1
158s + 1
155s + 1

9s
17s
18s
(
60.7609
0.4674)e
(
15.5485
0.1196)e
(
16.952
0.1304)e

147s + 1
156s + 1
157s + 1
D(s) =

15s
18s
10s
(
13.8824
0.1176)e
(
18.5142
0.1569)e
(
38.173
0.3235)e

153s + 1
151s + 1
146s + 1

( 15.4112s 0.1204)e14s ( 17.7792s 0.1389)e13s ( 34.3704s 0.2685)e7s

156s + 1
159s + 1
144s + 1

12407

(50)

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

Figure 6. Closed-loop response plot for example 3 (solid, proposed method; dash, Shen et al.30).

Table 7. Comparison of IAE Values for Both Control Schemes of Example 3


normalized30
performance indices
yr1
IAE

IAEa
a

proposed method

y1

y2

y3

y4

y1

y2

y3

y4

48.3300

6.4670

1.9900

2.4620

34.3100

2.1870

5.3270

5.3070

yr2

6.7320

47.2300

4.8830

3.6730

3.0770

32.0900

4.8850

5.1690

yr3

2.4360

2.2010

48.6100

3.4640

3.8350

3.6666

33.5900

2.5010

yr4

2.6990

2.3990

3.9540

50.6700

3.8380

4.0400

2.5020

36.9400

238.2

183.2640

Sum of main and interaction responses.

The unit set-point changed in r1 at t = 0, r2 at t = 1000, r3 at t =


2000, and r4 at t = 3000. Compared to Shen et al.29 method, the
overall performances are improved by the proposed method.
IAE values are listed in Table 7. The three simullation examples

Decoupled process models and controller settings for the


proposed and normalized30 decoupling methods are listed in
Table 6. Figure 6 shows the closed-loop responses for the
proposed method and for the normalized decoupling method.30
12408

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

ij = element of relative normalized gain


ij = average residence time
ij = average residence time when other loops are closed
ij = relative gain array element

show that the nominal performance of the proposed simplied


decoupler with ETF based controller method is better than the
ideal, inverted, and normalized decoupling and centralized
control systems. The performance by the robustness analysis
shows that the proposed method gives a better robust
performance compared to that of the other decoupling control
techniques.

Subscripts

6. CONCLUSIONS
A decoupler with a decentralized control system is designed
based on ETF models of MIMO systems. It is shown that, for
the examples considered here, the decoupler with decentralized
control system (designed based on ETFs) reduced the
interaction and gives better responses when compared to that
of the centralized control system, ideal, inverted, and
normalized decoupling control methods. The proposed
decoupler with the decentralized control system is easy to
design when compared to the method of EOTF (derived using
Maclaurin series) and Cai et al.6 It is shown that the perfect
control assumption4 is valid when introducing the decoupler in
to the control system. The independent controllers are
designed using the SIMC method based on the calculated
ETFs. The simulation examples show the better performance of
the proposed method compared with the centralized control
system, ideal, inverted, and normalized decoupling methods.

i,j = loop representation

REFERENCES

(1) Luyben, W. L. Simple Method for Tuning SISO Controllers in


Multivariable Systems. Ind. Eng. Chem. Proc. Des. Dev. 1986, 25, 654
669.
(2) Shen, S. H.; Yu, C. C. Use of Relay Feedback Test for Automatic
Tuning of Multivariable Systems. AIChE J. 1994, 40, 627646.
(3) Loh, P. A.; Hang, C. C.; Quek, K. C.; Vasnani, U. V. Auto-tuning
of Multi-loop Proportional-Integral Controllers Using Relay Feedback.
Ind. Eng. Chem. Res. 1993, 32, 11021107.
(4) Vu, T. N. L.; Lee, M. Independent Design of Multi-loop PI/PID
Controllers for Interacting Multivariable Processes. J. Process Control
2010, 922933.
(5) Tavakoli, S.; Griffin, I.; Fleming, P. J. Tuning of Decentralized PI
(PID) Controllers for TITO Processes. Control Eng. Pract. 2006, 14
(9), 10691080.
(6) Cai, W. J.; Ni, W.; He, M. J.; Ni, C. Y. Normalized Decoupling
A New Approach for MIMO Process Control System Design. Ind. Eng.
Chem. Res. 2008, 47 (19), 73477356.
(7) Gagnon, E.; Pomerleau, A.; Desbiens, A. Simplified, Ideal, or
Inverted Decoupling? ISA Trans. 1998, 37 (4), 265276.
(8) Weischedel, K.; McAvoy, T. J. Feasibility of Decoupling in
Conventionally Controlled Distillation Columns. Ind. Eng. Chem.
Fundam. 1980, 19 (4), 379384.
(9) Yunhui, L. U. O.; Hongbo, L. I. U.; Lei. J. I. A. Improved Inverted
Decoupling Control Using Dead-time Compensator for MIMO
Processes. Proceedings of the 29th Chinese Control Conference, China,
2010.
(10) McAvoy, T. J. Interaction Analysis: Principles and Applications;
Instrument Society of America: Research Triangle Park, NC, 1983.
(11) Wade, H. L. Inverted Decoupling: A Neglected Technique. ISA
Trans. 1997, 36 (1), 310.
(12) Kumar, N.; Pandit, M.; Chidambaram, M. Multivariable Control
of Four-Tank System, . Indian Chem. Eng., Sect. A 2004, 46 (4), 216
221.
(13) Maghade, D. K.; Patre, B. M. Decentralized PI/PID Controllers
based on Gain and Phase Margin Specifications for TITO Processes.
ISA Trans. 2012, 51 (4), 550558.
(14) Huang, H. P.; Jeng, J. C.; Chiang, C. H.; Pan, W. A Direct
Method for Multi-Loop PI/PID Controller Design. J. Process Control
2003, 13 (8), 769786.
(15) Xiong, Q.; Cai, W. J. Effective Transfer Function Method for
Decentralized Control System Design of Multi-input Multi-output
Processes. J. Process Control 2006, 16 (8), 773784.
(16) He, M. J.; Cai, W. J.; Ni, W.; Xie, L. H. RNGA Based Control
System Configuration for Multivariable Processes. J. Process Control
2009, 19 (6), 10361042.
(17) Naini, N. M.; Fatehi, A.; Sedigh, A. K. InputOutput Pairing
Using Effective Relative Energy Array. Ind. Eng. Chem. Res. 2009, 48
(15), 71377144.
(18) Shen, Y.; Cai, W. J.; Li, S. Multivariable Process Control:
Decentralized, Decoupling, or Sparse? Ind. Eng. Chem. Res. 2010, 49
(2), 761771.
(19) Hu, W.; Cai, W. J.; Xiao, G. Decentralized Control System
Design for MIMO Processes with Integrators/Differentiators. Ind. Eng.
Chem. Res. 2010, 49 (24), 1252112528.
(20) Kumar, V. V.; Rao, V. S. R.; Chidambaram, M. Centralized PI
Controllers for Interacting Multivariable Processes by Synthesis
Method. ISA Trans. 2012, 51 (3), 400409.
(21) Seborg, D. E.; Edgar, T. F.; Mellichamp, D. A. Process Dynamics
and Control, 2nd ed.; John Wiley & Sons Asia Pte. Ltd.: Singapore,
2009.

AUTHOR INFORMATION

Corresponding Author

*E-mail: chidam@iitm.ac.in.
Notes

The authors declare no competing nancial interest.

NOMENCLATURE
DRGA = dynamic relative gain array
IAE = integral of absolute error
IMC = internal model control
ISE = integral of the squared error
ISP = industrial scale polymerization
SIMC = simplied IMC
G, Gc = process and controller transfer function matrices
gij, gc,i,j, gij = process, controller, and equivalent transfer
function models
geff
ij , g*
ii = eective open-loop transfer function and decoupled
process model
dij = decoupler elements
= process and eective steady-state gain
kpij, kpij
ij, ij = process and eective time constant
ij, ij = process and eective time delay

Y = closed-loop process response vector


yi = closed-loop process response
tc = closed time constant of the model
Kci = Controller gain
Ii, Di = integral and derivative time
= relative normalized gain array (RNGA)
= relative gain array (RGA)
= relative average residence time array (RARTA)
KN = normalized gain matrix
KNij = normalized gain
s = laplace domain
t = time
uj = manipulated variable
ij = relative average residence time
12409

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

Industrial & Engineering Chemistry Research

Article

(22) Bequette, B. W. Process Control: Modeling, Design, and


Simulation, 1st ed.; Prentice Hall: Upper Saddle River, NJ, 2003.
(23) Wang, Q. G.; Huang, B.; Guo, X. Auto-tuning of TITO
Decoupling Controllers from Step Tests. ISA Trans. 2000, 39 (4),
407418.
(24) Skogestad, S. Simple Analytical Rules for Model Reduction and
PID Controller Tuning. J. Process Control 2003, 13 (4), 291309.
(25) Morari, M.; Zariou, E. Robust Process Control; Prentice Hall:
Englewood Clis, NJ, 1989.
(26) Maciejowski, J. M. Multivariable Feedback Design; AddisonWesley: New York, 1989.
(27) Chien, I. L.; Huang, H. P.; Yang, J. C. A Simple Multi-Loop
Tuning Method for PID Controllers with No Proportional Kick. Ind.
Eng. Chem. Res. 1999, 38 (4), 14561468.
(28) Xiong, Q.; Cai, W. J.; He, M. J. Equivalent Transfer Function
Method for PI/PID Controller Design of MIMO Processes. J. Process
Control 2007, 17 (8), 665673.
(29) Garrido, J.; Vazquez, F.; Morilla, F. An Extended Approach of
Inverted Decoupling. J. Process Control 2011, 21 (1), 5568.
(30) Shen, Y.; Cai, W. J.; Li, S. Normalized Decoupling Control for
High-Dimensional MIMO Processes for Application in Room
Temperature Control HVAC Systems. Control Eng. Pract. 2010, 18
(6), 652664.

12410

dx.doi.org/10.1021/ie301448c | Ind. Eng. Chem. Res. 2012, 51, 1239812410

You might also like