See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/222820180

An extended approach of inverted decoupling

Article in Journal of Process Control · January 2011

DOI: 10.1016/j.jprocont.2010.10.004

CITATIONS READS

99 295

3 authors, including:

Juan Garrido Francisco Vazquez

University of Cordoba (Spain) University of Cordoba (Spain)
61 PUBLICATIONS 841 CITATIONS 95 PUBLICATIONS 1,052 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Juan Garrido on 13 January 2024.

The user has requested enhancement of the downloaded file.

Juan Garrido, Francisco Vázquez, Fernando Morilla, An extended approach of inverted decoupling, Journal of
Process Control, Volume 21, Issue 1, January 2011, Pages 55-68, ISSN 0959-1524,
http://dx.doi.org/10.1016/j.jprocont.2010.10.004.
(http://www.sciencedirect.com/science/article/pii/S0959152410001903)

AN EXTENDED APPROACH OF INVERTED DECOUPLING

Juan Garridoa*, Francisco Vázqueza, Fernando Morillab

a
Department of Computer Science and Numerical Analysis, University of Córdoba,
Campus de Rabanales, 14071, Córdoba, Spain
* Fax: (+34)957218729; e-mail: juan.garrido@uco.es
b
Department of Computer Science and Automatic Control, UNED, Juan del Rosal 16,
28040, Madrid, Spain

Abstract

This paper presents an extension of the inverted decoupling approach that allows for
more flexibility in choosing the transfer functions of the decoupled apparent process. In
addition, the expressions of the inverted decoupling are presented for general nxn
processes, highlighting that the complexity of the decoupler elements is independent of
the system size. The realizability conditions are stated in order to select a proper
configuration, and the different possible cases for each configuration are shown.
Comparisons with other works demonstrate the effectiveness of this methodology,
through the use of several simulation examples and an experimental lab process.

Keywords: decoupling control, inverted decoupling, simplified decoupling, PID

control.

1. Introduction

Multi-input multi-output (MIMO) processes consist of several measurement and control

signals, and there are often complicated couplings between them, which can cause
difficulties in feedback controller design. Control engineers traditionally solve these
problems using single-loop PID controllers because they can be easily understood and
implemented [1]. These decentralized approaches have evolved through years of
experience, and they are adequate when the interactions in different channels of the
process are modest [2-3]. Nevertheless, when interactions are significant, the
decoupling is often treated inefficiently, e.g., by detuning control loops. In these cases, a
full matrix controller (centralized control) is advisable.

There are two approaches of centralized control: a pure centralized strategy [4-9] or a
decoupling network combined with a diagonal decentralized controller [10-14]. A
decentralized control system with a decoupling matrix can be designed by combining a
diagonal controller C(s) with a block compensator D(s) in such a way that the controller
sees the apparent process G(s)·D(s) as a set of n completely independent processes. The
essence of decoupling is the imposition of a calculation net that cancels the existing
process interaction, allowing for independent control of the loops. Although model
predictive control (MPC) is becoming the standard technique to solve multivariable

-1-
control problems in the process industry, several authors [6, 7, 12] claim that when
MPC is used today, it is mostly used on a higher level to give setpoints to the PID
controllers that are operating on the basic level. They are operating in a supervisory
mode with sampling times that are longer than in the PID controllers at the lower level.
And there are some difficulties in dealing with the interaction at the MPC level because
the bandwidths of the MPC loops are limited. Therefore, the centralized control using
PID controllers and decoupler networks is an interesting strategy in the process
industry.

The theory of decoupling control has been addressed in the literature [15-18]. Some
decoupling schemes are static [19], and others are dynamic [6], [13], [14], [20]. Most of
these methodologies focus on systems with two inputs and two outputs (TITO systems).

Most decoupling approaches use a conventional decoupling scheme in which the

process inputs are derived by a time-weighted combination of feedback controller
outputs (Figure 1). In this case, the design of the decoupler network for an nxn process
is obtained from (1), generally specifying n elements of the decoupler D(s) or the n
desired transfer functions of the apparent process Q(s). The most extended forms of
conventional decoupling were termed ideal and simplified decoupling in [10]. This
approach has received considerable attention in both control theory and industrial
practice for several decades.

INSERT HERE FIGURE 1

D(s) = G -1 (s)·Q(s) (1)

In ideal decoupling, the goal is to make the apparent processes as simple as the diagonal
elements of the process matrix G(s). The main inconveniences of this method are the
complexity of the decoupler elements (2) and realizability problems.

Adj ( G(s) ) (2)

D(s) = ·G diag (s)
G(s)

On the other hand, with simplified decoupling, n elements of the decoupler, generally
the diagonal ones, are set to unity. Decoupler network design is easier, but the
complexity of the apparent process (3) is greater since it consists of the determinant of
G(s) and the g ii (s) cofactor, Gii(s).

G(s) (3)
q i (s) =
G ii (s)

An alternative means of decoupling, called inverted decoupling, derives a process input

as a time-weighted combination of one feedback controller output and the other process
inputs. It is rarely mentioned in the literature [11], [21], [22], [23], and in those cases, it
is only applied to TITO processes using the scheme depicted in Figure 2. In this case, it
is possible to keep the same apparent process of ideal decoupling while using the simple
decoupler elements of simplified decoupling [11].

-2-
In [21], a comparative study of simplified, ideal and inverted decoupling is presented,
and the practical advantages of inverted decoupling are mentioned. Chen and Zhang
[23] improved upon the inverted decoupling technique for a class of stable linear
multivariable processes with multiple time delays and non-minimum-phase zeros.

INSERT HERE FIGURE 2

Nevertheless, the authors have not found any published works in which inverted
decoupling is applied to processes that are larger than a 2x2 system. In this work, the
formulation of a general inverted structure for decoupling of nxn processes is presented.
It is shown that the complexity of the transfer functions of the decoupler elements and
of the decoupled equivalent open loop process is independent of the system size. In
conventional decouplings, when the input and output number of the square system
increases, the complexity of the decoupler elements or the decoupled process is
increased as well, and so they are more difficult to implement.

In addition, this formulation is carried out using decoupler elements that differ from
unity, in such a way that the decoupled apparent processes can be specified with more
flexibility. The paper is structured as follows. Section 2 presents the general formulation
of inverted decoupling for nxn processes and the different possible configurations.
Furthermore, a study focused on TITO processes is carried out in order to show all
possible cases. The expressions for 3x3 processes are also obtained. In Section 3, the
realizability conditions are stated and some rules are suggested for selecting a proper
configuration. In Section 4, the performance of the proposed decoupling methodology is
tested and compared with other methods using several simulation examples and a real
quadruple tank process. Finally, conclusions are summarized in Section 5.

2. Inverted decoupling for nxn processes

Currently, most papers in the literature regarding inverted decoupling only apply the
approach to 2x2 processes (Figure 2), and only the expressions for this case are
presented. In this section, the formulation for the general case of nxn processes is
calculated, and then, the equations for the cases n=2 and n=3 are given.

In order to study inverted decoupling for a square process with n inputs and n outputs,
the use of a matrix representation is proposed as shown in Figure 3. There are n
elements of the decoupler (Dd(s) matrix), which try to directly connect the decoupler
inputs “m” with the process inputs “u”, while the rest of the elements (Do(s) matrix)
feedback the input process “u” toward the decoupler inputs in order to decouple the
system.

INSERT HERE FIGURE 3

The whole decoupler D(s) is split into two matrices: a matrix Dd(s) in the direct path
(between controller outputs c and process inputs u) and a matrix Do(s) in a feedback
loop (between process inputs u and controller outputs c). The Dd(s) matrix must have
only n non-zero elements, since there must be only a direct connection for each process
input. Note that these relationships are not required in the Do(s) matrix. Additionally,
since the signal flow direction in Do(s) is opposite that of Dd(s), the corresponding
elements of Do(s) that must equal zero are the transpose non-zero elements of Dd(s).

-3-
For instance, in a 3x3 process, if element Dd(3,1) is specified as a direct path between
u 3 and c 1 , there will not be feedback from u 3 toward c 1 , and thus, the element Do(1,3)
must be zero.

Following the decoupler representation given in Figure 3, the expression of the whole
decoupler matrix D(s) is obtained as follows:

D(s)=Dd(s)·(I-Do(s)·Dd(s))-1 (4)

The transfer function matrix D(s) of conventional decoupling is related to the inverted
decoupling structure according to (4). As it is a complex expression, it is easier to work
with its inverse, which is very simple, as follows:

D-1 (s)=(I-Do(s)·Dd(s))·Dd -1 (s)=Dd -1 (s)-Do(s) (5)

Inverting equation (1) and substituting it into (5), the following expression is obtained:

Dd -1 (s)-Do(s)=Q-1 (s)·G(s) (6)

This last expression can be used to calculate the elements of the inverted decoupling. Its
main advantage is its simplicity, regardless of the size of the system, because the Q(s)
matrix is chosen to be diagonal and the resulting subtraction of the inverse of Dd(s) and
Do(s) is a transfer matrix with only one element to be calculated in each position.

Note that Dd(s) has to be non-singular because it is inverted, and therefore, when its
elements are chosen, only one element in each row and column can be selected.
Consequently, for an nxn process there are only n! possible choices of Dd(s).

To name these possibilities, the authors propose a notation in which the indicated
number corresponds to the column with the selected element. For instance, in a 3x3
process, configuration 1-2-3 means that elements Dd(1,1), Dd(2,2) and Dd(3,3) are
chosen; configuration 3-1-2 means that elements Dd(1,3), Dd(2,1) and Dd(3,2) are
selected, and so on. Each configuration has a different set of decoupler elements, which
is interesting since some choices can result in non-realizable decoupler elements. Thus,
the configuration can be selected depending on the realizability, which will be discussed
later.

2.1. Inverted decoupling for 2x2 processes

In this section, a detailed study of inverted decoupling for 2x2 processes is carried out
using expression (6). In this case (n=2), there are two possible configurations to choose
for the Dd matrix: diagonal elements (configuration 1-2) or off-diagonal elements
(configuration 2-1). Hereafter, the Laplace variable s is omitted for simplicity.

a) Configuration 1-2

Using this configuration, the following expressions (8) for the elements of the inverted
decoupling are obtained from (7), which is derived from (6). These are the expressions
for general inverted decoupling with configuration 1-2. q 1 and q 2 are the desired

-4-
equivalent open loop transfer functions to be controlled by the decentralized control,
and they can be specified freely as long as the decoupler elements are realizable.
However, in most examples found in the literature on inverted decoupling, two of these
elements are set to unity, and so only two elements need to be implemented.

 1  g g12 
-do12   11
 dd
 11  =  q1 q1 
(7)
 1   g 21 g 22 
 -do 21   
 dd 22   q 2 q 2 
q1 -g12
dd11 = do12 =
g11 q1
(8)
-g q2
do 21 = 21 dd 22 =
q2 g 22

Usually, the two non-zero elements equal to one are the elements in the direct way, that
is, the elements of the Dd matrix. However, this case is only one of the four possible
cases according to the two elements chosen to be equal to unity. The control schemes of
these different cases are depicted in Figure 4. In an nxn process, the number of possible
cases for each configuration is nn. Nevertheless, the realizability of the decoupler
elements in each configuration is always the same, independent of the selected case. For
example, as shown in Table 1, for inverted decoupling of TITO processes using
configuration 1-2, the pair of decoupler elements that differ from one have the following
expressions:

-g12 -g 21
and (9)
g11 g 22

INSERT HERE FIGURE 4

Decoupler elements Decoupled process

dd11 =dd 22 =1
g 0 
Case 1 -g12 -g Q=  11 
do12 = , do 21 = 21  0 g 22 
g11 g 22
dd11 =do 21 =1
g 0 
Case 2 -g -g Q=  11 
do12 = 12 , dd 22 = 21  0 -g 21 
g11 g 22
do12 =do 21 =1
 -g 0 
Case 3 -g -g Q=  12 
dd11 = 12 , dd 22 = 21  0 -g 21 
g11 g 22
do12 =dd 22 =1
 -g 0 
Case 4 -g12 -g Q=  12 
dd11 = , do 21 = 21  0 g 22 
g11 g 22

Table 1 - Cases of 2x2 inverted decoupling with two unitary elements (configuration 1-
2)

b) Configuration 2-1

-5-
In this configuration, the general expressions (11) for the inverted decoupling are
obtained from (10), which is derived from (6). Note that care should be taken in
inverting the Dd matrix.

 1   g11 g12 
 -do11 dd   q q1 
 21   1
= (10)
 1  g g 22 
 -do 22   21 
 dd12   q2 q 2 

-g11 q2
do11 = dd12 =
q1 g 21
(11)
q1 -g
dd 21 = do 22 = 22
g12 q2

If two of these elements are set to unity, there are again four possible cases (as with
configuration 1-2). However, the pair of decoupler elements that differ from one is
given by (12), which is the inverse of (9) in the other configuration.

-g11 -g 22
and (12)
g12 g 21

The cases in which the elements of the Dd matrix are equal to unity (case 1 in
configuration 1-2, for instance) are the only ones found in the literature.

2.2. Inverted decoupling for 3x3 processes

For 3x3 processes, the procedure is the same: we obtain the expressions from (6) after
choosing the configuration. In this case, there are six (3!) possible configurations
according to the three elements of Dd chosen to be non-zero. In the following, two
examples of these configurations are shown. Using configuration 1-2-3 (diagonal
elements of Dd) in (6), equation (13) is obtained. From this, the expressions of the
decoupler elements (14) are easily derived.

 1  g g12 g13 
 -do12 -do13   11 
 dd11   q1 q1 q1 
 1  g g 22 g 23 
 -do 21 -do 23  =  21  (13)
 dd 22   q2 q2 q2 
 1   g 31 g 32 g 33 
 -do31 -do32   
 dd 33   q 3 q3 q 3 

q1 -g12 -g13
dd11 = do12 = do13 =
g11 q1 q1
-g 21 q2 -g 23
do 21 = dd 22 = do 23 = (14)
q2 g 22 q2
-g 31 -g 32 q3
do31 = do32 = dd 33 =
q3 q3 g 33

-6-
Using configuration 2-1-3 (15) and operating in the same way, the corresponding
expressions of the decoupler elements (16) are achieved.

 1  g g12 g13 
 -do11 -do13   11 
 dd 21   q1 q1 q1 
 1  g g 22 g 23 
 -do 22 -do 23  =  21  (15)
 dd12   q2 q2 q2 
 1   g 31 g 32 g 33 
 -do31 -do32   
 dd 33   q 3 q3 q 3 

-g11 q2 -g13
do11 = dd12 = do13 =
q1 g 21 q1
q1 -g 22 -g 23
dd 21 = do 22 = do 23 = (16)
g12 q2 q2
-g 31 -g 32 q3
do31 = do32 = dd 33 =
q3 q3 g 33

These expressions are as simple as in 2x2 processes. If three elements are fixed to unity,
there are 27 cases for each configuration. Nevertheless, we study only the case in which
the elements of the Dd matrix are equal to unity. In this case and using configuration 1-
2-3, the following expressions are obtained:

-g12 -g13
dd11 =1 do12 = do13 =
g11 g11
-g 21 -g 23
do 21 = dd 22 =1 do 23 = (17)
g 22 g 22
-g 31 -g 32
do31 = do32 = dd 33 =1
g 33 g 33

The apparent processes are the diagonal elements of G(s), as obtained with ideal
decoupling, and the decoupler elements are as simple as the simplified decoupler
elements for 2x2 processes.

In [24], an interaction compensator is designed for a 3x3 process as a static

compensator. This design is carried out as in the simplified decoupling case. The
expressions of simplified decoupling for 3x3 processes are shown in (18) where Gij is
the cofactor corresponding to g ij (s) of G(s). Then, the realization is made as with the
inverse decoupling structure using the static simplified decoupling elements for 3x3
processes (18). It is said that the resulting interaction compensator obtains the apparent
process of the ideal decoupling case in steady-state conditions. This last statement is
wrong, because it would only be correct for TITO processes, not for systems with
greater dimensions.

In the case of 3x3 processes, the cofactor of g 11 is G11= g 22 ·g 33 -g 32 ·g 23 . Therefore, it is

easy to find that the expressions of the inverted decoupler elements in (17) are much
simpler than the simplified decoupler elements in (18). In addition, the apparent
processes are given by equation (3), which is much more complex than the diagonal
elements of G(s).

-7-
  G 21 G 31 
 1 
 G 22 G 33 
 G12 G 32 
D=  11 1  (18)
G G 33 
 G13 G 23 
 11 1 
G G 22 

2.3. General expressions of inverted decoupling

From (6) and the equations of the previous sections, it is possible to obtain the general
expressions of inverted decoupling for nxn processes. If the configuration {p 1 -p 2 -…-p i -
…p n-1 -p n } is chosen, the elements that differ from zero of the Dd and Do matrices are
given by (19) and (20), respectively. The transfer functions of the apparent process q i (s)
can be selected in any way that assures the realizability of these decoupler elements.

qj
ddij = ∀i ; j = pi (19)
g ji

−g
doij = ij ∀i, j ; i ≠ pj (20)
qi

The following advantages of inverted decoupling are concluded from these general
expressions:

- The apparent processes do not contain sums of transfer functions, and therefore, tuning
of the diagonal controllers is easier. In multivariable processes with strong cross
couplings, even if the elements of the system have simple dynamics, conventional
decoupling may result in complicated diagonal apparent processes consisting of parallel
coupled processes that may have different signs and different time delays [12]. An
example of such a diagonal element is:

1.2 1
G(s)= e-4s - e-s (21)
( 0.5s+1)( 0.7s+1) ( 3s+1)( 2s+1)

which consists of a difference between two fairly simple transfer functions. The step
response of this process is shown in Figure 5. If PID controllers are used to control a
system with diagonal elements like (21), design methods that rely on simple process
dynamics, like step response methods [25], are not appropriate.

INSERT HERE FIGURE 5

- Decoupler elements do not contain either sum of transfer functions, so they are easy to
design. Using conventional decoupling, in some cases, it is possible to have decoupler
elements with complexity similar to that of process (21), and therefore, they are difficult
to implement.

-8-
- The complexity of the decoupler elements and apparent processes is always the same,
independent of the size of the process. With conventional decoupling, these elements
become more complex as the system size increases.

Nevertheless, inverted decoupling presents an important disadvantage: due to stability

problems, it cannot be applied to processes with RHP zeros in the determinant of the
transfer matrix G(s). For internal stability, these RHP zeros should appear in the
apparent process. In a conventional decoupling structure, RHP zeros of the determinant
of G(s) can be included in the desired equivalent open loop transfer functions q i (s).
However, this is not possible using inverted decoupling, because such RHP zeros would
appear as unstable poles in some decoupler elements do ij (s). If the multivariable RHP
zero is associated with a single output and is therefore included in the process transfer
functions of the same row, inverted decoupling can be applied because the RHP zero
will be cancelled.

Therefore, as mentioned in [23], it seems that inverted decoupling cannot be applied in

general to a process with a determinant that has RHP zeros. To apparently solve it, in
[23], the decoupler structure is changed from inverted decoupling to conventional ideal
decoupling without feedback, and so, the inverted decoupling structure is not used.

When the non-zero elements of the Dd matrix in (19) are fixed to unity, the expressions
of the decoupler elements and apparent processes q i (s) are given by (22), (23) and (24).

ddij =
1 ∀i ; j=pi (22)
−g
doij = ij ∀i, j ; i ≠ p j ; k pk =i (23)
gik
qj = g ji ∀j ; i pi =
j (24)

In this case (non-zero dd ij =1), inverted decoupling presents several practical advantages
over conventional decoupling, which makes this structure very interesting from a
practical point of view. These advantages [11] are the following:

- The apparent process seen by each controller is the same as that obtained if there
was no decoupling and the alternate controllers were in the manual mode.
- Inverted decoupling can often be implemented within a DCS using PID function
blocks with feedforward inputs. This will automatically provide such features as
initialization and bumpless transfer between manual and automatic.
- The antireset windup feature of the PID, combined with its feedforward input,
can be used to directly take into account the saturation of the manipulated
variables when inverted decoupling is implemented.
- When decoupler outputs are used as cascade setpoints to lower level controllers,
each decoupled control loop is immune to abnormalities (e.g., a valve at a limit
or a secondary controller in manual) in the secondary of the opposite control
loops.

Conventional decoupler networks present several practical disadvantages: problems like

bumpless transfer or antireset windup are more difficult to solve, and the controller
parameters should change when some decoupler element is disenable, since the apparent
process will be different. In addition, for high dimensional systems (3x3 or 4x4) the

-9-
design of conventional decouplers becomes more complicated and important
approximations are usually necessary. For these reasons, many industrial multivariable
processes are preferred to be controlled by decentralized PID controllers due to the
easiness of implementation and tuning.

Nevertheless, using inverted decoupling (with non-zero dd ij =1), as the apparent process
is the same for such processes, their performance could be easily improved simply
adding decoupling blocks (lead-lags plus time delay blocks). This is important, because
it means that inverted decoupling can be used in most modern DCS systems using the
standard blocks that are already available in their function block library. Thanks to the
easiness of implementation and achieving bumpless transfer, it would be easy to
alternate between a decentralized or centralized control without modifying the PID
parameters. For this reason, inverted decoupling may be a good and easy way to
improve the performance of industrial multivariable processes with interaction
problems.

3. Decoupler realizability

The realizability requirement for the decoupler is that all of its elements must be proper,
causal and stable. For processes with time delays or non-minimum-phase zeros, direct
calculation of the decoupler can lead to elements with prediction or right-half-plane
(RHP) poles. This problem is discussed in [23] for inverted decoupling of 2x2
processes.

In this section, the conditions that a specified configuration needs to satisfy in order to
be realizable are discussed. In addition, the constraints on the apparent process are
indicated. There are three aspects to take into account and to be inspected for each row:

- Non-causal time delays θi j must be avoided in decoupler elements. If g ik is the

transfer function of the row i with the smallest time delay θik , the element dd ki of Dd
should be different from zero. In addition, the time delay of the q i apparent process ( q q )
i

must fulfill (25):

qqq
ik ≤ qi ≤ min(
j≠k ij ) (25)

- Decoupler elements must be proper, that is, the relative degrees r ij must be greater
than or equal than zero. Similar to the case of time delays, if g ik is the transfer function
of the row i with the smallest relative degree r ik , the element dd ki of Dd should be
different from zero. In addition, the relative degree of the q i apparent process (r qi ) must
fulfill (26):

rik ≤ rqi ≤ min(rij ) (26)

j≠k

- When some transfer function g im has a RHP zero, the element dd mi of Dd should not
be different from zero in order to avoid this zero becoming a RHP pole in some
decoupler element. When the same zero appears in all elements of the same row, it is
necessary to check the multiplicity ηij in each element. Again, similar to the other cases,
if g ik is the transfer function of the row i with the smallest RHP zero multiplicity ηik , the

- 10 -
element dd ki should be non-zero. This RHP zero must appear in the q i apparent process
with a multiplicity ( ηq ) that fulfils:
i

ηik ≤ ηqi ≤ min(

j≠k
ηij ) (27)

From (25), (26) and (27), it is found that when the smallest value (time delay, relative
degree or RHP zero multiplicity) is shared by two or more transfer functions of the
same row, there are more possibilities for the configuration, but the flexibility (time
delay or relative degree) of the apparent process q i is limited to this common smallest
value of the row.

When two or more elements of Dd must be selected in the same column to satisfy the
previous conditions in all rows, there are no realizable configurations. Then, it is
necessary to insert an additional diagonal block N(s) between the system G(s) and the
inverted decoupler D(s) in order to modify the process and to force the non-realizable
elements into realizability. Then, inverted decoupling can be applied to the new process
GN(s)=G(s)·N(s).

N(s) is a diagonal block with the necessary extra dynamics. If there are no realizability
problems in row i, the N(i,i) element is equal to the unity. If the non-realizability comes
from an element with a non-causal time delay, an additional time delay ( e−θ s ) is inserted
i

in the corresponding diagonal element of N(s) to obtain realizability. If it comes from a

RHP zero z, which has become an unstable pole, the following element is used in N(s):

ηi
 −s + z 
 * 
(28)
 s+z 

where z* is the complex conjugate of z. If it comes from a properness problem, a simple

stable pole with the adequate multiplicity can be inserted as follows:

1
(29)
( λ s + 1)
ri

For illustration, considering the following example [5]

 e −2 s −e −6 s 
 
s+2 s+2
G(s) =   (30)
 ( s − 0.5 ) ⋅ e −3 s ( s − 0.5) ⋅ e−8 s
2

 
 ( s + 2 )2 2 ( s + 2)
3

 

This process has a multivariable RHP zero at s=0.5. Nevertheless, it is associated with a
single output, the second one, and therefore, inverted decoupling can be applied.
However, this RHP zero appears in the two process transfer functions of the second row
with different multiplicity. According to the previous RHP zero condition, the element
dd 21 should be selected to be non-zero in the Dd matrix because element G(2,1) has the
smallest RHP zero multiplicity. In addition, it has the smallest time delay of the second

- 11 -
row. In the first row, due to time delay condition, the element dd 11 should be selected to
be non-zero in the Dd matrix. Since elements dd 11 and dd 21 are in the same column, no
configuration is initially realizable. To achieve realizability, an extra time delay of 4
units has to be added in the first input. In this case, the new process to be decoupled is
given by (31), and using configuration 2-1, the element dd 12 can be selected in the first
row. Then, according to (11), the decoupler matrices are given by (32) and the apparent
decoupled process is composed by off diagonal elements of (31). The RHP zero appears
in the apparent process of the second output, which is necessary for internal stability.

 e −6 s −e −6 s 
 
s+2 s+2
G (s) = 
N  (31)
 ( s − 0.5 ) ⋅ e −7 s ( s − 0.5) ⋅ e−8 s
2

 
 ( s + 2 )2 2 ( s + 2)
3

 

1 0 
0 1  
Dd=   Do=  0 −0.5 ( s − 0.5 ) ⋅ e  −s
(32)
1 0  
 ( s + 2) 

Generally, it is preferable to add the minimum extra dynamics. Therefore, after

checking the necessary additional dynamics of each configuration, we choose the one
with fewer RHP zeros or time delays in N(s).

The problem of determining the minimal N(s) for a given configuration can be
formulated as a linear programming problem. For instance, the next example of a 3x3
process (33) shows the procedure when it is only necessary to add extra time delays in
N(s). This process, the Tyreus distillation column [26], does not have any
configurations in which all inverted decoupling elements are realizable. According to
the previous condition, the elements dd 11 , dd 23 and dd 33 should be selected to be non-
zero in the Dd matrix because they have the smallest time delays by row. However, two
of them are in the same column, so no configuration is initially realizable.

 
 1.986e 
-0.71s
-5.24e-60 s -5.984e-2.24 s
 66.7 s + 1 400 s + 1 14.29 s + 1 
 
G T ( s ) =  
-0.0204e-0.59 s 0.33e-0.68 s -2.38e-0.42 s
2 
(33)
 ( 7.14 s + 1) ( 2.38s + 1) ( + ) 
2 2
1.43 s 1
 -0.374e-7.75 s 11.3e-3.79 s 9.811e -1.59 s 
 
 22.22 s + 1 ( 21.74s + 1)
2
11.36 s + 1 


To force realizability, the diagonal block of extra time delays

N(s)=diag( e − n1s , e − n2 s , e − n3s ) is proposed, and the goal is to specify the smallest n i ≥0 that
obtains realizability for the inverted decoupling. The time delay matrix Θ of the new
process G(s)·N(s) is given by (34).

- 12 -
  0.71 + n1 60 + n2 2.24 + n3 
 
=
Θ  0.59 + n1 0.68 + n2 0.42 + n3  (34)
 7.75 + n 3.79 + n 1.59 + n3 
 1 2

Configuration 1-2-3 is selected, and consequently, the diagonal elements of (34) should
be the smallest. According to (17), where the same configuration is used, the time delay
matrix of the Do matrix is given by (35), and its time delays must be equal to or greater
than zero in order to obtain realizability in the decoupler ( Θ Do ≥ 0 ).

 0 59.29 + n2 − n1 1.53 + n3 − n1 
 
Θ Do =
 −0.09 + n1 − n2 0 −0.26 + n3 − n2  (35)
 6.16 + n − n 2.2 + n2 − n3 
 1 3 0 

This condition can be represented as follows:

 −1 1 0   −59.29 
   
 −1 0 1  n   −1.53 
 1 −1 0   1   0.09 
A·Θ N ≥ b ⇔  · n2  ≥   (36)
 0 −1 1  n   0.26 
 1 0 −1  3   −6.16 
   
 0 1 −1  −2.2 

Then, the problem can be formulated as the following linear programming minimization
(37), which obtains the smallest sum of time delays.

min ∑ ni
(37)
subject to : − A·Θ N ≤ −b, ni ≥ 0

For the example under study, the solution is n 1 =0.09, n 2 =0 and n 3 =0.26, and
configuration 1-2-3 is the only one that achieves realizability. Using other possible
configurations, there is no feasible region in problem (37), independent of N(s).

This procedure can be similarly applied to realizability problems that arise from relative
degrees or RHP zeros. When the inclusion of an additional block N(s) is necessary to
obtain realizability, one of the main inconveniences is a loss of the practical advantages
of the inverted decoupling, as mentioned in [11].

4. Examples

In this section, the proposed methodology is applied to three simulated processes of

different dimensions. Additionally, its effectiveness is verified in a real quadruple tank
plant.

4.1. Example 2x2: An industrial-scale polymerization reactor

This process [7] is an industrial-scale polymerization reactor given by (38), where the
time scales are in hours. The two controlled variables are measurements of the reactor

- 13 -
condition, and the two manipulated variables are the setpoints of the two reactor feed
flow loops.

 22.89 ⋅ e −0.2 s −11.64 ⋅ e −0.4 s 

 
4.572 s + 1 1.807 s + 1
G R (s) =   (38)
 4.689 ⋅ e −0.2 s 5.80 ⋅ e −0.4 s 
 
 2.174 s + 1 1.801s + 1 

In this case, due to the time delays of the system, there are no realizable configurations
for inverted decoupling. To obtain realizability, it is necessary to add an extra time
delay associated with the first input. The diagonal block N(s) is given by n 11 (s)=e-0.2s
and n 22 (s)=1. Then, the new apparent process is the following:

 22.89 ⋅ e −0.4 s −11.64 ⋅ e −0.4 s 

 
4.572 s + 1 1.807 s + 1
G R (s) = 
N
 (39)
 4.689 ⋅ e −0.4 s 5.80 ⋅ e −0.4 s 
 
 2.174 s + 1 1.801s + 1 

As there are no RHP zeros in the determinant of G(s), inverted decoupling is applicable.
If configuration 1-2 is chosen and the two elements of the Dd(s) matrix are fixed to one,
according to Table 1, the apparent processes q 1 (s) and q 2 (s) are given by g 11 (s) and
g 22 (s) of (39). The other two decoupler elements of Do(s) are collected in Table 2.

After determining the decoupler network, the parameters of the decentralized controller
can be tuned independently for the corresponding q i (s). Therefore, the existing SISO
PID tuning methods can be directly applied to guarantee the stability and performance
of each loop. In this work, gain margin specifications are used to tune the PID
controllers according to the tuning rules in [27]. A gain margin of 5 is specified for both
loops. The parameters of the two resulting PI controllers are shown in Table 2.

PI parameters
Method Decoupler IAE 1 IAE 2 μ RS μ RP
K p1 K p2 T i1 T i2
Dd=I

 11.64 ( 4.572 s + 1 ) 
 
0
22.89 ( 1.807 s + 1 )
Proposed
0.157 0.244 4.57 1.8 1.27 1.27 0.2 0.8
(n 11 =e-0.2s) Do=  
 −4.689 (1.801s + 1) 0

 5.8 ( 2.174 s + 1) 


K =
(
 0.3137 1+ 1
4.572s
) (
0.2203 1+
2.174s
1
) 

( ) ( ) 
Xiong - - - - 1.31 1.22 0.39 1.23
 −0.0369 1+ 1 1


0.2439 1+
1.807s 1.801s

 −1.237 ( 2.174 s + 1) e −0.2 s

T

 
1
Simplified ( 1.801s + 1 )
decoupling D=   -0.066 0.119 3.35 1.24 1.34 1.3 0.2 0.81
 0.508 ( 4.572 s + 1) e −0.2 s 1

 ( 1.807 s + 1 ) 

- 14 -
 Lee
(λ 1 =1.3, - 0.133 0.19 6.47 2.61 4.46 2.03 0.3 1.34
λ 2 =1.3)

Table 2 – Different control methods for Example 1 with performance and robustness
indices

Next, in order to verify the nominal control system performance, the closed loop system
responses are shown in Figure 6. There are unit step changes at t=1 h, in the first
reference, and at t=25 h, in the second one. The IAE of each loop is obtained as
performance index and collected in Table 2. For comparison, other control
methodologies are also shown in the figure: the centralized PID control of Xiong [7], a
multiloop PI controller based on Lee’s method [28], and a simplified decoupling using
the third configuration in [10]. They are tuned trying to achieve similar performance (a
gain margin of 5 in each loop). All parameters of these other controllers are shown in
Table 2 with the corresponding IAE indices. With regard to simplified decoupling, it is
interesting to note the greater complexity of the apparent decoupled processes (40).

−(13.92 s 2 + 12.56 s + 2.685) ⋅ e −0.4 s

q simp1 (s)=
s 3 + 1.327 s 2 + 0.5498s + 0.06721
(40)
(5.996 s 2 + 5.411s + 1.157) ⋅ e −0.4 s
q simp2 (s)= 3
s + 1.569 s 2 + 0.8172 s + 0.1413

INSERT HERE FIGURE 6

The proposed design presents better performance. In the outputs, the inverted
decoupling is better in terms of overshoot, decoupling and settling time. It also has the
best average IAE index. The control signals are less aggressive, and their peaks are
smaller. The response of simplified decoupling is quite similar.

In order to evaluate the robustness of the controllers, a μ-analysis is carried out in

presence of diagonal multiplicative input uncertainty. Multiplicative input uncertainty is
represented as illustrated in Figure 7, where Δ I (s) is the disturbance and W I (s) and
W P (s) are the diagonal weights for uncertainty and performance, respectively. In this
example, the selected weights are

(0.1s+0.2)
WI (s)=w I (s)·I= ·I
0.05s+1
(41)
(s/2.2+0.3)
WP (s)=w P (s)·I= ·I
s

The weight w I (s) can be loosely interpreted as the process inputs increase by up to
200% uncertainty at high frequencies and by almost 20% uncertainty in the low
frequency range. The performance weight w P (s) specifies integral action, a maximum
peak for σ ( S ) of M s =2.2 and a bandwidth of about 0.3 rad/s.

INSERT HERE FIGURE 7

- 15 -
 To achieve robust stability the necessary and sufficient condition [16] is

=µRS µ [ -WI (s)TI (s)] < 1 ∀ω (42)

where μ is the structured singular value (SSV) and T I (s)=K(s)G(s)(I+K(s)G(s))-1 is the

input complementary sensitivity function. To evaluate if the closed loop system will
respect the desired performance even in presence of diagonal multiplicative input
uncertainty, the necessary and sufficient condition [16] is

 -WI (s)TI (s) -WI (s)K(s)S(s) 

µRP µ < 1 ∀ω
WP (s)S(s) 
(43)
 WP (s)S(s)G(s)

where S(s)= (I+G(s)K(s))-1 is the sensitivity function and T(s)=G(s)K(s)(I+G(s)K(s))-1

is the complementary sensitivity function.

Figure 8 shows the SSV for robust stability and robust performance for the different
controllers. The robust stability (RS) is clearly smaller than one for all frequencies,
indicating that the systems will remain stable in spite of an uncertainty of 20% on each
process input. The peak values are shown in Table 2. The proposed method and the
simplified decoupling have the smallest values. The robust performance analysis shows
that the proposed method and simplified decoupling satisfy the robust performance (RP)
condition (43). For the other controllers, the performance will deteriorate at frequency
range of [1-10] rad/s, where the peaks appear. These values are also collected in Table
2. The multiloop PI controller of Lee has a good robust stability; however, it shows the
worst robust performance, even at low frequencies, due to the strong interactions in the
first loop.

INSERT HERE FIGURE 8

4.2. Example 3x3: Tyreus distillation column

This process [26] is a sidestream column separating a ternary mixture, where a feed
containing 10 percent benzene, 45 percent toluene and 45 percent o-xylene is separated
in a single column into three product streams. The transfer function matrix is given by
(33), where the controlled and manipulated variables are y 1 (toluene impurity in the
distillate); y 2 (benzene impurity in the sidestream); y 3 (toluene impurity in the bottom);
u 1 (reflux ratio); u 2 (sidestream flow rate); u 3 (reboil duty).

As was explained in Section 3, there are no realizable configurations for this process
due to the time delays. Thus, it is necessary to include an additional block N(s) with
delays. To achieve realizability by adding the minimum quantity of delays, the best
choice is configuration 1-2-3 with n 11 (s)=e-0.09s, n 22 (s)=1 and n 33 (s)=e-0.26s. The new
apparent process to design the inverted decoupling is:

- 16 -
  −0.8 s

 1.986e −5.24e −60 s −5.984e −2.5 s 
 66.7 s + 1 400 s + 1 14.29 s + 1 
 
−0.0204e −0.68 s 0.33e −0.68 s −2.38e −0.68 s
G TN ( s ) =   (44)
 ( 7.14 s + 1) ( 2.38s + 1) (1.43s + 1) 
2 2 2

 −0.374e −7.84 s 11.3e −3.79 s 9.811e −1.85 s 

 
 ( 21.74s + 1) 11.36 s + 1 
 22.22 s + 1
2

Therefore, using configuration 1-2-3 and fixing the three elements of the Dd(s) matrix
to one, the open loop processes are the diagonal elements of (44), and the decoupler
matrix Do(s) can be calculated according to (17), obtaining:

 2.638 ( 66.7 s + 1) e 3.013 ( 66.7 s + 1) e 

−59.2 s −1.7 s

 0 
 ( 400 s + 1) (14.29 s + 1) 
 0.062 ( 2.38s + 1)2 7.212 ( 2.38 s + 1)
2 
Do(s) =  0  (45)
 ( 7.14 s + 1)2 (1.43s + 1) 2

 
 0.038 (11.36 s + 1) e −1.213 (11.36 s + 1) e
−5.99 s −1.94 s

 22.22 s + 1 ( 21.74 s + 1)
0 
 
2

Although the process is a 3x3 system, the complexity of the elements of (45) is as
simple as that of the inverted decoupler elements for 2x2 processes. Three PI controllers
are independently tuned for the corresponding decoupled processes q i (s) based on the
IMC rules of Rivera [29]. Closed-loop time constants of 15, 20 and 18 s are specified in
each loop, respectively. The parameters are listed in Table 3.

PI parameters
Method IAE 1 IAE 2 IAE 3 μ RS μ RP
K p1 K p2 K p3 T i1 T i2 T i3
Proposed
2.25 0.77 0.07 67.1 5.1 12.3 15 21.3 22.9 0.28 0.86
(n 11 =e-0.09s, n 33 =e-0.26s)

Wang - - - - - - 38.7 31 38.3 0.22 1.24

Liu - - - - - - 17.6 25.1 22.9 0.23 0.77

Lee
2.14 -0.04 0.13 61.1 2.39 37.5 65.1 22.8 133 0.31 1.8
(λ 1 =15,λ 2 =20,λ 3 =18)

Table 3 – PI parameters, performance and robustness indices for the different control
methods for Example 2

The closed loop system responses (outputs and control signals) are shown in Figure 9.
There are unit step changes at t=0 s in the first reference, at t=333 s in the second one
and at t=666 s in the third one.

For comparison, other control methodologies are also shown in the figure: the pure
centralized control of Wang in [18], the analytical decoupling control of Liu in [30], and
a multiloop PI controller. Some elements of the controllers of Wang, K(s), and Liu,
C(s), are shown in (46) and (47), respectively. In them, all of the elements have a high
order in the numerator and denominator, about third, fourth, fifth or sixth order plus
time delay. Their higher complexity is clear in comparison with the elements of the

- 17 -
 proposed control in (45). The multiloop PI controller is tuned based on Lee’s method
[28], and using the same closed-loop time constants as the proposed control. According
to the relative gain array (RGA), the chosen pairing has been y 1 -u 1 , y 2 -u 3 and y 3 -u 2 .
The PI parameters are collected in Table 3 with the performance and robustness indices
of the different methodologies.

−69.82 s + 10.68s + s + 0.01277

3 2

k 21 ( s ) = e−
10.52 s
,
1041s + 126.5s + 22.49 s
3 2

24260 s − 10870 s + 2651s +

6 5 4

−110 s + 55.22 s + s + 0.0006456

3 2

k12 ( s ) = e−
8.376 s
,
−57140 s − 17230 s − 3031s +
6 5 4
(46)
−376.2 s − 23.97 s − 0.05722 s
3 2

−32.71s − 3.641s + s + 0.03188

3 2

k33 ( s ) =
4964 s + 411s + 102.3s
3 2

14543s 2 + 256.3578s + 0.5502

c11 ( s ) = f1 e −0.09 s ,
(15s + 1)( 438.7353s + 1)
4773900 s 6 − 6620600 s 5 − 3286200 s 4
−532380 s 3 − 41045s 2 − 523.1791s − 0.296
c12 ( s ) = f 2
−3.73 s
e ,
(12 s + 1) ( 611700 s + 109510 s
2 4 3

+12128s 2 + 465.9313s + 1)
16790 s + 1582.9 s + 39.2646 s + 0.0885
3 2

c23 ( s ) = f3 , (47)
(18s + 1)( 511.4853s + 440.0233s + 1)
2


1 1 1
=f1 = , f2 = , f3
e −0.8 s e −0.68 s e −1.85 s
1− 1− 1−
15s + 1 (12 s + 1)2 18s + 1

INSERT HERE FIGURE 9

The proposed design achieves perfect decoupling performance without overshoot or

inverse response in the outputs; although the control signals are more oscillatory than in
Wang’s control. In addition, it obtains the smallest IAE. The response is quite similar
than that of Liu’s controller. Nevertheless, it is important to note that the complexity of
the controller of Wang or Liu is much greater than that of the proposed control. In
addition, these design procedures are more complex than that of the proposed inverted
decoupling. The response of the multiloop controller is very good in the second loop;
however, it shows important interactions in the other outputs.

In order to investigate the robustness of the controllers, a μ-analysis is carried out as it

was in the previous example. The selected weights are given by (48). Figure 10 shows
the SSV for robust stability and robust performance for the different controllers. The
robust stability is easily satisfied by all of controllers. The smallest values are achieved
by the methods of Wang and Liu, although all values are very close. The robust
performance analysis shows that the Liu’s control and the proposed controller satisfy

- 18 -
the robust performance condition (43). In spite of its high complexity, the performance
of Wang’s controller will deteriorate at frequency range of [0.01-0.1] rad/s, where a
peak appears. The multiloop PI controller has a good robust stability; however, due to
the strong interactions, it shows the worst robust performance at frequencies below 0.1
rad/s.

(2s+0.2)
WI (s)=w I (s)·I= ·I
s+1
(48)
(s/2.75+0.025)
WP (s)=w P (s)·I= ·I
s

INSERT HERE FIGURE 10

4.3. Example 4x4: HVAC process

This interactive 4x4 process is the experimental centralized HVAC system of four
rooms that appears in [31]. Air temperature control for four rooms is a complicated
multivariable control problem because there are interactions among the air flow rates to
the different rooms. The air temperature in the four rooms is controlled by regulating
the position of four variable-air-volume (VAV) dampers. The 4x4 model is:

 −0.098 ⋅ e −17 s −0.036 ⋅ e −27 s −0.014 ⋅ e −32 s −0.017 ⋅ e −30 s 

 
 122 s + 1 149 s + 1 158s + 1 155s + 1 
 −0.043 ⋅ e −25 s −0.092 ⋅ e −16 s −0.011 ⋅ e −33 s −0.012 ⋅ e −34 s 
 147 s + 1 130 s + 1 156 s + 1 157 s + 1 
G H (s) =   (49)
 −0.012 ⋅ e −31s −0.016 ⋅ e −34 s −0.102 ⋅ e −16 s −0.033 ⋅ e −26 s 
 153s + 1 151s + 1 118s + 1 146 s + 1 
 −32 s

 −0.013 ⋅ e −0.015 ⋅ e −31s −0.029 ⋅ e −25 s −0.108 ⋅ e −18 s 
 
 156 s + 1 159 s + 1 144 s + 1 128s + 1 

In [31], using the normalized decoupling control system design rules, an apparent
decoupled process (50) is proposed. Then, a normalized decoupling network is designed
using the ETF method [14]. By selecting a gain margin of 5 and a phase margin of 2π/5
rad in all of the loops, a diagonal PI controller is designed. The parameters are collected
in Table 4.

 e −21.82 s e −21.32 s e −22.21s 

e −23.12 s
Q H (s) = diag  , , ,  (50)
113.83s + 1 121.37 s + 1 113.9 s + 1 123.55s + 1 

In process (49), the only configuration for inverted decoupling that achieves
realizability without adding extra dynamics is configuration 1-2-3-4. Using this
configuration, the proposed extended approach of inverted decoupling is applied in such
a way that the same apparent process (50) is specified in (19) and (20), resulting in the
following Dd and Do matrices:

 −10.2 (122 s + 1) −10.87 (130 s + 1) −9.8 (118s + 1) −9.26 (128s + 1) 

Dd(s) = diag   (51)
−4.82 s −5.32 s −6.21 s −5.12 s
e , e , e , e
 (113.88s + 1) (121.4s + 1) (113.92s + 1) (123.5s + 1) 

- 19 -
  ( 4.098 s + 0.036 ) −5.18 s
(1.594 s + 0.014 ) −10.2 s
(1.935 s + 0.017 ) −8.18 s 
e e e
 
0
149 s + 1 158 s + 1 155 s + 1
 
 ( 5.219 s + 0.043 ) e −3.68 s
0
(1.335 s + 0.011)
e
−11.7 s
(1.456 s + 0.012 )
e
−12.7 s

 147 s + 1 156 s + 1 157 s + 1 
Do(s) =
 (1.367 s + 0.012 ) (1.822 s + 0.016 ) ( 3.759 s + 0.033 ) 
 e
−8.79 s
e
−11.8 s
0 e
−3.79 s

 153s + 1 151s + 1 146 s + 1 
 (1.606 s + 0.013 ) (1.853s + 0.015 ) ( 3.583s + 0.029 ) 
 e
−8.88 s
e
−7.88 s
e
−1.88 s

 
0
156 s + 1 159 s + 1 144 s + 1
(52)

In order to achieve the same performance obtained with the normalized decoupling in
[31] and because the apparent decoupled process is practically identical, the same PI
parameters are used in the proposed control. Figure 11 shows the closed loop system
response (outputs and control signals) of the designed inverted decoupling control in
comparison with that of the normalized decoupling of [31]. A multiloop PI controller
based on the Lee’s method [28] is also shown.

PI parameters
Method IAE 1 IAE 2 IAE 3 IAE 4 μ RS μ RP
K p1 K p2 K p3 K p4 T i1 T i2 T i3 T i4

Proposed 1.64 1.79 1.61 1.68 69.4 67.8 70.7 73.5 80.4 78.1 81.7 85.1 0.22 0.97

Normalized
1.64 1.79 1.61 1.68 69.4 67.8 70.7 73.5 87.2 84.5 87.9 90.7 0.24 1.04
Decoupling
Lee - - - -
101.8 108.3 108.2 117.4 99.9 99.1 95.2 95.8 0.41 1.22
(λ i =45, ∀i ) 20.5 23.5 19.3 19.2

Table 4 – PI parameters, performance and robustness indices for the different control
methods for Example 3

INSERT HERE FIGURE 11

There is a unit step change in each reference every 500 seconds, and at t=2000 s, there
is a unit step in all process inputs at the same time as the input disturbance.
The performances of the different methodologies are quite similar with the same settling
time. However, the proposed control achieves perfect decoupling, while the multiloop
control presents some interactions. Normalized decoupling also shows some little
interactions. The IAE indices are collected in Table 4. They are very similar; however,
the proposed control obtains the smallest ones because it achieves perfect decoupling.

Although the complexities of both decouplers are practically the same, another
advantage of the proposed methodology over normalized decoupling is its direct method
of carrying out the decoupler network design. In the normalized decoupling design, the
procedure is slightly more complex with the calculation of the normalized gain matrix,
the RGA, the RNGA and the RARTA.

Using the weights in (53), a μ-analysis in presence of diagonal multiplicative input

uncertainty shows that the three controllers have similar robust stability and robust
performance. Results are depicted in Figure 12 and the peak values are collected in

- 20 -
Table 4. The multiloop control has a worse robust performance in a small range of
frequencies due to the interactions.
(10s+0.2)
WI (s)=w I (s)·I= ·I
5s+1
(53)
(s/2+0.008)
WP (s)=w P (s)·I= ·I
s

INSERT HERE FIGURE 12

4.4. Experimental process: quadruple tank system

The experimental process is a quadruple tank plant [32] in the lab of the Computer
Science Department of the University of Córdoba. The outputs are the level of the lower
tanks inside the range of [0-35] cm, and the inputs are the flow references of the
secondary control loops that regulate the operation of the pumps, in the range of [0-200]
cm3/s. The plant was configured in order to show interaction problems without having
multivariable RHP zeros, and then, the process was identified around the operation
point h=[20 20] cm and u=[135 135] cm3/s. The resultant model is given by (54), and it
has a RGA of 2.29.

 0.3284 0.2454 
 184.5s + 1 (184.5s + 1)( 535.1s + 1) 
G Q (s) =  (54)
 0.2457 0.3378 
 
 (185s + 1)( 503.2 s + 1) 185s + 1 

Due to relative degrees, configuration 1-2 must be chosen for realizability without
adding extra dynamics. If the Dd(s) matrix is fixed to the unitary matrix, according to
Table 1, the apparent processes q 1 (s) and q 2 (s) are given by g 11 (s) and g 22 (s) of (54),
and the other two decoupler elements of Do(s) are:

-g12 -2454
do12 = =
g11 3284 ( 535.1s + 1)
(55)
-g 21 -2457
=
do =
g 22 3378 ( 503.2 s + 1)
21

After determining the decoupler elements, the parameters of a PI decentralized

controller are independently tuned for the corresponding q i (s) using the IMC tuning
rules of [29]. A closed loop time constant of 120 s is specified in both loops to obtain a
settling time of about 600 s. The parameters of the two resulting PI controllers are
shown in Table 5 with the performance and robustness indices.

PI parameters
Method IAE 1 IAE 2 μ RS μ RP
K p1 K p2 T i1 T i2
Proposed 4.68 4.56 184.4 185 1002 1.27 0.35 0.97

Decentralized 4.68 4.56 184.5 185 1895 1.22 0.37 0.99

- 21 -
 Table 5 – PI parameters and performance and robustness indices for a quadruple tank
process

Figure 13 shows the resultant response of the closed loop system for a step of 5 cm in
the references. For comparison, we also show the response of the decentralized
controller obtained only with the PI controllers in Table 5, without the decoupler
elements (55). Therefore, it is like two independent SISO controllers, one for g 11 (s) and
the other for g 22 (s).

With inverted decoupling, a better response is achieved with a very good decoupling
performance and a smaller settling time in both loops. The decentralized control reaches
the references later, and the rejection of the interactions is very slow, and therefore, the
corresponding IAE indices are greater than those of the proposed control. Using the
model in (54), a μ-analysis in presence of diagonal multiplicative input uncertainty is
carried out. The robust stability and robust performance indices in Table 5 for both
controllers are quite similar.

INSERT HERE FIGURE 13

5. Conclusions

A new generalized approach of the inverted decoupling technique for nxn processes has
been developed in this work. The problem is approached from a compact matricial
formulation, and we demonstrate that the transfer functions of the decoupler elements
and the decoupled equivalent open loop process have the same complexity, regardless
of the system size. This is a great advantage over other conventional decoupling
methods. In addition, the methodology allows for more flexibility in choosing the
decoupled apparent processes, and since they are usually very simple, the tuning of
decentralized controllers is much easier. Therefore, it is concluded that the proposed
methodology has important design advantages that make it very easy to apply when the
process does not have multivariable RHP zeros.

A study of all possible configurations for inverted decoupling was carried out, and the
realizability conditions for applying this decoupling method were expounded. The
expressions of inverted decoupling for 2x2 and 3x3 processes are shown in more detail.
The methodology is illustrated with three simulation examples of different sizes (2x2,
3x3 and 4x4). PI decentralized controllers were used to control the apparent decoupled
processes. The system time responses are compared with other works, demonstrating
that this decoupling achieves similar or better performance. Additionally, an
experimental lab process of four tanks was used to verify the effectiveness of this
methodology.

Acknowledgments

This work was supported by the Spanish CICYT under Grant DPI 2007-62052. This
support is very gratefully acknowledged. Moreover, J. Garrido expresses appreciation
for the FPU fellowship (Ref. AP2006-01049) from the Spanish Ministry of Education.

References

- 22 -
[1] F. C. Shinskey, Sistemas de control de procesos. Aplicación, diseño y sintonización. Tomo 1,
McGrawhill, 1996.
[2] F. Vázquez, F. Morilla, S. Dormido, An iterative method for tuning decentralized PID
controllers, Proceeding of the 14th IFAC World Congress (1999) 491-496.
[3] F. Vázquez, Diseño de controladores PID para sistemas MIMO con control descentralizado, PhD
Thesis, UNED, 2001.
[4] Q. W. Wang, B. Zou, T. Lee, Q. Bi, Auto-tuning of Multivariable PID Controllers from
Decentralized Feedback, Automatica 33 (1997) 319-330.
[5] Q. W. Wang, Y. Zhang, M. S. Chiu, Decoupling internal model control for multivariable
systems with multiple time delays, Chem. Eng. Sci. 57 (2002) 115-124.
[6] Q. W. Wang, Y. Zhang, M. S. Chiu, None-interacting control design for multivariable industrial
processes, J. Process Control 13 (2003) 253-265.
[7] Q. Xiong, W. J. Cai, M. J. He, Equivalent transfer function method for PI/PID controller design
of MIMO processes, J. Process Control 17 (2007) 665-673.
[8] F. Morilla, F. Vázquez, J. Garrido, Centralized PID Control by Decoupling for TITO Processes,
Proceedings of 13th IEEE International Conference on Emerging Technologies and Factory
Automation (2008) 1318-1325.
[9] J. Garrido, F. Morilla, F. Vázquez, Centralized PID Control by Decoupling of a Boiler-Turbine
Unit, 10th European Control Conference, Budapest (2009) 4007-4012.
[10] K. Waller, Decoupling in distillation, AIChE J. 20 (1974) 592-594.
[11] H. L. Wade, Inverted decoupling: a neglected technique, ISA Trans. 36 (1997) 3-10.
[12] P. Nordfeldt, T. Hägglund, Decoupler and PID controller design of TITO systems, J. Process
Control 16, (2006) 923-936.
[13] S. Tavakoli, I. Griffin, P. J. Fleming, Tuning of decentralised PI (PID) control for TITO
processes, Control Eng. Pract. 14 (2006) 1069-1080.
[14] W. J. Cai, W. Ni, M. J. He, C. Y. Ni, Normalized Decoupling – A new approach for MIMO
process control system design, Ind. Eng. Chem. Res. 47 (2008) 7347-7356.
[15] J. Maciejowski, Multivariable Feedback Design, Addison-Wesley, 1989.
[16] S. Skogestad, I. Postlethwaite, Multivariable Feedback Control: Analysis and Design, John
Wiley & sons, 1996.
[17] G. C. Goodwin, S. F. Graebe, M. E. Salgado, Control System Design, Prentice-Hall, 2001.
[18] Q. W. Wang, Decoupling Control, Lecture Notes in Control and Information Sciences, 285,
Springer-Verlag, 2003.
[19] K. J. Åström, K. H. Johansson, Q. W. Wang, Design of decoupled PI controller for two-by-two
systems, IEE Proc. Contr. Theor. Appl. 149 (2002) 74-81.
[20] F. Vázquez, F. Morilla, Tuning decentralized PID controllers for MIMO systems with
decouplers, Proceeding of the 15th World Congress of the IFAC, Barcelona, Spain (2002).
[21] E. Gagnon, A. Pomerleau, A. Desbiens, Simplified, ideal or inverted decoupling?, ISA Trans. 37
(1998) 265-276.
[22] J. Zheng, G. Guo, Y. Wang, Feedforward decoupling control design for dual-actuator system in
hard disk drives, IEEE Trans. Magn. 40 (2004) 2080-2082.
[23] P. Chen, W. Zhang, Improvement on an inverted decoupling technique for a class of stable linear
multivariable processes, ISA Trans. 46 (2007) 199-210.
[24] R. Garduno, K. W. Lee, Compensation of control-loop interaction for power plant wide-range
operation, Control Eng. Pract. 13, (2005) 1475-1487.
[25] T. Hägglund, K. J. Åström, Revisiting the Ziegler-Nichols step response method for PID control,
J. Process Control 14 (2004) 635-650.
[26] B. D. Tyreus, Multivariable control system design for an industrial distillation column, Ind. Eng.
Chem. Process Des. Dev. 18 (1979) 177-182.
[27] A. O’Dwyer, PI and PID controller tuning design for processes with delay to achieve constant
gain and phase margin for all values of delay, Proceedings of the Irish Signals and Systems
Conference (2001) 96-100.
[28] M. Lee, K. Lee, C. Kim, J. Lee, Analytical design of multiloop PID controllers for desired
closed-loop responses, AIChE J. 50 (2004) 1631-1635.
[29] D. E. Rivera, M. Morari, S. Skogestad, Internal Model Control. 4. PID controller design, Ind.
Eng. Chem. Process Des. Dev. 25, (1986) 252-265.
[30] T. Liu, W. Zhang, F. Gao, Analytical decoupling control strategy using a unity feedback control
structure for MIMO processes with time delays, J. Process Control 17 (2007) 173-186.

- 23 -
[31] Y. Shen, W. Cai, S. Li, Normalized decoupling control for high-dimensional MIMO processes
for application in room temperature control HVAC systems, Control Eng. Pract. 18 (2010) 652-
664.
[32] K. H. Johansson, The quadruple-tank process: a multivariable laboratory process with an
adjustable zero, IEEE Trans. Control Syst. Technol. 8 (2000) 456-465.

- 24 -
Table Captions

Table Captions

Table 1 - Cases of 2x2 inverted decoupling with two unitary elements (configuration 1-
2).

Table 2 – Different control methods for Example 1 with performance and robustness
indices.

Table 3 – PI parameters, performance and robustness indices for the different control
methods for Example 2.

Table 4 – PI parameters, performance and robustness indices for the different control
methods for Example 3.

Table 5 – PI parameters and performance and robustness indices for a quadruple tank
process.
Table1

Decoupler elements Decoupled process

dd11 =dd 22 =1
g 0 
Case 1 -g12 -g Q=  11 
do12 = , do 21 = 21  0 g 22 
g11 g 22
dd11 =do 21 =1
g 0 
Case 2 -g -g Q=  11 
do12 = 12 , dd 22 = 21  0 -g 21 
g11 g 22
do12 =do 21 =1
 -g 0 
Case 3 -g -g Q=  12 
dd11 = 12 , dd 22 = 21  0 -g 21 
g11 g 22
do12 =dd 22 =1
 -g 0 
Case 4 -g -g Q=  12 
dd11 = 12 , do 21 = 21  0 g 22 
g11 g 22
Table2

PI parameters
Method Decoupler IAE1 IAE2 μRS μRP
Kp1 Kp2 Ti1 Ti2
Dd=I

 11.64  4.572 s  1  
 
0
22.89  1.807 s  1 
Proposed
0.157 0.244 4.57 1.8 1.27 1.27 0.2 0.8
(n11=e-0.2s) Do=  
 4.689 1.801s  1  0

 5.8  2.174 s  1 


K 

 0.3137 1+ 1
4.572s
 
0.2203 1+
1

2.174s
 

    
Xiong - - - - 1.31 1.22 0.39 1.23
 0.0369 1+ 1 1


0.2439 1+
1.807s 1.801s

 1.237  2.174 s  1 e 0.2 s

T

 
1
Simplified 1.801s  1
decoupling D=   -0.066 0.119 3.35 1.24 1.34 1.3 0.2 0.81
 0.508  4.572 s  1 e 0.2 s 
1
 1.807 s  1 
Lee
(λ1=1.3, - 0.133 0.19 6.47 2.61 4.46 2.03 0.3 1.34
λ2=1.3)
Table3

PI parameters
Method IAE1 IAE2 IAE3 μRS μRP
Kp1 Kp2 Kp3 Ti1 Ti2 Ti3
Proposed
2.25 0.77 0.07 67.1 5.1 12.3 15 21.3 22.9 0.28 0.86
(n11=e-0.09s, n33=e-0.26s)

Wang - - - - - - 38.7 31 38.3 0.22 1.24

Liu - - - - - - 17.6 25.1 22.9 0.23 0.77

Lee
2.14 -0.04 0.13 61.1 2.39 37.5 65.1 22.8 133 0.31 1.8
(λ1=15,λ2=20,λ3=18)
Table4

PI parameters
Method IAE1 IAE2 IAE3 IAE4 μRS μRP
Kp1 Kp2 Kp3 Kp4 Ti1 Ti2 Ti3 Ti4

Proposed 1.64 1.79 1.61 1.68 69.4 67.8 70.7 73.5 80.4 78.1 81.7 85.1 0.22 0.97

Normalized
1.64 1.79 1.61 1.68 69.4 67.8 70.7 73.5 87.2 84.5 87.9 90.7 0.24 1.04
Decoupling
Lee
-20.5 -23.5 -19.3 -19.2 101.8 108.3 108.2 117.4 99.9 99.1 95.2 95.8 0.41 1.22
(λi=45, i )
Table5

PI parameters
Method IAE1 IAE2 μRS μRP
Kp1 Kp2 Ti1 Ti2

Proposed 4.68 4.56 184.4 185 1002 1.27 0.35 0.97

Decentralized 4.68 4.56 184.5 185 1895 1.22 0.37 0.99

Figure 1. Conventional decoupling control system
Figure 2. Inverted decupling control system of a TITO process.
Figure 3. Matrix representation of inverted decoupling.
Figure 4. Control schemes of the four cases of 2x2
Figure 5. Step response of system (21).
Figure 6. Outputs and control signals of example 1
Figure 7. System with multiplicative input uncertainty
Figure 8. SSV for RS and RP in example 1
Figure 9. Outputs and control signals in example 2
Figure 10. SSV for RS and RP in example 2
Figure 11. Outputs and control signals in example 3
Figure 12. SSV for RS and RP in example 3
 Figure 13. Outputs and control signals in example 4
View publication stats