Journal of Process Control 16 (2006) 923–936

www.elsevier.com/locate/jprocont

Decoupler and PID controller design of TITO systems

Pontus Nordfeldt, Tore Hägglund *

Department of Automatic Control, Lund Institute of Technology, Lund University, Box 118, S-221 00 Lund, Sweden

Received 24 January 2006; received in revised form 12 June 2006; accepted 14 June 2006

Abstract

This paper treats controller design and tuning for systems with two input signals and two output signals in the process industry. Two
design methods that can be combined to form a core in an algorithm for automatic design and tuning for the considered systems are
presented. The proposed controller consists of a decoupler and a diagonal PID controller. The decoupler has the property that as little
dynamics as possible is introduced in the decoupler. The PID design method is based on exhaustive search. The methods are combined
and tested in an industrial environment in a paper mill.
 2006 Elsevier Ltd. All rights reserved.

Keywords: TITO control; Process control; PID control; Multi-variable control; Decoupling

1. Introduction selectors, and sequence functions are also used to obtain

the desired overall control function. These decentralized
Most control systems are complex and multi-variable, approaches have traditionally evolved through years of
i.e. they consist of several measurement signals and control experience. The approaches manage to handle the nonlin-
signals, and there are often complicated couplings between ear effects. On the other hand, the decoupling is often trea-
the different signals. A typical process control plant may ted inefficiently, e.g. by detuning control loops.
contain thousands of input and output signals. There is an unfortunate gap between the centralized
There are multi-variable control techniques available computational approaches of multi-variable control theory
today. For example, model-predictive control (MPC) has and the common practice to design local control loops
gained a lot of interest in recent years, both in academia disregarding couplings and interaction. Today it appears
and in industry. The general approach has proved to be that both approaches has reached a point of refinement
useful, but much remains to be done for its full utilization. where the gap can be reduced from both sides.
One important problem is to limit the computational com- It is very common that different process values interact,
plexity in presence of logical functions and mode switching. which means that the different control loops disturb each
When MPC is used today, it is mostly used on a higher other. The problem of interacting control loops is very
level to give setpoints to the PID controllers that are oper- common in process control. Today, the problem is solved
ating on the basic level. by tuning the most important loop to give good perfor-
Process control problems are traditionally solved using mance, while the other loop is detuned in such a way that
single-loop PID controllers that are connected through the interaction with the first loop becomes acceptable. This
well-known couplings such as cascade control, feed-for- solution is, of course, far from optimal. One of the loops is
ward control, ratio control, split-range control, etc. Logic, forced to perform poorly in order to compensate for the
interaction.
*
Corresponding author. Tel.: +46 46 2228798; fax: +46 46 138118.
For this reason, a research project has been initiated
E-mail addresses: [email protected] (P. Nordfeldt), tore@con- where the goal is to develop a two-times-two controller
trol.lth.se (T. Hägglund). which manages to decouple the control of a TITO (two

0959-1524/$ - see front matter  2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2006.06.002
924 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936

input two output) system. The goal is also that the control-
−1
ler should be useful in an industrial environment, which
means that the nonlinear functions for anti-windup and
mode switches should be worked out. Another prerequisite
is that the controller must be easy to operate, which means Σ PID
that controller parameters and decoupling functions must
Decoupler TITO system
be determined automatically.
A natural question is – Why only two times two? There Σ PID
are at least two good reasons for restricting the controller
to two inputs and two outputs. One reason is the complex-
ity. The complexity of the nonlinear functions of the con-
troller, such as saturations, increases drastically when the −1
number of signals is increased. Another reason is that
Fig. 1. Block diagram of the proposed control structure for TITO
many control problems consists of two variables that inter- systems.
act, and these problems can be solved efficiently with a two
times two controller.
This paper summarizes the first part of the project. In pler, and the outputs from the decoupler form the two
the next section, a more detailed problem description is inputs to the process. With the structure depicted in
presented. A new decoupling procedure has been devel- Fig. 1 the control problem is separated into two parts,
oped, with the feature that as little dynamics as possible one part that concerns decoupling and a second part that
is introduced in the control loops by the decoupler. It is concerns control of decoupled loops. It turns out that it
presented in Section 3. There are many methods available is possible to find automatic design methods for these
for design of PID controllers, but most of them fail to han- two parts. The main contribution of this paper is that it
dle decoupled systems. The reason for this is discussed in provides such methods.
Section 4, which also gives a new design procedure suitable
for the project. The new decoupler and design procedure 2.1. General disturbances in the process industry
have been tested in a paper mill. The results are presented
in Section 5. More details are given in the thesis [8]. In the process industry (for which this work is aimed)
robustness and disturbance attenuation are often the pri-
2. Problem description mary objective of control, and this should be reflected in
the specifications on which the design method is based.
Many TITO systems found in the process industry have Below is a general description of what disturbances we
an additional property apart from being TITO systems. should expect on a process in the process industry, and a
They are close to being linear square stable non-singular short discussion.
systems as defined in Definition 1. The feedback system with input signal, output signal
and disturbances is depicted in Fig. 2. A similar figure
Definition 1 (Linear square stable non-singular systems). A and a discussion of specifications are present in [14].
linear square stable non-singular system is stable and has The load disturbance (or process-input disturbance) l is
the same number of input signals and output signals. It is generally a low-frequency disturbance and process output
linear and can be represented by a linear square stable non- disturbances n and measurement disturbances m are often
singular transfer-function matrix. That the transfer-func- high-frequency disturbances.
tion matrix is non-singular means here that it is not It is undesirable to have high-frequency disturbances
singular for any positive finite frequency on the real axis. amplified in the control signal u since it may damage the
The work presented in this paper is focused on TITO actuator and/or give a high-energy consumption. Thus,
systems. When they are mentioned in the text it is assumed the transfer functions from n and m to u
that they have the properties stated in Assumption 1. Gnu ¼ Gmu ¼ CðI þ GCÞ1 ;
Assumption 1 (TITO systems). It is assumed that the
treated TITO systems are linear square stable non-singular n
l
systems with two input signals and two output signals. y
r u
C G
Most systems in process control are nonlinear, but they −
can be linearized using, e.g. gain scheduling.
The first step of this project was to find a suitable con-
trol structure. The structure depicted in Fig. 1 is sometimes
used for control of TITO systems. Two PID controllers, m
that takes the two measurement signals as inputs, are used. Fig. 2. Block diagram of the closed-loop system with relevant
The output signals from the PID controllers enter a decou- disturbances.
 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936 925

where G is the process transfer function and C is the control- Σ C D G

ler transfer function, should have low high-frequency gain.
Ideally, all disturbances in the output signal y should be
attenuated. Thus, the transfer functions from n, m, and l
to y −1

Gny ¼ ðI þ GCÞ1 ; Fig. 3. Block diagram of the closed-loop system.

1
Gmy ¼ GCðI þ GCÞ ;
1
Gly ¼ ðI þ GCÞ G;
• The product GD should be diagonal.
should be small or at least bounded. • It must not be of high-pass character.
The transfer function Gny is often called the output sen- • It should contain as little dynamics and time delay as
sitivity function S and the transfer function Gmy is often possible.
called the output complementary sensitivity function T [14]. • It should fulfill the special requirements of automatic
methods presented in the previous section.
2.2. Stability and a stability margin
Below follows a matrix theory description of the prob-
Stability of the closed-loop system is, of course, neces- lem and an answer to the question of how the decoupler
sary but it is also necessary that the system be robust should be chosen. Definition 2 and Proposition 1 are well
enough not to be pushed over to instability by small mod- known and documented in the literature, see for example
eling errors or small nonlinearities. It is thus necessary to [6].
have a measure of the system stability margin, a measure
Definition 2. The adjoint of a matrix A, denoted adj(A), is
of how far it is from instability.
the transposed matrix of cofactors of A
2.3. Special demands on automatic methods ðadjðAÞÞij ¼ Aji ;
where Aji are the cofactors of A.
The specification that the method should work as an
automatic method gives rise to some special demands. If Proposition 1
a design method or a decoupling method is to be used
for automatic tuning it must be simple. In this context this A  adjðAÞ ¼ adjðAÞ  A ¼ detðAÞ  I:
means that the method should not require qualitative The first step in the search for a decoupler design
choices to be made by the user. It is, for example, not good method is to find an expression for the whole space of
if approximations that depend on the process model struc- possible decouplers, from which a suitable one could
ture have to be done. It is not a problem if the method be chosen. Proposition 2 gives an expression for this
requires complicated computations, as long as these can space.
be performed by a computer in a numerically stable way
without qualitative choices and in a limited amount of Proposition 2. All matrices D that make the system GD
time. If the method requires parameter tuning by an oper- diagonal can be factorized as the adjoint of G times a
ator, there must be default values that always work fairly diagonal matrix K. D is then given by
well.
D ¼ adjðGÞ  K:
2.4. Previous work
Proof. The proposition follows directly from Proposition
1. h
There has been a lot of research in the area of PID con-
troller design for several years, and during very recent years Proposition 2 may seem trivial but it is important
some good results have been obtained for SISO (single input because it shows the whole space of possible decouplers.
single output) systems [9,1]. Attempts have also been made The question is how K should be designed to satisfy the
to find results for multi-variable systems [2,10–13]. None of requirements stated above and in [8].
these attempts fulfil the requirements of this project. More All real processes attenuate sufficiently high frequencies
attention is payed to this subject in Sections 3.2 and 4.1. and can thus be represented by models of low-pass charac-
ter. This means that every component of the transfer-func-
3. The decoupler design method tion matrix of the process is of low-pass character. This, in
turn implies that all cofactors of the transfer-function
The structure of the closed-loop system is depicted in matrix and thus the adjoint of the process transfer-function
Fig. 3. A design method for the decoupler D is described matrix is of low-pass character. Below K = I and thus
in this section. The decoupler D should satisfy the follow- D = adj(G) is taken as a starting point in the search for a
ing demands: suitable D.
926 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936

The decoupler objective states that the decoupler should A time delay of 1.3 s can be removed from the left col-
contain as little time delay as possible. This can be achieved umn, and a delay of 2.6 s from the right column. The left
by a modification of K. It is obvious from Proposition 2 column has a common zero in s = 10 and a pole in
that a common factor of the elements of a decoupler col- s = 2. The right column has a common zero in s = 4
umn can be canceled out by putting its inverse as a factor and a pole in s = 1. Using Method 1, K is therefore
in the corresponding diagonal element of K. Since it is modified to
D = adj(G) Æ K that eventually is implemented it is not a 0 1
s þ 2 1:3s
problem if K contains non-implementable elements (like B s þ 10 e 0 C
inverted time-delays) as long as D does not. This means K¼B @
C;
s þ 1 2:6s A
that a time delay corresponding to the shortest time delay 0 e
among the column elements can be removed from each ele- sþ4
ment by multiplying the corresponding diagonal element of the decoupler D becomes
K with the inverse of this time delay. 0 1
In the same way as time delay is removed from the 1 1 0:2s
B s þ 15 e C
decoupler, poles and zeroes can be removed from the decou- sþ5
D¼B
@ 1
C;
A
pler columns. When poles are removed from the decoupler 1
it might also be necessary to put extra low-pass filters into sþ5 s þ 10
K. Otherwise D might get a high-pass character.
The decoupler design method is summarized in Method 1. which obviously contains less dynamics than adj(G).
The decoupled loop becomes
Method 1 (The decoupler design method )
 
gd 11 0
(1) Start with K = I, then D = adj(G). GD ¼ ;
0 gd 22
(2) Remove the largest common time delay of each
decoupler column by multiplying the corresponding where
diagonal elements of K by inverted time delays.
(3) Remove common poles and zeros of the decoupler sþ4 sþ4
columns by multiplying the corresponding diagonal gd 11 ¼ e2:6s  e2:8s ;
s3 þ 26s2 þ 175s þ 150 s3 þ 11s2 þ 35s þ 25
elements of K by the inverse of the poles and zeros, ðs þ 10Þ 1
possibly also multiplying diagonal elements of K by gd 22 ¼ e1:5s þ 2 e1:3s :
s3 þ 12s2 þ 45s þ 50 s þ 17s þ 30
low-pass filters to avoid giving the decoupler high-
pass character.
The dynamics of the diagonal elements of the transfer-
All the steps of Method 1 are easy to automate which function matrix are quite complicated. Hence, PID control-
makes the method very suitable in an algorithm for auto- ler design methods that rely on simple dynamics cannot be
matic tuning. In an implementation it may be advanta- used. A new PID tuning method that was developed to
geous to normalize the columns of the decoupler, but it is work in this situation is presented in Section 4.
not done here. Method 1 is not restricted to TITO systems,
but the decoupler D may get a complicated for systems of 3.1. Approximate decoupler design method
higher dimension.
The proposed decoupler design method is illustrated in In this section it is assumed that the transfer-function
Example 1. matrix elements can be approximated with first-order plus
dead-time models.
Example 1 (Decoupler design). Consider the process 0 1
k 11 sL11 k 12 sL12
0 1 B T 11 s þ 1 e e C
sþ4 ðs þ 4Þ T 12 s þ 1
e2:6s e2:8s GðsÞ ¼ B@ k 21
C:
A
B ðs þ 1Þðs þ 10Þ ðs þ 1Þðs þ 5Þ C sL21 k 22 sL22
G¼B @ ðs þ 10Þ
C:
A
e e
1:3s s þ 10 1:3s
T 21 s þ 1 T 22 s þ 1
e e
ðs þ 2Þðs þ 5Þ ðs þ 2Þðs þ 15Þ It is motivated to look at this case because the use of first-
order plus dead-time models is very common in the process
If K is chosen as the identity matrix I, the decoupler D be- industry.
comes the adjoint adj(G) of the process transfer function: The adjoint of G is
0 1 0 1
s þ 10 sþ4 k 22 k 12
e1:3s e2:8s sL22
 esL12 C
B ðs þ 2Þðs þ 15Þ ðs þ 1Þðs þ 5Þ C B T 22 s þ 1 e T 12 s þ 1
D ¼ adjðGÞ ¼ B
@
C:
A adjðGðsÞÞ ¼ B
@
C:
A
s þ 10 1:3s sþ4 2:6 k 21 k 11
e e  esL21 e sL11
ðs þ 2Þðs þ 5Þ ðs þ 1Þðs þ 10Þ T 21 s þ 1 T 11 s þ 1
 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936 927

Common time delays of the model adjoint columns are re- A dead time of 4 s can be removed from the left column
moved during decoupler design, according to the method and a dead time of 2 s from the right column. Further, if
proposed above. Common poles are also removed accord- D is approximated to
ing to the proposed method. The number of common col- 0 1
2 2
umn poles can be increased if a certain approximation is B ð5s þ 1Þð2s þ 1Þ C
6s þ 1
used. Each decoupler column element has a pole with a D¼B @
C;
A
time constant Tij. The shortest of these in each column is 1 3 s
e
called Ts and the longest is called Tl. The elements with 5s þ 1 ð6s þ 1Þð3s þ 1Þ
the long time constants are approximated with second-or- a pole corresponding to the time constant T = 5s can be re-
der transfer functions using the approximation moved from the left column, and a pole corresponding to
1 1 the time constant T = 6s can be removed from the right
 : column. Hence, if K is modified to
T l s þ 1 ðT s s þ 1ÞððT l  T s Þs þ 1Þ
 
Then the pole ð5s þ 1Þe4s 0
K¼ ;
1 0 ð6s þ 1Þe2s
ðT s s þ 1Þ
the decoupler D becomes
can be removed from the column. 0 1
2
The approximate design method is summarized in B 2s þ 1 2 C
Method 2. D¼B @
C;
A
3 s
1 e
Method 2 (Approximate decoupler design method). 3s þ 1
which obviously contains less dynamics than adj(G).
(1) Create a first-order plus dead-time model of the
process.
(2) Start with K = I, then D = adj(G). 3.2. Brief history and contributions
(3) Remove the largest common time delay of each
decoupler column by multiplying the corresponding Many textbooks and papers have treated decoupling
diagonal element of K by the inverse of that time in the past, including those by Maciejowski [7], Good-
delay. win et al. [3], Wang and Yang [11], Wang et al.
(4) Use the approximation [10,13,12].
A common approach to decoupling is to use the inter-
1 1 actor matrix theory, see [3]. This theory is based on poly-

T l s þ 1 ðT s s þ 1ÞððT l  T s Þs þ 1Þ nomial representations. Since it is common to have
process models with time delays in process control, this
on the element with the longest time constant in each
means that one is forced to use discrete-time models. To
column.
be able to use continuous-time representations, it was
(5) Remove the common pole of each decoupler column
decided not to use the interactor matrix theory in this
by multiplying the corresponding diagonal element of
project.
K by the inverse of that pole.
The existing decoupler design methods do not satisfy
the decoupler objectives described above [8]. Therefore,
The method is demonstrated in Example 2. a new method for decoupler design was developed. The
contributions are that a clear expression for the whole
Example 2 (Approximate decoupler design). Consider the
space of possible decouplers for a linear square stable
process
non-singular system is provided, and that very simple
0 1
3 3s 2 2s
rules of how to choose a decoupler among those are
B 9s þ 1 e e C developed.
6s þ 1
G¼B @ 1
C:
A
2
e4s e4s 4. PID design method
5s þ 1 7s þ 1
If K is chosen as the identity matrix I, the decoupler D 4.1. The problem
becomes
0 1 There are many PID design methods [9,1]. These meth-
2 4s 2 2s
B 7s þ 1 e e C ods are normally based on the idea of first approximating
6s þ 1
D ¼ adjðGÞ ¼ B@ 1
C:
A the process dynamics with a simple model, and then basing
4s 3 3s the design on this model. This approach works well on
e e
5s þ 1 9s þ 1 SISO systems in the process industry, since these systems
928 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936

are often well described by simple models. An example of measure of the cost of decoupling to judge the benefits of
that is methods that use step responses for tuning [15,4,5]. diagonal decoupling. This is a topic for further research
When PID controllers are to be used for multi-variable in the project.
control of processes with strong cross couplings the situa- If PID controllers are used to control a system with
tion is different. In many cases the system has to be decou- diagonal elements like (1), design methods that rely on sim-
pled. Even if the elements of the system have simple ple process dynamics, like step response methods, are not
dynamics, decoupling may result in complicated diagonal appropriate. Because of that, a PID design method that
elements consisting of parallel coupled processes that might does not rely on simple process dynamics was developed
have different signs and different time delays. An example and is presented in this section.
of such a diagonal element is
4.2. The controller
1:2 1
G¼ e4s  e1s : ð1Þ
ð0:5s þ 1Þð0:7s þ 1Þ ð3s þ 1Þð2s þ 1Þ The PID controller is described by
The step response and the Bode plot of this process is  
1
shown in Figs. 4 and 5, respectively. The process shows a C ¼K 1þ þ T ds ; ð2Þ
typical non-minimum phase behaviour which limits the T is
control performance. This means that the decoupling has where K is the proportional gain, Ti is the integral time,
a price. The cost of decoupling is discussed in e.g. [3]. To and Td is the derivative time. A pure PID controller would
overcome this problem, it is sometimes suggested to use have infinite high-frequency gain. It is both undesirable
partial decoupling, e.g. by creating a triangular system in- and impossible to realize such a controller. Therefore, a
stead of a diagonal one, see [3]. In this project, we have low-pass filter is required. A second-order low-pass filter
decided to keep the diagonal structure even when the cost
of decoupling is high. The major reason is the advantages 1
F ¼ ð3Þ
obtained by having the diagonal system from an operating ðsT f þ 1Þ2
point of view, see Fig. 1. However, it is of interest to get a
is used in this project.
0.5
0.4 4.3. The optimization criteria
0.3
0.2 The chosen optimization criteria is to minimize the inte-
0.1 grated absolute error
0 Z 1
–0.1 IAE ¼ jeðtÞjdt; ð4Þ
–0.2 0

–0.3 where e(t) is the control error at step load disturbances,

–0.4
0 5 10 15 20 25 30
subject to bounds on the sensitivity function and the com-
plementary sensitivity function. The bounds on the sensi-
Fig. 4. Step response of the process (1). tivity functions can be interpreted as two circles in the
complex plane that the Nyquist curve of the open-loop sys-
tem has to stay outside. A larger circle that encircles these
100 two are constructed and called the M-circle [1]. The bounds
on the sensitivity functions then means that the Nyquist
curve should stay outside the M-circle.
|G(iω)|

10–1 The optimization criteria has been motivated and used

before [1] but a short motivation is in place anyway. The
general requirements of a design method for the process
10–2 –2 industry were described in Section 1, and below is a
10 10–1 100 101
description of how these requirements are reflected in the
0 optimization criteria.
–500 Fig. 2 shows the general disturbances l, n, and m. The
arg G(iω) [deg]

–1000 bound on the sensitivity functions bounds the transfer

–1500 functions Gny and Gmy from n and m to the output signal y.
–2000
The second-order low-pass filter (3) gives the controller
a low high-frequency gain. Together with the bounds on
–2500
10–2 10–1 100 101 the sensitivity functions this gives the high-frequency
ω [rad/s]
region of the transfer functions Gnu and Gmu from n and
Fig. 5. Bode plot of the process (1). m to the control signal u low gain. This is important, since
 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936 929

n and m are high-frequency disturbances. Furthermore, the non-integrator poles and the zeros of the process, so this
transfer function Gly from the low-frequency disturbance l region is limited.
to the output signal y is small since the effect of this distur-
bance is minimized. 4.4.3. Parameter K
Good stability margin is given by the fact that the For every pair of Td and Ti, a gain K that gives the sys-
Nyquist curve is kept at a distance from the critical point tem the pre-specified maximum values of the sensitivity
1 by the M-circle. functions, without making the system unstable, has to be
found. For stable processes this corresponds to finding a
4.4. The design method K that puts the Nyquist curve on the edge of the M-circle
without making it encircle the point 1. An algorithm that
An upper bound on the sensitivity functions is specified. checks if the 1 point is encircled has to be used.
The space of possible controllers is discretized in the A large K, Kmax is chosen as a starting value. K is
parameters Ti, Td, and K. For each combination of Ti decreased until the point 1 is not encircled and the
and Td, a gain K is found, that puts the Nyquist curve of Nyquist curve is outside the M circle. If K is lowered under
the open-loop system on the edge of the M-circle in such a certain bound Kmin, without making the system satisfy
a way that the Nyquist curve does not encircle the 1 these specifications, the conclusion is drawn that no stable
point, if possible. For each controller a step load distur- closed-loop system exists for the present combination of Ti
bance is simulated and the integrated absolute error, and Td. Subsequently K is gently increased until the
IAE, is calculated. The controller that gives the smallest Nyquist plot is close to the edge of the M circle. Kmax
IAE is chosen. and Kmin works as upper and lower bounds on K. This is
The method was implemented in Matlab/Simulink and necessary for the method algorithm to work but it is also
tested in several examples, see [8]. The controller design natural to have some bounds on the controller gain.
time was about 1.5 min which is acceptable since the design
is performed off-line. The design time can probably be sig- 4.4.4. IAE
nificantly shortened if effort is put into time optimization of The integrated absolute error IAE (see (4)) is calculated
the code. by the integration of a simulation of a step load distur-
bance response. The controller with the smallest IAE is
4.4.1. The sign then chosen. To improve the accuracy the algorithm can
Since the algorithm should be able to handle processes be repeated with the intervals of Td and Ti centered around
with different signs of the low-frequency gain, a sign is the Td and Ti values of the first controller and with a nar-
added to the PID controller. The output of the process rower grid.
after a step change of the control signal is simulated. If
the output goes to a positive value or towards plus infinity 5. Decoupling and tuning – industrial example
the sign is chosen positive. If the output goes to a negative
value or towards minus infinity the sign is chosen negative. The decoupling procedure presented in Section 3 and the
In either case the controller is connected to the process PID controller design method presented in Section 4 have
using negative feedback. been tested in a paper mill. The results of these tests are
presented in this section.
4.4.2. Parameters Td and Ti
5.1. The process
The controller has one pole in the origin, two filter poles
and two zeros. The zeros are located in:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi At Stora Enso Publication Paper, Hylte Mill in Sweden
1 1 1 paper is manufactured from wood and recycling paper.
z¼   ð5Þ This involves transportation of pulp at several locations
2T d 4T 2d T i T d
in the mill. The transportation is carried out in pipes and
If Ti is less than 4Td the zeros are complex conjugated with the flow is controlled by pumps and valves.
a real part a = 1/2Td. The imaginary part will increase In this case a pipe transporting pulp from a recycling-
with decreasing Ti. If Ti is greater than 4Td, the zeros will pulp tower to a pulp container is considered. The flow
be real and centered around a = 1/2Td.
1/2Td is swept over the frequency region of interest. This
Recycling−
region could, for example, be 0.001–1000 Hz with the grid pulp
points spread in a logarithmic fashion. In this way many tower Container
processes can be covered. PT FT
For each value of Td, Ti is swept over a reasonable
region. In most cases it is not interesting to get a controller
with zeros that have very large imaginary parts or a con-
troller with zeros at frequencies far below or above the Fig. 6. A schematic picture of the process.
930 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936

and the pressure in the pipe are to be controlled. The con-

troller objective is to hold a constant pressure in the pipe D
and make the flow follow a reference signal from the level
controller in the container downstream. There are two PIC FIC

PT FT
Flow
450
440
430 Fig. 10. A picture of the proposed controller design.
420
410
0 50 100 150 200 250
actuators on the pipe, a pump and a valve. The pressure
Input Signal to the Valve and the flow are measured a small distance upstream from
40
38 the valve. A schematic picture of the process is shown in
36 Fig. 6.
34
32
The picture is schematic, but it contains the relevant
0 50 100 150 200 250 parts of the process. The effects of other parts of the pro-
Fig. 7. Experiment pointing out the backlash. cess are seen as disturbances. This process is a TITO system
with the two input signals, to the pump and to the valve,
and the two measured signals of the pressure and the flow.
PIC FIC It has cross couplings since both the pump and the valve
FX affect both the flow and the pressure in the pipe. Since it
is a TITO system with cross couplings the methods pro-
PT FT posed in previous sections are appropriate.
A SIEMENS control system is used for control of the
part of the mill where the pipe is located. The control sys-
Fig. 8. A picture of the original controller design. tem has PID controllers, filter blocks, and delay blocks that

Pressure
1

0.8

0.6

100 200 300 400 500 600 700 800 900

Flow
500
450
400
350
100 200 300 400 500 600 700 800 900

Input Signal to the Pump

30

25

20
100 200 300 400 500 600 700 800 900

Input Signal to the Valve

40

35

30

100 200 300 400 500 600 700 800 900

Fig. 9. Data for system identification.

 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936 931

may be used for control. The sections below describe how 5.2. Backlash in the valve
the process is controlled today and how a new controller
may be designed using the methods proposed in previous There is a backlash in the valve which was revealed by a
sections. simple experiment. Fig. 7 shows the input to the valve and
The process signals were logged during the experiments the flow response during the experiment. The response
as described below. In all the experiment plots the time is in the measured flow is quite slow. The reason for this is
given in seconds, the pressure unit is bar, the flow unit is that the flow signal is filtered with a low-pass filter with
m3/h, the signal to the pump is given in percent of capacity the time constant 20 s. The experiment shows that the
and the signal to the valve is given in percent of open- backlash has a magnitude of about 3 percent.
ness. Some of the experiments were performed in closed A backlash of this magnitude limits the achievable
loop. In these cases the reference signals were noted by performance of the system and makes it necessary
hand. to use a robust controller. This situation is expected

Pressure

0.8

10 20 30 40 50 60 70 80 90 100 110 120

Flow
650

600

550
10 20 30 40 50 60 70 80 90 100 110 120

Input Signal to the Pump

60

55

50

45
10 20 30 40 50 60 70 80 90 100 110 120

Input Signal to the Valve

45

40

35

30
10 20 30 40 50 60 70 80 90 100 110 120

Pressure Controller Output

62
60
58
56
54
10 20 30 40 50 60 70 80 90 100 110 120

Flow Controller Output

62
60
58
56
54
10 20 30 40 50 60 70 80 90 100 110 120

Fig. 11. Data from test with the decoupled system. At the approximate times of 37 s and 85 s the pressure controller output is raised by 3%.
932 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936

and common in process control, and the design meth- It means that the pressure control is considered more
ods proposed in previous sections give a robust con- important than the flow control.
troller. The PI controllers have been tuned manually, and the
parameters are K = 0.2 and Ti = 2 s in the pressure control-
5.3. Original design ler. In the flow controller the parameters are K = 0.6 and
Ti = 28 s. The signals are normalized in the controllers
The original control structure is depicted in Fig. 8. Two and the controller gains K above are given for the normal-
PI controllers are used to control the process. A first-order ized system. The controller gains are K = 4 in the pressure
low-pass filter FX with the time constant T = 20 s is placed controller and K = 0.07 in the flow controller if physical
in the flow loop before the controller. It is put there to units are used instead of normalized. In the following sec-
reduce the effects of cross couplings and of the backlash. tions it is assumed that physical units are used.

Pressure

0.85
0.8
0.75
0.7
20 40 60 80 100 120

Flow

600

550

500
20 40 60 80 100 120

Input Signal to the Pump

52
50
48
46
44
42
20 40 60 80 100 120

Input Signal to the Valve

46
44
42
40
38
20 40 60 80 100 120

Pressure Controller Output

62
60
58
56
54
20 40 60 80 100 120

Flow Controller Output

62
60
58
56
54
20 40 60 80 100 120

Fig. 12. Data from test with the decoupled system. At the approximate times of 29 s and 82 s the flow controller output is raised by 3%.
 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936 933

5.4. System identification This was done by fitting first-order plus dead-time models
to the step responses of the experimental data.
The design methods proposed in previous sections are
model-based. Thus, a model of the process had to be 5.5. Controller design
found. Simple experiments were carried out to be used
for system identification. Fig. 9 shows the experimental The proposed control structure is depicted in Fig. 10. A
data that was used. The data was captured in open loop. decoupler and PID controllers were designed by the meth-
Two steps, a negative step followed by a positive step in ods derived in previous sections (the columns of the decou-
the signal to the valve were performed, and both the pres- pler were also normalized). The decoupler was built up by
sure and the flow were logged. Then the same was done filter blocks and delay blocks in the SIEMENS control sys-
with the signal to the pump. This gave the required data tem. It had the structure
and a linear process transfer-function matrix was identified 0 1
1:36s 1:06
as B 2:00e C
0:9s þ 1
0 1 D¼B @
C:
0:018e2s 0:02e1:4s 0:950 0:6s A
B 0:42s þ 1 0:682 e
0:9s þ 1 C 0:42s þ 1
G¼B @ 4:1e2:64s
C:
12e4s A The PID controllers in the SIEMENS control system are
3s þ 1 3s þ 1 on the following form

Pressure
0.9

0.8

0.7

0.6

0.5
100 200 300 400 500 600

Filtered Flow
1000

800

600

400

200
100 200 300 400 500 600

Input Signal to the Pump

80
70
60
50
40
30
100 200 300 400 500 600

Input Signal to the Valve

70
60
50
40
30
20
100 200 300 400 500 600

Fig. 13. Data from reference tests with the old controller. The dashed lines indicate the reference signals for the pressure and the flow. The solid lines
show the logged signals.
934 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936
 
1 s The sampling time of the system was 1 s. Since the
C ¼K 1þ þ Td :
T is 1 þ T d s=v design was performed in continuous time and the actual
control was performed in discrete time some discretization
The measurement signals were filtered by second-order
effects were expected. A specific analysis of this was not
low-pass filters
made. The controller design methods were developed to
1 give a robust controller which handled the various sources
F ¼ :
ðsT f þ 1Þ2 of uncertainty.

The only difference between this controller structure and 5.6. Test of decoupling
the controller structure described in previous sections
is the extra filters on the D-part of the PID controllers. The decoupler was first implemented and tested. When
In the experiments the parameter v was chosen sufficiently the loops are decoupled the pressure controller output
large to make the filters negligible, such that the developed should affect only the pressure, and the flow controller out-
methods could be used without restrictions. put should affect only the flow. The first test was to do
The proposed method with Ms and Mp values of 1.2 for some steps in the pressure controller output. Fig. 11 shows
both controllers gave the controller parameters. In the the result. At the approximate times of 37 s and 85 s the
pressure loop they were K = 3.60, Ti = 1.76 s, Td = pressure controller output is raised by 3%.
0.703 s and Tf = 0.141 s, and in the flow loop they were It is obvious from the experiment that the cross coupling
K = 0.0144, Ti = 4.49 s, Td = 0.703 s and Tf = 0.0873 s. from the pressure controller output to the flow is negligible.
Furthermore, v = 100 was chosen in both loops. It is interesting to see that the decoupler action is as intui-

Pressure
0.9

0.8

0.7

0.6

0.5
100 200 300 400 500 600

Flow
1000

800

600

400

200
100 200 300 400 500 600

Input Signal to the Pump

70
60
50
40
30
20
100 200 300 400 500 600

Input Signal to the Valve

70
60
50
40
30
20
100 200 300 400 500 600

Fig. 14. Data from reference tests with the proposed decoupled controller. The dashed lines indicate the reference signals for the pressure and the flow.
The solid lines show the logged signals.
 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936 935

tively clear as increasing the signal to the pump and closing 6. Summary and future work
the valve a bit, when its input signals demand that the pres-
sure should be raised and flow should be kept at the same TITO systems with strong cross couplings are an impor-
level. tant class of systems in the process industry. Today these
In the second test the pressure controller output was systems are not dealt with in an effective way. A new
kept constant and some steps were performed in the second decoupler design method and a new PID controller design
input signal. At the approximate times of 29 s and 82 s the method have been proposed in this paper. The methods ful-
flow controller output is raised by 3%. The results are fill the special requirements of an automatic method (see
shown in Fig. 12. Section 2) and give the closed-loop system good distur-
It is obvious that cross coupling from the flow controller bance attenuation properties.
output to the pressure is negligible and it is interesting The methods were developed to work together, but each
again to see that the action of the decoupler is quite intui- of them may also be useful alone in other contexts. Most
tive. When its input signals demand that the pressure PID design methods work for processes with simple
should be kept and flow should be raised, it increases the dynamics but the one proposed in this paper works for pro-
signal to the pump and opens the valve a bit. cesses with complicated dynamics as well, which of course
is a great advantage. The proposed decoupler design
method may also be used in other contexts where a decou-
5.7. Flow setpoint changes
pler is needed.
The proposed methods were tested in both simulations
Under normal operating conditions the pressure should
and on a real process in an industrial environment. It
be kept at a constant value while the flow should follow
was observed that the proposed methods gave the closed-
external setpoint changes. Thus, it is interesting to see what
loop systems good disturbance attenuation properties, as
happens when this is done in closed loop. It was tested
they were designed to do. During the industrial tests the
with both the old controller and the new decoupled con-
advantage of having decoupled loops when independent
troller. Fig. 13 shows the results when the old controller
setpoint changes should be followed was also obvious
is used. The reference value for the pressure is kept at
(see Figs. 13 and 14).
0.7 bar and the reference value for the flow is altered in
There are some natural areas for continuation of the
steps of 120 m3/h. The old controller does not manage to
project. The first step in the algorithm is to find a model
keep the pressure close to the reference value. The flow
of the TITO system by system identification. That part is
was logged between the low-pass filter and the flow con-
not covered here and should receive future attention. The
troller when the old controller was used. This makes the
following two steps are the decoupler design and the PID
flow changes appear to be a bit smoother and slower than
controller design for which methods were proposed in
they really are.
this paper (see Sections 3 and 4). The last step of the
Fig. 14 shows the results for a similar experiment with
algorithm is design for anti-windup and for bumpless
the proposed decoupled controller. This controller
transfer between manual and automatic mode. Other
manages to keep the pressure much closer to the reference
interesting research topics are to find good measures for
value and is able to perform faster setpoint changes in the
the cost of decoupling and to design pre-filtering of
flow.
setpoints.

5.8. Conclusion
Acknowledgement
Simulations are a good first test of a control strategy,
but they should always be followed by real-world experi- This work was supported by the Swedish Research
ments, in this case industrial tests. There are differences Council (VR).
between simulations and industrial tests. In the latter there
are real disturbances, nonlinearities like the backlash, and
unmodeled dynamics. These things are often hard to incor- References
porate in simulations.
[1] K.J. Åström, T. Hägglund, Advanced PID Control, ISA – The
In this section industrial tests of the developed control
Instrumentation, Systems, and Automation Society, Research Trian-
strategy were presented. The conclusion of the tests is that gle Park, NC, 2005.
the proposed control strategy works well. The pressure [2] K.J. Åström, K.H. Johansson, Q. Wang, Design of decoupled PI
loop and the flow loop are decoupled and the disturbance controller for two-by-two systems, IEE Proceedings on Control
attenuation properties of the controller are good. Further, Theory and Applications 149 (2002) 74–81.
[3] G.C. Goodwin, S.F. Graebe, M.E. Salgado, Control System Design,
the controller design methods fulfill the special require-
Prentice-Hall, Upper Saddle River, NJ, 2001.
ments of automatic methods (see Section 2), which is nec- [4] T. Hägglund, K.J. Åström, Revisiting the Ziegler–Nichols tuning
essary if they are to be used in an automatic design and rules for PI control, Asian Journal of Control 4 (4) (2002)
tuning algorithm. 364–380.
936 P. Nordfeldt, T. Hägglund / Journal of Process Control 16 (2006) 923–936

[5] T. Hägglund, K.J. Åström, Revisiting the Ziegler–Nichols step [11] Q.-G. Wang, Y. Yang, Transfer function matrix approach to
response method for PID control, Journal of Process Control 14 (6) decoupling problem with stability, Systems & Control Letters 47
(2004) 635–650. (2002) 103–110.
[6] P. Lancaster, Theory of Matrices, Academic Press, New York, 1969. [12] Q.-G. Wang, Y. Zhang, M.-S. Chiu, Decoupling internal model
[7] J. Maciejowski, Multivariable Feedback Design, Addison-Wesley control for multivariable systems with multiple time delays, Chemical
Publishing Company, 1989. Engineering Science 57 (2002) 115–124.
[8] P. Nordfeldt, PID control of TITO systems. Licentiate thesis ISRN [13] Q.-G. Wang, Y. Zhang, M.-S. Chiu, None-interacting control design
LUTFD2/TFRT-3238-SE, Department of Automatic Control, Lund for multivariable industrial processes, Journal of Process Control 13
Institute of Technology, Sweden, December 2005. (3) (2003) 253–265.
[9] S. Skogestad, Probably the best simple PID tuning rules in the world, [14] K. Zhou, J.C. Doyle, Essentials of Robust Control, Prentice-Hall,
in: AIChE Annual Meeting, Reno, NV, USA, November 2001. Upper Saddle River, NJ, 1998.
[10] Q.-G. Wang, B. Huang, X. Gu, Auto-tuning of tito decoupling [15] J.G. Ziegler, N.B. Nichols, Optimum settings for automatic control-
controllers from step tests, ISA Transactions 39 (2000) 407–418. lers, Transactions of ASME 64 (1942) 759–768.