You are on page 1of 16

Journal of Process Control 17 (2007) 157172

www.elsevier.com/locate/jprocont

A dynamic operability analysis approach for nonlinear processes


Osvaldo J. Rojas a, Jie Bao
a

a,*

, Peter L. Lee

School of Chemical Sciences and Engineering, The University of New South Wales, Sydney, NSW 2052, Australia
b
Chancellery, Level 2, 160 Currie Street, The University of South Australia, Adelaide, SA 5000, Australia
Received 22 May 2006; received in revised form 25 August 2006; accepted 9 September 2006

Abstract
Current process operability indicators are mostly restricted to linear approximations of the process dynamics. Other operability analysis approaches that have the capability to include full nonlinear process models rely on mixed integer dynamic optimisation techniques
which, in general, require large amount of computations. In this paper we propose a dynamic operability analysis approach for stable
nonlinear processes that can be readily applied during process design and can be solved eciently using a limited amount of computations. The process nonlinear dynamics are approximated by a series interconnection of static nonlinearities and linear dynamics, represented by the so-called HammersteinWiener models. These type of models can often be obtained during process design where detailed
steady-state nonlinear models are available, combined with some (usually limited) information on the process dynamics. Using an
extended internal model control (IMC) framework, we investigate the interaction between the static nonlinearities and linear dynamics
on the operability of the process. The framework extends the well-known equivalence between operability and invertibility of linear processes to nonlinear systems. In particular, by exploiting some results from the theory of passive systems we provide conditions that guarantee the existence of the inverse of the static nonlinearities. We show that the inverse can be attained inside a specic input/output
region. This region imposes a constraint on the maximum magnitude of the signals that appear in the closed-loop and represents the
eect of the static nonlinearities on the operability of the overall process. Dynamic operability is then quantied using a linear matrix
inequality (LMI) optimisation approach that minimises a given performance criterion subject to the constraint imposed by the static
nonlinearities.
 2006 Elsevier Ltd. All rights reserved.
Keywords: Process operability; Nonlinear systems; HammersteinWiener model; Passivity; Static nonlinearity; Input constraints

1. Introduction
Process operability refers to the inherent property of a
process to achieve acceptable control performance in spite
of unknown but bounded disturbances and model uncertainty, using the available manipulated variables and sensor measurements [1]. Process design decisions, such as
the number of trays in a distillation column, the energy
integration interconnection between equipment and the
presence of recycle streams, can have a signicant impact
on the operability characteristics of the process (e.g., [2
4]). Ignoring operability considerations during process
*

Corresponding author. Tel.: +61 2 9385 6755; fax: +61 2 9385 5966.
E-mail address: j.bao@unsw.edu.au (J. Bao).

0959-1524/$ - see front matter  2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2006.09.001

design may lead to costly retro-tting and re-design if the


process is found to be dicult to control during
commissioning.
A large number of operability indicators have been proposed and developed in the literature to assist in assessing
the operability properties of the process during the design
stage. These indicators include the process singular values
[1,5,6], the relative gain array (RGA) [7,8], the dynamic relative gain array, the closed-loop disturbance gain [9], etc.
Most of these operability indicators can be easily computed, especially when they refer only to the steady-state
information of the process. However, these operability
indicators are generally limited to linear systems. Also, they
usually reveal the eect on operability of only some process
characteristics, e.g. non-minimum phase zeros, delays,

158

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

input constraints, etc. Thus, they may fail to quantify the


combined eect of these factors on dynamic operability.
Dynamic operability indicators for nonlinear processes
are still not widely available. This is partly due to the fundamental complexities of nonlinear dynamics. One (indirect) approach to study the operability of nonlinear
processes is to quantify the degree of nonlinearity of the
process so as to assess whether linear control is able to
achieve the desired performance. This approach has been
pursued using a variety of nonlinearity measures [1014].
Even though the question whether linear control is sucient to control the process is relevant per se, it does not
fully identify the inherent operability properties of the nonlinear process.
A dierent approach to process operability analysis is
based on mixed integer dynamic optimisation (MIDO).
This approach has been explored by several authors, notably by Perkins and co-workers in the context of their simultaneous process and control design methodology [1517].
The aim of this methodology is to determine both the best
process design as well as the best control structure and
parameters in one single integrated framework. The most
appealing advantage of this approach is that it can incorporate, with great exibility, a variety of process and control considerations. For example, both economic indexes
(e.g. expected total annualised cost) and dynamic performance indexes (e.g. integral square error) can be included.
Structural process and control design decisions can be
included via boolean variables. Actuator constraints and
process variables constraints can be considered explicitly.
Finally, in principle, the approach can make use of detailed
nonlinear process models, thus leading to potentially more
realistic analysis results. Unfortunately, the main disadvantage of the optimisation approach is that the complexity of
the resulting optimisation problem grows very quickly even
for moderately small dimensional processes [17]. Sakizlis
et al. have conceded in a recent overview of the approach
[15] that one of the challenges that lies ahead in this area
is the need for a rigorous and ecient solution of the
underlying optimisation problem. To make the problem
tractable several simplifying assumptions are usually
required, such as assuming that the process is controlled
with PI controllers, considering only a nite time horizon
in the optimisation, etc. These simplications may lead to
an unrealistic estimate of the process operability. In addition, the results may depend heavily on the specic time
domain prole assumed for the disturbance and reference
signals.
Recently, Georgakis and co-workers have proposed a
geometric approach for process operability analysis
[1820]. The advantage of this approach is that it is applicable to both linear and nonlinear processes and requires
only limited computation. In its simplest version the geometric approach considers only the steady-state nonlinear
information of the process [20]. The user denes the available input space (AIS) for the manipulated variables and
the desired output space (DOS) for the controlled vari-

ables. Using the nonlinear steady-state information the


AIS can be mapped into the output space and compared
to the DOS; equivalently, the DOS can be mapped into
the input space and compared to the AIS. This procedure
provides an indication of whether the desired range of
operation can be achieved in steady-state. However, the
approach is essentially open-loop. It does not guarantee
that the closed-loop will have acceptable operability properties. An extension to the steady-state geometric approach
that considers the stability of the closed-loop has recently
been developed by Rojas et al. [21]. Ekawati and Bahri
[22] have also used the geometric interpretation of the operating spaces, such as the DOS and AIS, to simplify the
dynamic operability framework [23] for regulatory cases.
In this paper, we propose a methodology to analyse the
dynamic operability of stable nonlinear processes. Our
main objective is to go beyond existing operability analysis
approaches by incorporating explicit information on the
process nonlinearity. We also aim at developing an
approach that can be readily applied during process design
and that can be solved eciently using limited amount of
computations.
Detailed steady-state nonlinear models of the process
are routinely accessible during process design. These models are well understood and have been extensively tested
and validated. Indeed, several commercial process design
software, such as Aspen, have a comprehensive library
of detailed nonlinear steady-state models for the most common process units in industry. On the other hand, only limited information on the process nonlinear dynamics is
usually available during the process design stage. However,
some studies have shown that the process nonlinear
dynamics can often be described with sucient accuracy
using static nonlinearities in conjunction with linear
dynamics [2427]. Linear approximations to the process
dynamics can be obtained from process owsheet data
using, for example, the approach proposed by Lewin and
co-workers [28,29]. Thus, in principle, relying only on owsheet data, one may derive a nonlinear dynamic model of
the process based on a series interconnection of static nonlinearities and linear dynamics as follows:
y ho fGp shi fugg
m

1
m

where hi fg : R ! R and ho fg : R ! R are input and


output static nonlinearities, respectively, and Gp(s) is a linear multivariable transfer function. The model in (1) is said
to have a HammersteinWiener type structure.
The operability analysis approach described in this
paper considers a process model as in (1) and is based on
an extended internal model control (IMC) framework (similar to Chan et al. [30] for multivariable control design). In
this way, one can investigate how the static nonlinearities
and the linear dynamics Gp(s) interact with each other to
aect the overall operability of the process. A key concept
related to the issue of process operability is that of process
invertibility [5,31,1]. The close connection between process
operability and process invertibility is highlighted by the

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

IMC framework. Ideally, to achieve perfect control, the


IMC controller should implement a perfect inverse of the
process. However, both theoretical and practical issues
pose limitations on the feasibility of implementing the
inverse of a dynamic system (e.g. non-minimum phase
zeros, time-delays, unstable poles, limited bandwidth of
physical components, model uncertainties and actuator
constraints). Thus limitations on process invertibility lead
to limitations on closed-oop performance and operability
[31].
Of particular interest to our approach is to determine
when it is possible to calculate the inverse of the static nonlinearity h{}. In many cases, especially when h{} is a multivariable nonlinear map, it may be either impractical or
not possible to derive an analytical inverse h1{}. We
address the issue by exploiting some results from the theory
of passive systems [3234]. In particular, based on an incremental passivity property we derive sucient conditions
that guarantee the existence of the inverse h1{}. It is
shown that h1{} can only be attained inside a specic
inputoutput region. This region imposes a constraint on
the maximum magnitude of the signals that appear in the
closed-loop. This constraint represents the eect of the static nonlinearity h{} on the operability of the overall nonlinear process.
To quantify the dynamic operability of the process we
consider the worst-case integral square error (wISE) when
the disturbances that aect the process are assumed to have
bounded energy. A key issue is to determine which system
norm is the most adequate to express the required constraints caused by the conditions on the invertibility of
the static nonlinearity h{}. A good candidate to achieve
this objective is the generalised H2 norm introduced by
Wilson [35] and Rotea [36]. The generalised H2 norm
has been found useful in process operability analysis for
linear processes [37] and has the advantage of being amenable to very ecient computation via a Linear Matrix
Inequality (LMI) approach [38,39].
The paper is organised as follows: Section 2 describes
the framework proposed to analyse the dynamic operability of stable nonlinear processes. Section 3 studies the conditions required to guarantee the existence of the inverse of
the static nonlinearities on the basis of a passivity analysis.
Section 4 presents the methodology that we propose to
quantify the operability of the nonlinear process. Finally,
Section 5 illustrates the application of the proposed
dynamic operability approach to a nonlinear neutralisation
process and a multi-unit reactor-separation process.
2. A framework for dynamic operability analysis
In order to study the dynamic operability of a stable
nonlinear process P we will adopt an internal model control (IMC) framework as shown in Fig. 1. The IMC framework has several advantages that makes it especially
suitable for operability studies. First, when there is no process-model mismatch, i.e. P0 = P, the closed-loop relation

159

Fig. 1. Internal model control scheme for stable processes.

between the process output y(t) and the reference signal


r(t) is ane with respect to the IMC controller Q:
yt PQrt

This facilitates the design of the IMC controller Q to meet


the required control objectives. Notice that in the nonlinear
case, the same closed-loop relation (2) holds except that PQ
denotes the series interconnection of the nonlinear systems
P and Q. Secondly, when the process is linear, a necessary
and sucient condition that guarantees the internal stability of the closed-loop in Fig. 1 is that the IMC controller Q
be stable. Unfortunately, this simple condition cannot be
extended directly to the general nonlinear case, and additional assumptions need to be considered [40]. However,
the dynamics of nonlinear processes with a HammersteinWiener model are linear, thus the stability conditions
for the IMC closed-loop shown in Fig. 1 are essentially the
same as those for linear systems.
The closed-loop relation in (2) highlights one of the fundamental properties of feedback control, namely that a
feedback control loop implicitly implements the inverse
of the process (usually over a limited frequency range).
When the process P admits a realisable and stable right
inverse P1 then, in principle, one could select Q = P1
and perfect control [31,1] is attainable, i.e. y(t) = r(t)"t.
From an operability point of view, processes that admit
perfect control are those with ideal operability characteristics. However, it should come as no surprise that this
ideal behaviour is rarely achievable in practice due to several limitations. These limitations include unstable zeros,
time-delays, actuator constraints and model uncertainty
[4143,1] and, eectively, they indicate that a perfect
inverse of the process P is not attainable.
There exists a large body of literature that has addressed
the question of quantifying in detail what can be achieved
with feedback control when a linear process P admits no
exact inverse [41,44,31,45,46,42]. However, only limited
studies are available for nonlinear processes [47]. The current results in both the linear and nonlinear case point consistently to the conclusion that what is achievable with
feedback control depends almost entirely on the inherent
properties of the plant and is independent of the controller.
Thus, operability is an inherent property of the process.
Considering the IMC framework shown in Fig. 1, we
can state the problem of assessing the dynamic operability
of a nonlinear stable process P as follows:

160

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

Dynamic operability assessment problem (Stable nonlinear case)


For a given performance criterion for operability, nd
the realisable and stable IMC controller Q that achieves
the best performance.
The above denition does not specify the type of operability analysis carried out on the nonlinear process P. This
issue lies entirely on the selection of the performance criterion for operability. Depending on the specic application
considered, one performance criterion may be more appropriate than others. Observe that the optimality implied
when referring to the best performance achieved by the
IMC controller Q in the denition ensures that the operability is determined by the process inherent properties and
not by the controller. To illustrate this point consider, for
example, the integral square error (ISE) as the chosen performance criterion for operability. The operability assessment problem in this case amounts to nd Q such that:
Z 1
2
opt
J , min
krt  ytk dt
3
Q

When the process P is linear and the reference signal r(t) is


assumed to be a unit step, the optimal Q that minimises the
ISE is given by [48,49]:
Qopt s P 1
m s

where the process P(s) is decomposed as:


P s LsP m s

and Pm(s) represents the minimum phase part of P(s) and


L(s) is an all-pass factor. Observe that the optimal IMC
controller QOPT inverts only the invertible part of the process, namely its minimum phase factor. This result conrms
that the operability of the process is independent of the
controller and is an inherent property of P(s). Another
interesting observation is that if the linear process P(s) is
minimum phase and admits a right inverse, then the operability analysis based on the ISE yields the result
QOPT(s) = P(s)1 (since Pm(s) = P(s)). Thus perfect control is attainable and JOPT = 0.
Extending dynamic operability analysis to general nonlinear processes is a formidable task. One of the main difculties is that nonlinear optimisation problems are

dicult to solve numerically, let alone being amenable to


an analytical solution for QOPT such as (4). In this paper,
we propose a methodology to study the dynamic
operability of stable nonlinear processes that can be modelled using a series interconnection of static nonlinearities
and a linear dynamic block such as in (1). To simplify
the discussion, we will focus on the following Hammerstein-type model:
y Gp shfug

where it is assumed that ho{} = I and hi{} = h{}. However, the method can readily be extended to models having
both input and output static nonlinearities.
One of the key elements of the methodology is the use of
an extended version of the IMC framework shown in Fig. 1
in order to take full advantage of the specic structure of
the Hammerstein-type process model. In particular, we
consider an IMC controller Q formed by the series interconnection of a linear system Qp(s) and a static nonlinearity g{}, i.e.
Q gfQp sg

This structure allows one to analyse the operability of the


nonlinear process P by separating the inuence of the static
nonlinearity h{} from that of the linear dynamics Gp(s).
Notice that the stability of the closed-loop is guaranteed
if Qp(s) is stable, regardless of the choice for g{}. However,
both static nonlinearities g{} and h{} are required to generate a bounded output whenever a bounded input is
applied.
Fig. 2 shows the equivalent closed-loop when Q has the
form shown in (7) and the process P can be approximated
by a Hammerstein-type model. When there is no processmodel mismatch, we obtain the following closed-loop
relation:
y Gp shfgfQp srgg

In the next Section we will study in more detail how the


interaction between the static nonlinearity h{} and the linear dynamics Gp(s) aect the operability of the nonlinear
process P. We will see that the presence of the static nonlinearity h{} in the process model will impose additional

Fig. 2. Closed-loop with extended IMC architecture.

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

limitations on the operability of P compared to the case of


P being a purely linear process P = Gp(s).

161

Denition 3.1 (Passive static nonlinearity). A static nonlinearity y h hfug : Rm ! Rm is said to be passive if:
T

3. The eect of the static nonlinearity on dynamic operability


In order to assess the dynamic operability of the process
P, we need to determine the best static nonlinearity g{}
and best linear transfer function Qp(s) with respect to the
chosen performance criterion. In most practical cases, the
performance criterion for operability will include (directly
or indirectly) a term related to the size of the process output y(t). This is the case, for example, if we choose the ISE
as the performance criterion for operability. Thus, for most
meaningful performance criteria the best static nonlinearity g{} is one that implements a right inverse of the process static nonlinearity h{}, i.e. g{} = h1{}. However,
similar to the dynamic case discussed earlier, the inverse
of a multivariable nonlinear static map h{} may not exist
or may be dened only over a limited range.
If the inverse of the static nonlinearity exists and it is
known, then one could select g{} = h1{}, thus eectively
removing h{} from the closed-loop. In this case the operability of the process P is only determined by its linear subsystem. However, even in the ideal case when the inverse
h1{} exists and is readily available, the nonlinearity
may still impose restrictions on the dynamic operability
of the overall process P. This issue arises primarily from
the fact that the static nonlinearity yh = h{u} (see Fig. 2)
may be dened only over a certain limited domain of the
input u 2 Rm and/or may span only a limited range of
y h 2 Rm , that is:
hfug : Xu  Rm ! Xy h  Rm

This is the result of physical limitations in the process, but


it can also account for limitations in the actuators, output
constraints, etc. Accordingly, the known inverse will be dened on the following range:
gfrh g , h1 fg : Xy h ! Xu

10

Under these conditions, the region Xy h places a constraint


on the output rh(t) of the linear sub-system Qp(s) in
Fig. 2. An example of this situation is when the static nonlinearity h{u} is a saturation function modelling actuator
limitations. In this case, the inverse of h{u} is readily available: in the linear range where u 2 [umin, umax] the inverse of
h{u} is the identity, whilst outside the linear range the inverse is not dened. It is well known that input constraints
may impose severe restrictions on the operability of an
otherwise linear process [31,5054] and, in general, cannot
be ignored.
In many practical cases the static nonlinearity h{} is not
invertible in the entire domain Xu. We will address the issue
of nding the invertible region in Xu by exploiting some
results from the theory of passive systems [3234]. For
the sake of completeness we present the following denitions:

hfug u P 0 8u 2 Xu

11

where h{u}T is the transpose of the vector h{u}.


Denition 3.2 (Incrementally passive static nonlinearity). A
static nonlinearity y h hfug : Rm ! Rm is said to be incrementally passive if:
hfu Dug  hfugT Du P 0

8u; u Du 2 Xu

12

Based on these denitions we have the following result:


Theorem 3.3 (Desoer and Vidyasagar (1975)). Consider
a nonlinear mapping y h hfug : Xu  Rm ! Xy h  Rm .
Assume that h{u} is incrementally passive and that it satisfies
the following Lipschitz continuity condition:
khfu Dug  hfugk 6 ckDuk

13

for all u 2 Xu and u + Du 2 Xu with 0 6 c < 1. Under these


conditions, h{u} is a bijection from Xu to Xyh (i.e. h{u} is
one-to-one and onto). Thus, h{}1 exists and is a well defined map from Xy h to Xu.
The above theorem outlines general conditions for the
existence of the inverse of the static nonlinearity h{}. However, the result is dicult to verify in practice, since it
requires checking the incremental passivity condition (12)
for all possible values of u 2 Xu and all possible increments
Du such that u + Du 2 Xu. If further assumptions are considered, namely that the static nonlinearity h{} is continuously dierentiable, then conditions for the existence of the
inverse h1{} can be derived in terms of the Jacobian of
the static nonlinearity h{}. This is described in the following theorem.
Theorem 3.4. Consider the nonlinear mapping y h
hfug : Xu  Rm ! Xy h  Rm . Assume that h{} is continuously differentiable over Xu  Rm . Assume there exists a
non-empty and convex region Hu  Xu defined by

 


 oh
oh



14
P 0 and   < 1
Hu , u 2 X u 
ou
ou
and let Hrh  Xy h be the region that results from mapping Hu
into the output space of h{}, i.e.
Hrh , hfHu g

15

Then, the static nonlinearity inverse h1{} exists and it is a


well defined map from Hrh to Hu.
Proof. Since h{u} is continuously differentiable we can
apply the Mean Value Theorem [55] and write:

oh
hfu Dug  hfug 
Du
16
ou uu0
for all u 2 Xu and u + Du 2 Xu, where u0 2 Rm is a point on
the straight line segment from u to u + Du. It is then clear
that:

162

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172


  
oh  oh
  
Du
khfu Dug  hfugk 
ou  6 oukDuk

17

where we have omitted the explicit reference to the point at


which the Jacobian is evaluated for ease of notation. If we
let c k oh
k then (17) shows that the static nonlinearity is
ou
Lipschitz continuous as per (13). Now, by means of (16)
we have that:
T

hfu Dug  hfug Du DuT

ohT
Du
ou

It is sometimes useful to consider a scaled version of the


static nonlinearity h{}:
19

where u = K u and K is a constant matrix. The reason for


this is that by appropriately choosing the scaling matrix K
it is possible to guarantee that the region Hu in Theorem
3.4 is never empty in the vicinity of a certain point of interest u = u* 2 Xu. This is shown in the following Lemma.
Lemma 3.5. Consider a static nonlinearity y h
hfug : Xu  Rm ! Xy h  Rm with the same assumptions
outlined in Theorem 3.4. Consider the scaled nonlinearity
y h h0 fu0 g : X0u  Rm ! Xy h  Rm defined in (19). Let

K


1
oh
ouuu

Once the region of invertibility of h{} has been identied, the problem of studying the dynamic operability of
a stable nonlinear process P becomes:

18

Thus, if there exists a nonempty and convex region Hu dened as in (14) in which the Jacobian is positive semi-definite, i.e. zT oh
z P 0 8z 2 Hu then the static nonlinearity
ou
h{u} is incrementally passive for all u 2 Hu and
u + Du 2 Hu. Using the result in Theorem 3.3 we conclude
that the static nonlinearity inverse h1{} exists and it is a
well dened map from Hrh to Hu, where Hrh is given in
(15). This concludes the proof. h

y h h0 fu0 g , hfKu0 g

Hence, we conclude that the Jacobian of the scaled nonlinearity h 0 {u 0 } is positive denite for all u 0 inside the local region Ku0 . Based on the denition of H0u we have that
Ku0  H0u . Thus H0u is not empty. In addition, H0u contains
the point u 0 = u 0 * since u0 2 Ku0 . h

20

for a certain point of interest u = u* 2 Xu. Then the region of


invertibility H0u of the scaled nonlinearity h 0 {u 0 } contains the
point u 0 * = K1u* and a non-empty neighbourhood of
u 0 = u 0 *.
Proof. Let Ku0 be a small neighbourhood around the point
u 0 = u 0 *. In the interior of Ku0 the nonlinear mapping h 0 {u 0 }
can be approximated by the rst term of a Taylor series
around u 0 = u 0 *, i.e.
 

oh
y h h0 fu0 g  y h
KDu0 ; 8u0 Du0 2 Ku0
ou 

Dynamic operability analysis problem: Consider a stable


nonlinear process P and assume that P can be modelled
as in (6). For a given performance criterion, the dynamic
operability of P can be assessed by identifying the realisable and stable transfer function Qp(s) that achieves
the best performance subject to the constraint:
rh t 2 Hrh  Xy h

8t

We conclude this Section with a brief discussion on how


one can calculate the inverse g{rh(t)} = h1{rh(t)} in a practical set-up. However, we emphasise that the region of
invertibility of the static nonlinearity h{} dened in Theorem 3.4 is independent of the specic method one may wish
to use to calculate the inverse. In a practical set-up, one
possibility is to implement the static nonlinearity g{} via
a high-gain closed-loop conguration as shown in Fig. 3.
It can be shown see Rojas et al. [21] that if rh(t) is equal
to a constant value r* such that r* is contained inside the
region Hrh then the closed-loop in Fig. 3 is asymptotically
stable. Although obtained using a Lyapunov stability argument, the result in Rojas et al. [21] relies on an incremental
passivity condition for h{} similar to that required in Theorem 3.4 and 3.3. The asymptotic stability of the closedloop in Fig. 3 implies that, in steady-state, y 0hss r , hence
uss = h1{r*}. Thus, the conguration in Fig. 3 achieves an
implicit inverse of the static nonlinearity h{} in steadystate. Observe that with the implementation depicted in
Fig. 3, g{} is, strictly speaking, no longer a static nonlinearity but becomes a dynamic nonlinear system. However,
from an inputoutput point of view the relation
u(t) = g{rh(t)} is, essentially, static provided the closedloop in Fig. 3 can be made arbitrarily faster than the
dynamic response of Qp(s). This can be achieved by using
an arbitrarily large gain e > 0 in Fig. 3.

uu

21
Thus,
oh0

ou0

 

oh
K;
ouuu

8u0 2 Ku0

22

Substituting (20) in the above expression we have:


oh0
I > 0;
ou0

8u0 2 Ku0

23

24

Fig. 3. Implementation of g{} via a high-gain feedback loop.

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

4. Quantifying process operability via LMI synthesis


Given the dynamic operability analysis framework
described in Section 3, we now focus on the issue of quantifying in detail the operability of P. The denition of the
dynamic operability assessment problem requires the signal
rh(t) to be constrained inside a certain pre-specied region.
This constraint on rh(t) represents the inuence that the static nonlinearity h{} has on the operability of the overall
nonlinear process P. Intuitively, we see that the smaller
the invertibility region Hrh the stronger the limitation
imposed by the static nonlinearity on the operability of
the process. Hence, a key issue in quantifying the dynamic
operability of P is to nd an eective and ecient way to
deal with the constraint rh t 2 Hrh  Xy h .
One possible approach is to consider the constraint
rh t 2 Hrh explicitly by solving a constrained optimisation
problem in the time domain. This is similar to the optimisation approach adopted in the simultaneous process and
control design methodology of Perkins and co-workers
[1517]. However, as mentioned earlier, the solution to
these type of nonlinear constrained optimisation problems
requires a large number of computations and, therefore, is
limited to relatively small-dimensional systems.
An alternative is to solve an optimisation problem in the
frequency domain, replacing the time-domain hard constraint rh t 2 Hrh by a related (though generally more conservative) constraint on the system gain of the IMC
controller Qp(s). This approach requires a limited amount
of computations and can be easily implemented using commercial numerical computation software (such as Matlab). The key issue is to determine which system norm
is appropriate to approximate the hard time-constraint
rh t 2 Hrh .
To discuss this point in more detail we rst need to
dene a reference point against which all vector and system
norm calculations are performed. Thus, without loss of
generality, it is assumed that the process P initially rests
at a given operating point u ; y h ; y  . Based on the chosen
operating point and the available information on the largest expected value of each process variable, we dene a new
set of scaled variables given by
ut  u
umax  u
y t  y h
~y h t , h
y hmax  y h
yt  y 
~y t ,
y max  y 
~
ut ,

25

Accordingly, we also have that:


rt  y 
y max  y 
rh t  y h
~rh t ,
y hmax  y h
~rt ,

26

163

From this point onwards we assume that Qp(s), Gp(s), g{}


and h{} have been scaled based on the above denitions.
Similarly, the constraint region Hrh is also assumed to be
scaled accordingly. We now consider the case when the
output disturbance d(t) in Fig. 2 has zero mean and can
be scaled as follows:
~ , dt
dt
d max

27

Thus, with reference to the IMC closed-loop in Fig. 2, we


~
have that ~rt dt,
since it is assumed that r(t) = y*.
A natural choice to express the constraint rh t 2
Hrh  Xy h is to consider the innity norm of ~rh t, i.e.


k~rh tk1 , max max j~rh;i sj
28
s

To illustrate, assume that ~rh t has only two dimensions.


Thus, to guarantee ~rh t 2 Hrh one needs to nd the largest
square centred at the origin and completely inscribed in
Hrh . If 2b > 0 is the length of the sides of the square, then
we require k~rh tk1 6 b. In addition, it is clear from the
~
scaling of the disturbance in (27) that kdtk
1 6 1. We conclude that ~rh t 2 Hrh if we impose the following constraint
on the system L1 norm of the IMC controller Qp(s):
Z 1
k~rh tk1
kqp tk1 ,
qp t dt sup
6b
29
~
~
kdtk
1
dt
1
where qp(t) is the impulse response of Qp(s). Unfortunately,
the results that can be obtained using the system L1 norm
are, in general, very conservative [49]. Also, the system L1
norm is not amenable to ecient and simple computation
[56]. A second choice is to consider a slight variation of
the innity norm of ~rh t dened in (28). In particular, consider the following signal norm:
k~rh tk12 , sup k~rh sk

30

where, in q
this
case, kk is the Euclidean vector norm, i.e.

T
k~rh sk ~rh s ~rh s. To illustrate, consider again a
two dimensional example. Thus, to guarantee ~rh t 2 Hrh
one needs to nd the largest circle centred at the origin
and completely inscribed in Hrh . If a > 0 is the radius of
the circle then we require k~rh tk12 6 a. Fig. 4 shows a
graphical interpretation of the signal norms k~rh tk1 and
k~rh tk12 discussed here when ~rh t has only two dimensions and when ~rh t is constrained inside an arbitrary region Hrh  Xy h . Fig. 4 shows that the innity-2 norm
k~rh tk12 is only slightly more conservative than the innity norm k~rh tk1 . Fig. 4 also shows that a time domain
optimisation approach that considers the constraint
~rh t 2 Hrh explicitly may, in principle, obtain dynamic
operability results that are more realistic. This is because
~rh t is allowed to vary inside the whole region Hrh and
not only inside its inner square or circle approximations.
However, as discussed earlier, a full time-domain optimisation approach exhibits other sources of conservatism
(e.g. nite horizon, xed disturbance prole, convex

164

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

Matrix Inequality (LMI) formalism [38] in combination


with convex optimisation routines [39]. Thus, we suggest
to use the generalised H2 system norm to approximate
the hard time-constraint rh t 2 Hrh  Xy h .
The main performance criterion for operability that we
will consider in our study is the worst-case integral square
of the error signal ~et ~rt  ~y t (wISE). When
~rh t 2 Hrh we have that g{} = h1{} and we can write:
^
^y s I  Gp sQp sds

34

where ^y s is the Laplace transform of the scaled variable


^ is the Laplace transform of dt.
~
~y t and ds
Thus the error signal is given by
^
^es I  Gp sQp sds

35

Fig. 4. Graphical interpretation of k~rh tk1 and k~rh tk12 for two
arbitrary regions Xyh and Hrh .

and the worst-case integral square error (wISE) is given by


the H1 norm of the sensitivity S(s) = I  Gp(s)Qp(s), i.e.

approximation of the constraints, etc.) and requires large


amount of computation. The degree of conservatism introduced by the innity-2 norm approximation in (30) of the
time domain constraint ~rh t 2 Hrh will clearly depend on
the specic shape of the constraint regions and the specic
operating point u ; y h ; y  chosen for operability analysis.
In a two dimensional case the required value of a > 0
can be easily found by direct inspection. In a higher dimensional case a can be determined by calculating the vector
norm of each point (in practice, of a grid of points) on
the contour of the constraint regions Hrh and then assigning to a the smallest of these values. If it is assumed that
~ has bounded energy:
the disturbance signal dt
s
Z

wISE , kSsk1 kI  Gp sQp sk1 sup

~
kdtk
2 ,

~ T dt
~ dt t < 1
dt

31

~
dt

k~etk2
~
kdtk
2

36

Both the H1 norm of the sensitivity S(s) and the generalised H2 norm of the IMC controller Qp(s) can be computed if some knowledge about the energy of the
~ exciting the closed-loop is available
disturbance signal dt
~
and provided this energy is bounded, i.e. kdtk
2 t < 1
see (31). In practice, this type of information can be easily
estimated from historical process data.
Thus, the dynamic operability of the nonlinear process
P can be quantied in terms of the worst-case integral
square error (wISE) as follows:
4.1. Dynamic operability analysis for nonlinear processes
(wISE)

then a system norm that relates the innity-2 norm of ~rh t


~ is the
in (30) to the above L2 norm of the disturbance dt
so-called generalised H2 norm given by
kQp skg sup
~
dt

k~rh tk12
~
kdtk
2

32

The generalised H2 norm was originally introduced by


Wilson [35] and subsequently discussed in detail by Rotea
[36]. Cao and Yang [37] have also suggested the generalised
H2 norm for operability analysis of linear systems when
the maximum amplitude of the manipulated variables is
limited. We see that the generalised H2 norm of the linear
system Qp(s) is, by denition, the largest vector norm of the
output signal ~rh t for all time t > 0, when the input signal
~
dt
has bounded L2 norm. Thus, to guarantee
k~rh tk12 6 a one can impose the following constraint
on the generalised H2 norm of Qp(s):
kQp skg 6 at1

33

One important advantage of the generalised H2 system


norm is that it can be computed eciently using a Linear

Consider a stable nonlinear process P and assume that P


can be modelled as in (6). Let the static nonlinearity h{}
have a well dened inverse for all rh t 2 Hrh with Hrh
dened in (15). Consider a given process operating point
u ; y h ; y  and re-scale the process model based on (25)
~
and (26). Assume that the scaled disturbance dt
has
~
bounded energy, i.e. kdtk2 6 t.
Then, the dynamic operability of the nonlinear process
P can be quantied by
J opt min kI  Gp sQp sk1
Qp s

37

subject to:
kQp kg 6 at1

38

where a is the radius of the largest hyper-sphere centred at


the origin that is entirely inscribed in Hrh .
The above operability analysis can be solved eciently
using a Linear Matrix Inequality (LMI) approach [38].
We next review the main steps required to obtain a solution
to the proposed operability analysis via LMI synthesis. The
IMC controller Qp(s) is rst parameterised as follows:

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

Qp s KsI Gp sKs

1

39

where K(s) is an equivalent closed-loop linear controller.


The process linear dynamics Gp(s) are then extended using
a state-space generalised plant G(s) formalism given by
[57]:

165

45
Observe that Gp(s) is required to be strictly proper in order
to guarantee the well-posedness of the closed-loop in Fig. 5
[39]. Similarly, let a state-space model of the lter We(s) in
(42) be:
46

40
where z(t) is a two component vector that contains the output signals that we wish to minimise or constrain (in our
case the error ~et and the input ~
ut to the linear plant
Gp(s)) whilst v(t) is the variable available for feedback to
the controller K(s). Fig. 5 shows a feedback interpretation
of the parameterisation of Qp(s) required for LMI synthesis. The shaded region indicates the generalised plant G(s)
considered in this case. In particular z(t) is given by

ze t
zt
41
~
ut
where ze(t) is a ltered version of the error ~et. The lter
We(s) is taken to be:


1
W e s diag
u1
42
su
This choice of We(s) is required to guarantee that the optimal IMC controller Qopt
p s that minimises the wISE performance criterion in (37) satises:
1
Qopt
p 0 Gp 0

43

so that, in steady-state, when r = y* we have:


1 


rhss Qopt
p 0y Gp 0 y y h

44

as required by the choice of the operating point of interest


u ; y h ; y  . Next, assume that the linear dynamics Gp(s)
have the following state-space representation:

Then, with reference to Fig. 5, the generalised plant G in


(40) is readily seen to be:

47

It can be shown that the H1 norm in (37) is smaller than


c > 0 if and only if there exists a symmetric matrix P such
that the following linear matrix inequalities are satised
[39]:
0 T
1
 P PA
 PB
T
 C
A
B T
T C
cI D
@ B P
A<0


C
D cI
P >0

48

 B;
 D
 C;
 is a state space representation of the senwhere A;
sitivity S(s). Similarly, the generalised H2 norm of Qp(s) is
smaller than a > 0 if and only if there exists a symmetric
matrix P such that the following linear matrix inequalities
are satised [39]:
!
~ PB
~TP P A
~
A
<0
~TP
B
I
!
~T
49
P C
>0
2
~
C aI
~ 0
D
~ B;
~ D
~ C;
~ is a state space representawhere, in this case, A;
tion of Qp(s). Using the LMIs in (48) and (49) in combination with the generalised plant G(s) description in (47) and
the parameterisation of Qp(s) in (39) gives rise to a set of bilinear matrix inequalities (BMIs). Unfortunately bi-linear
matrix inequalities cannot be solved eciently. However,
the BMIs can be converted into a set of LMIs by using
the following variable transformations similar to those proposed by Scherer et al. [39]:
^ , NAK M T NBK CX YBC K M T YA B2 DK C 2 X
A
^ , NBK YB2 DK
B
^ , C K M T DK C 2 X
C
^ , DK
D

Fig. 5. Parameterisation of IMC controller Qp(s) for LMI synthesis.

50

166

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

where (AK, BK, CK, DK) are the state-space matrices of the
controller K(s) in (39), N and M are auxiliary variables
and X and Y are auxiliary decision variables. With these
variable transformation, solving the dynamic operability
analysis (wISE) in (37) and (38) is equivalent to solving
the following convex optimisation problem:

with K chosen as in (20). Identify the region


Hrh h0 fH0u g in which the static nonlinearity h 0 {u 0 } is
guaranteed to be invertible based on the result in Theorem 3.4.
5. Calculate the required upper bound a > 0 for the gener~
alised H2 norm of Qp(s) assuming kdtk
2 1.
6. Solve the convex optimisation problem based on LMI
synthesis in Eqs. (51)(55) and compute the H1 norm
of the achieved sensitivity function S(s).
7. Repeat steps 46 for as many dierent operating points
as required. Alternatively, if dierent process designs
need to be compared for the same operating point

4.2. Dynamic operability analysis (wISE) via LMI synthesis


min

^ B;
^ D;X;Y
^ C;
^
A;

51

subject to the Linear Matrix Inequalities:


0

^ B2 C
^ T
AX XAT B2 C
B
^ A B2 DC
^ 2 T
B
A
B
B
^ 21 T
B1 B2 DD
@
^
C 1e X D12e C


T
^ 2 BC
^ 2 T
A Y YA BC
^ 21 T
YB1 BD
^ 2
C 1e D12e DC



cI
^ 21
D11e D12e DD

C
 C
C<0
 C
A

52

cI

and

^ B2 C
^ T
AX XAT B2 C

B
T
T
^ A B2 DC
^ 2 BC
^ 2 T
^ 2
@
A Y YA BC
A
^ 21 T
^ 21 T
B1 B2 DD
YB1 BD
0
1
X


B
C
I
Y

@
A>0
2
^
^
C 1u X D12u C C 1u D12u DC 2 a I
^ 21 0
D11u D12u DD

Linear Matrix Inequalities are symmetric; thus the terms


written with * in the above expressions need to be
replaced by the transpose of the corresponding sub-matrices below the main diagonal. The variables in bold face are
the decision variables in the optimisation. Finally, we provide a step by step procedure that summarises the dynamic
operability analysis for stable nonlinear processes proposed
in this paper.

C
 A<0
I

53

54
55
u ; y h ; y  , repeat steps 16 with a new Hammerstein
model for each new process design. However, to make
the results comparable use the same scaling for each
design in step 2.
8. Compare the H1 norm of the achieved sensitivity functions S(s) for each operating point of interest or alternative process design.

Procedure 4.1 (Dynamic operability analysis for nonlinear


systems).
5. Illustrative examples and discussion
1. Approximate the nonlinear process P using a Hammerstein-type model as in (6).
2. Select an operating point u ; y h ; y  at which to perform
the operability analysis.
3. Scale the process model based on (25) and (26).
4. Consider the scaled static nonlinearity
y h h0 fu0 g , hfKu0 g

56

To illustrate the dynamic operability analysis for nonlinear processes described in this paper, we present two case
studies. The rst example is a neutralisation process similar
to that studied by Lakshminarayanan et al. [25]. These type
of processes have been shown to be adequately described
by Hammerstein models such as that in (6). The second
example considers a more complex multi-unit reactor-sep-

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

167

3
2

1
0

2
1

1
0.5

0.5

0.5

0.5
1

u Base
2

0.5

0.5

0.5
u Base

u1 Acid

0.5
u1 Acid

Fig. 6. Static nonlinearity h{} for the neutralisation example: (a) yh1 and (b) yh2.

aration process similar to that investigated by Samyudia


et al. [58] and Lee et al. [59].

0:6055

0:0199

0:0059

6
6 0:0103
B6
6 0:2560
4

C

0:0807

0:1669

0:1235

0:0078

0:0219

0:0727

57

7
0:0269 7
7
0:1072 7
5

0:6474

0:1380

1:8553

0:4552

0:8054 0:8953

59

y h2 0:0031u1 u2 0:2896u21 0:1983u22

The state space matrices that dene the linear dynamics


Gp(s) of the process are given by
2
3
0:0283 0:0083 0:0045
0:0016
6
7
6 0:0120 0:0345 0:0049 0:0048 7
6
7
A6
0:0700 0:0874 0:0218 7
4 0:3695
5
0:8926

0:5663u1 u2 0:0322u31
0:0326u32 0:1987u1 u22  0:2144u21 u2 ;

5.1. Neutralisation process

y h1 u1  0:3056u2  0:349u21  0:1719u22

58

 0:5149u1 u2  0:0356u31 0:0677u32


 0:2032u1 u22 0:1452u21 u2 ;
The process inputs are the acid (u1) and base ow rates (u2),
whilst the process outputs are the level (y1) and the pH (y2)
of the liquid in the well stirred neutralisation tank. All
these variables are deviation variables from the nominal
operating conditions. The time base is seconds. To perform
the operability analysis we select the operating point of
interest to be the origin (Step 2). The model has been
appropriately scaled, thus we can omit Step 3 in this example. Fig. 6 shows the static nonlinearity h{} when the inputs u1 and u2 are contained in the available input space
(AIS) dened by the following region:
AIS , fu1 ; u2 j  1 6 u1 6 1 and

 1 6 u2 6 1g
60

The result of mapping the AIS into the process output space
and the static nonlinearity output space is shown in Fig. 7.

while the static nonlinearity h{} is given by

3
2.5
2.5
2
2

1.5
1
yh2

y2

1.5
1

0.5

0.5

0.5

0.5
1

0
y1

1.5
3

2.5

1.5

1
yh1

0.5

0.5

Fig. 7. Achievable output space (AOS) in the process output space (y1y2) and in the static nonlinearity output space (yh1yh2).

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

The static nonlinearity h{} in (59) is not amenable to a


simple analytical inversion. Thus, we will use the result of
Theorem 3.4 to determine the regions Hu (dened in the
input space of the static nonlinearity, see (14)) and Hrh
(dened in the output space of the static nonlinearity) in
which an inverse of the static nonlinearity exists and is well
dened. We nd that, in this example, the region Hu coincides with the available input space (AIS) in (60). Thus, the
region Hrh coincides with the achievable output space
(AOS) shown in Fig. 7(b). We conclude that, in this case,
the condition that guarantees the existence of the static
nonlinearity inverse does not impose any additional constraint to that already imposed by the static nonlinearity
output range Xy h , AOS since Hrh  Xy h .
Next, we proceed with Step 5 and estimate the radius a
of the largest circle centred at the origin that is entirely
inscribed inside the constraint region Hrh . This is shown
in Fig. 8 where it is found that a = 0.67. Observe from
Fig. 8 that using the generalised H2 norm of Qp(s) to
approximate the constraint ~rh t 2 Hrh introduces a certain
degree of conservatism. However, if we were to use the
innity norm of ~rh t, as in (28), to approximate the constraint we would have to inscribe a square (instead of a circle) centred at the origin inside the region Hrh . In either
case, we see that the results would be comparable. Still,
we prefer to use the generalised H2 norm due to its ecient
computation via LMIs.
Based on the estimated upper bound a of the generalised
H2 norm of Qp(s) we solve the dynamic operability analysis (worst-case ISE) via LMI synthesis in Eqs. (51)(55).
The solution to the optimisation problem yields a worstcase ISE of:
kSsk1 kI  Gp sQp sk1 1:6663

61

Fig. 9 shows the largest singular value in the frequency domain of the achieved sensitivity S(s) = I  Gp(s)Qp(s). This
frequency domain response provides additional informa2.5

1.5

rh

yh2

0.5

0.5
1
3

2.5

1.5

y h1

0.5

Fig. 8. The largest circle inscribed in Hrh .

0.5

10

0
Maximum singular Value (dB)

168

-10

-20

-30

-40

-50
10-4

10 -3

10 -2

10 -1

10 0

10 1

Frequency (rad/s)

Fig. 9. Frequency response of the sensitivity S(s) = I  Gp(s)Qp(s) for the


neutralisation example.

tion on the performance of the closed-loop. In particular,


for a given frequency, it indicates the level of attenuation
~ achieved
(or amplication) of the output disturbance dt
by the closed-loop. In addition, the plot in Fig. 9 provides
an indication of the achieved closed-loop bandwidth (w0)
or, equivalently, the achieved closed-loop settling time


 w40 . In this example, we see that w0  2 102 rad/s;
thus the closed-loop achieves a settling time of approximately 200 s.
5.2. Reactor-separation process
The reactor-separation process considered in this example is shown in Fig. 10. A similar multi-unit process was
studied by Samyudia et al. [58] and Lee et al. [59]. A brief
description of the process follows, which is adopted from
[59]: components A and B are fed into a jacketed CSTR
where a rst order, endothermic, irreversible reaction
A + B ! P + G takes place in the liquid phase. It is
assumed that component B is completely consumed by
the reaction. The majority of the by-product G is extracted
in a single stage liquidliquid extractor using solvent C.
The ranate from the extractor is a mixture of A and P.
A distillation column separates the product P as the overhead stream while the bottom ow of A is discarded. The
reactor-separation process considered in this example differs from that in [58,59] in that the bottom ow of component A from the distillation column is not recycled into the
CSTR. In addition, the process model does not include a
ash-drum unit used to purge the by-product G from the
extract ow.
A detailed description of the nonlinear dynamics of this
multi-unit process can be found in [59], including all the
relevant assumptions and parameters values. The process
nonlinear model contains 17 states and up to 5 inputoutput pairs. For illustration we will consider only 2 inputs
and 2 outputs. The inputs are the reux owrate L(u1)
and fresh feed F4 of component A into the CSTR(u2).

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

F4

169

Reflux L

B feed
Solvent C

Extract

Column

Steam

Extractor

Reactor

Steam

Fig. 10. Reactor-separator process owchart.

yh1

yh2

60

80

55

70

50

60

45

50

40

40

35

30

30
100

20
100
90

F4 (kmol/h)

560
555

80

550

70

545
60

90

F4 (kmol/h)

560
555

80

550

70

L (kmol/h)

540

545
60

540

L (kmol/h)

Fig. 11. Static nonlinearity h{} for the reactor-separation process: (a) yh1 and (b) yh2.

The outputs are the composition of A in the reactor xA3(y1)


and the distillate composition xD(y2). The time base is
hours.
The process nonlinear dynamics are approximated using
a Hammerstein-type model as in (6) (Step1). Fig. 11 shows
the static nonlinearity h{} of the model. The linear
dynamic part Gp(s) of the model is given by the following
transfer function matrix:
2
Gp s 4

is guaranteed to be invertible (Step 4). For convenience,


only points inside the Desired Input Space (DIS) shown
in Fig. 12(a) are considered to compute Hrh . The DIS in
Fig. 12(a) maps into a rectangular Desired Output Space
(DOS) given by
DOS , fy 1 ; y 2 j 0:2 6 y 1 6 0:35 and 0:8 6 y 2 6 1g
3

0:0068098s223:3
s1:274s234:8

10:7518s48:68s148:5
s227:9s106:2s43:12s4:051

2:2805s0:1355s2 3:862s3:898
s43:12s4:051s1:897s1:274s1:069

The operating point selected for operability analysis (Step


2) is given by

548:51
40:67
0:29



u
; yh
; y
63
83
56:9
0:9
After appropriately scaling the model variables (Step 3) we
compute the region Hrh in which the static nonlinearity h{}

64

62

Fig. 12(b) shows the invertibility region Hrh obtained in this


example. Observe that due to the scaling of the model variables (Step 3) the selected operating point in (63) is now
mapped into the origin. The required upper bound a > 0
for the generalised H2 norm of Qp(s) is given by the radius
of the largest circle entered at the origin and inscribed in
Hrh (Step 5). This is also shown in Fig. 12(b). We have that

170

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

Desired Input Space (DIS)

rh

100

95
5
90

h2

85

u*

Scaled y

u : Feed flowrate F4 of component A (kmol/h)

10

80

-5

75
-10
70

65
540

-15
542

544

546

548

550

552

554

556

-15

-10

-5

10

15

Scaled yh1

u1 : Reflux flowrate L (kmol/h)

Fig. 12. Desired Input Space (DIS) and the invertibility region Hrh of the static nonlinearity h{} for the reactor-separation process.

can be readily applied during process design and can be


solved eciently using a limited amount of computations.
By exploiting a Hammerstein/Wiener model approximation to the process dynamics in combination with an
extended IMC framework we have investigated the interaction between the static nonlinearities and the system
dynamics on the operability of the overall process. We have
shown that the eect of the static nonlinearities on operability can be translated into a time domain constraint on
the signals that appear in the closed-loop. The dynamic
operability of the process is then quantied using a linear
matrix inequality (LMI) optimisation approach that minimises the required performance criterion subject to the constraint imposed by the static nonlinearities.

10

Maximum singular Value (dB)

0
10
20
30
40
50
60
70
80

10

10

10

10

10

10

10

10

Frequency (rad/h)

Acknowledgements

Fig. 13. Frequency response of the sensitivity S(s) = I  Gp(s)Qp(s) for the
reactor-separation example.

a = 9.9. Based on the estimated upper bound a of the generalised H2 norm of Qp(s) we solve the dynamic operability
analysis (worst-case ISE) via LMI synthesis in Eqs. (51)
(55). The solution to the optimisation problem yields a
worst-case ISE of:
kSsk1 kI  Gp sQp sk1 1:5938

65

Fig. 13 shows the largest singular value in the frequency


domain of the achieved sensitivity S(s) = I  Gp(s)Qp(s).
Observe that the achieved closed-loop bandwidth (w0) is
w0  4 101 rad/h.
6. Conclusion
In this paper, we have described a dynamic operability
approach for stable nonlinear processes. The approach

The support of the Australian Research Council (Grant


DP0558755) is gratefully acknowledged.
References
[1] S. Skogestad, I. Postlethwaite, Multivariable Feedback Control,
second ed., John Wiley & Sons Ltd., 2005.
[2] Y.C. Cheng, C.C. Yu, Eects of feed tray locations to the design of
reactive distillation and its implication to control, Chemical Engineering Science 60 (17) (2005) 46614677.
[3] W.L. Luyben, B.D. Tyreus, M.L. Luyben, Plantwide Process Control,
McGraw-Hill, New York, 1998.
[4] K.L. Wu, C.C. Yu, Reactor/separator processes with recycle 1.
Candidate control structure for operability, Computers & Chemical
Engineering 20 (11) (1996) 12911316.
[5] M. Morari, Design of resilient processing plants III. A general
framework for the assessment of dynamic resilience, Chemical
Engineering Science 38 (11) (1983) 18811891.
[6] S. Skogestad, M. Morari, Eect of disturbance directions on closedloop performance, Industrial & Engineering Chemistry Research 26
(1987) 20292035.

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172


[7] E.H. Bristol, On a new measure of interaction for multivariable
process control, IEEE Transactions on Automatic Control 11 (1966)
133134.
[8] S. Skogestad, M. Morari, Implications of large RGA elements on
control performance, Industrial & Engineering Chemistry Research
26 (11) (1987) 23232330.
[9] M. Hovd, S. Skogestad, Simple frequency-dependent tools for
control-system analysis, structure selection and design, Automatica
28 (5) (1992) 989996.
[10] M. Guay, P.J. McLellan, D.W. Bacon, Measurement of nonlinearity
in chemical process control systems: the steady state map, The
Canadian Journal of Chemical Engineering 73 (6) (1995) 868882.
[11] A.J. Stack, F.J. Doyle, The optimal control structure: an approach to
measuring control-law nonlinearity, Computers & Chemical Engineering 21 (9) (1997) 10091019.
[12] A. Helbig, W. Marquardt, F. Allgower, Nonlinearity measures:
denition, computation and applications, Journal of Process Control
10 (23) (2000) 113123.
[13] T. Schweickhardt, F. Allgower, Quantitative nonlinearity assessment
an introduction to nonlinearity measures, in: P. Seferlis, M.C.
Georgiadis (Eds.), The Integration of Process Design and Control,
Computer-Aided Chemical Engineering, vol. 17, Elsevier BV,
Amsterdam, 2004, pp. 7695 (Chapter A3).
[14] M. Nikolau, P. Misra, Linear control of nonlinear processes: recent
developments and future directions, Computers & Chemical Engineering 27 (2003) 10431059.
[15] V. Sakizlis, J.D. Perkins, E.N. Pistikopoulos, Recent advances in
optimization-based simultaneous process and control design, Computers & Chemical Engineering 28 (10) (2004) 20692086.
[16] V. Sakizlis, J.D. Perkins, E.N. Pistikopoulos, Parametric controllers
in simultaneous process and control design optimization, Industrial &
Engineering Chemistry Research 42 (20) (2003) 45454563.
[17] J.D. Perkins, The integration of design and control the key to future
processing systems? In: 6th World Congress of Chemical Engineering,
Melbourne, Australia, 2001.
[18] C. Georgakis, D. Uzturk, S. Subramanian, D.R. Vinson, On the
operability of continuous processes, Control Engineering Practice 11
(2003) 859869.
[19] D. Uzturk, C. Georgakis, Inherent dynamic operability of processes:
general denitions and analysis of SISO cases, Industrial & Engineering Chemistry Research 41 (2002) 421432.
[20] D.R. Vinson, C. Georgakis, A new measure of process output
controllability, Journal of Process Control 10 (2000) 185194.
[21] O.J. Rojas, J. Bao, P.L. Lee, Linear control of nonlinear processes:
the regions of steady-state attainability. Industrial & Engineering
Chemistry Research 45 (22) 75527565.
[22] E. Ekawati, P.A. Bahri, The integration of the output controllability
index within the dynamic operability framework in process system
design, Journal of Process Control 13 (2003) 717727.
[23] P.A. Bahri, J.A. Bandoni, J.A. Romagnoli, Integrated exibility and
controllability analysis in design of chemical processes, AIChE
Journal 43 (4) (1997) 9971015.
[24] E. Eskinat, S.H. Johnson, W.L. Luyben, Use of Hammerstein models
in identication of nonlinear-systems, AIChE Journal 37 (2) (1991)
255268.
[25] S. Lakshminarayanan, S.I. Shah, K. Nandakumar, Identication of
Hammerstein models using multivariate statistical tools, Chemical
Engineering Science 50 (22) (1995) 35993613.
[26] R.K. Pearson, M. Pottmann, Gray-box identication of blockoriented nonlinear models, Journal of Process Control 10 (4) (2000)
301315.
[27] J. Abonyi, R. Babuska, M.A. Botto, F. Szeifert, L. Nagy, Identication and control of nonlinear systems using fuzzy Hammerstein
models, Industrial & Engineering Chemistry Research 39 (11) (2000)
43024314.
[28] O. Weitz, D.R. Lewin, Dynamic controllability and resiliency
diagnosis using steady state process owsheet data, Computers &
Chemical Engineering 20 (4) (1996) 325335.

171

[29] W.D. Seider, J.D. Seader, D.R. Lewin, Product and Process Design
Principles: Synthesis, Analysis, and Evaluation, John Wiley & Sons,
2003.
[30] K.H. Chan, J. Bao, W.J. Whiten, A new approach to control of
MIMO processes with static nonlinearities using and extended IMC
framework, Computers & Chemical Engineering 30 (2) (2005) 329
342.
[31] M. Morari, E. Zariou, Robust Process Control, Prentice Hall,
Englewood Clis, New Jersey, 1989.
[32] H.K. Khalil, Nonlinear Systems, third ed., Prentice Hall, Upper
Saddle River, New Jersey, 2002.
[33] M. Vidyasagar, Nonlinear Systems Analysis. Classics in Applied
Mathematics, second ed., SIAM, Philadelphia, 2002.
[34] C.A. Desoer, M. Vidyasagar, Feedback Systems: InputOutput
Properties, Academic Press, New York, 1975.
[35] D.A. Wilson, Convolution and Hankel operator norms for linear
systems, IEEE Transactions on Automatic Control 34 (1989) 9497.
[36] M.A. Rotea, The generalized H2 control problem, Automatica 29 (2)
(1993) 373385.
[37] Y. Cao, Z. Yang, Multiobjective process controllability analysis,
Computers & Chemical Engineering 28 (2004) 8390.
[38] S. Boyd, L. El Ghaoui, E. Feron, V. Balakrishnan, Linear Matrix
Inequalities in System and Control Theory, SIAM Studies in Applied
Mathematics, vol. 15, SIAM, Philadelphia, 1994.
[39] C. Scherer, P. Gahinet, M. Chilali, Multiobjective output-feedback
control via LMI optimization, IEEE Transactions on Automatic
Control 42 (7) (1997) 896911.
[40] B.D.O. Anderson, From Youla-Kucera to identication, adaptive
and nonlinear control, Automatica 34 (12) (1998) 14851506.
[41] M.M. Seron, J.H. Braslavsky, G.C. Goodwin, Fundamental Limitations in Filtering and Control, Springer-Verlag, London, 1997.
[42] J. Chen, Logarithmic integrals, interpolation bounds, and performance limitations in MIMO feedback systems, IEEE Transactions on
Automatic Control 45 (6) (2000) 10981115.
[43] G.C. Goodwin, S.F. Graebe, M.E. Salgado, Control System Design,
Prentice Hall, Upper Saddle River, NJ, 2001.
[44] J.S. Freudenberg, D.P. Looze, Right half plane poles and zeros and
design tradeos in feedback systems, IEEE Transactions on Automatic Control 30 (6) (1985) 555565.
[45] R.H. Middleton, Trade-os in linear control system design, Automatica 27 (2) (1991) 281292.
[46] J. Chen, Sensitivity integral relations and design trade-os in linear
multivariable feedback systems, IEEE Transactions on Automatic
Control 40 (10) (1995) 17001716.
[47] M.M. Seron, J.H. Braslavsky, P.V. Kokotovic, D.Q. Mayne, Feedback limitations in nonlinear systems: from Bode integrals to cheap
control, IEEE Transactions on Automatic Control 44 (4) (1999) 829
833.
[48] J. Chen, L. Qiu, O. Toker, Limitations on maximal tracking accuracy,
IEEE Transactions on Automatic Control 45 (2) (2000) 326331.
[49] W.Z. Zhang, J. Bao, P.L. Lee, Process dynamic controllability
analysis based on all-pass factorization, Industrial & Engineering
Chemistry Research 44 (2005) 71757188.
[50] E. Zariou, H.W. Chiou, On the dynamic resiliency of constrained
processes, Computers & Chemical Engineering 20 (4) (1996) 347
355.
[51] Y. Cao, D. Biss, J.D. Perkins, Assessment of inputoutput controllability in the presence of control constraints, Computers & Chemical
Engineering 20 (4) (1996) 337346.
[52] J. Chen, S. Hara, G. Chen, Best tracking and regulation performance
under control energy constraint, IEEE Transactions on Automatic
Control 48 (8) (2003) 13201336.
[53] T. Perez, G.C. Goodwin, M.M. Seron, Performance degradation in
feedback control due to constraints, IEEE Transactions on Automatic Control 48 (8) (2003) 13811385.
[54] R. Baker, C.L.E. Swartz, Rigorous handling of input saturation in
the design of dynamically operable plants, Industrial & Engineering
Chemistry Research 43 (18) (2004) 58805887.

172

O.J. Rojas et al. / Journal of Process Control 17 (2007) 157172

[55] T.M. Apostol, Calculus: Multi-Variable Calculus and Linear Algebra


with Applications, vol. 2, John Wiley & Sons, 1969.
[56] M.A. Dahleh, I.J. Diaz-Bobillo, Control of Uncertain Systems. A
Linear Programming Approach, Prentice-Hall, 1998.
[57] K. Zhou, J.C. Doyle, K. Glover, Robust and Optimal Control,
Prentice-Hall, Upper Saddle River, NJ, 1996.

[58] Y. Samyudia, P.L. Lee, I.T. Cameron, M. Green, Application of


multi-unit control analysis and design to a reactor/separation process.
In: AIChE Symposium Series, number 316, 1997.
[59] P.L. Lee, H.Z. Li, I.T. Cameron, Decentralized control design for
nonlinear multi-unit plants: a gap metric approach, Chemical
Engineering Science 55 (2000) 37433758.

You might also like