You are on page 1of 8

Chemical Engineering Science 55 (2000) 2393}2400

On the computation of a nonlinearity measure using


functional expansions
Kenneth R. Harris!,1, M. Celeste Colantonio", Ahmet Palazogy lu!,*
!Department of Chemical Engineering and Materials Science, University of California at Davis, 1 Shields Avenue, Davis, CA 95616-5294, USA
"Centre for Process Systems Engineering, Imperial College, London SW7 2BY, UK
Received 4 August 1998; accepted 11 July 1999

Abstract
Nonlinearity measures quantify the degree of nonlinearity in dynamic systems. In this work, we present a novel approach to the
computation of these measures by functional expansions. Rather than calculating an exact nonlinearity measure which is often
computationally intensive, functional expansion (FEx) models are used to derive approximate lower and upper bounds on the
measure. The resulting bounds are algebraic in nature and provide insight into the e!ect of process parameters on the nonlinearity
measure. These concepts are illustrated by a benchmark CSTR problem. ( 2000 Elsevier Science Ltd. All rights reserved.
Keywords: Nonlinearity measures; Nonlinear operators; Functional expansion models

1. Introduction
Nearly all chemical processes exhibit some form of
nonlinearity that is rooted in reaction kinetics, thermodynamics and/or underlying transport processes. Common practice in process control is to implement linear
control laws for such processes due to the simplicity of
the resulting design and the transparency of the available
analysis tools. Theoretically, nonlinear control laws will
improve the overall quality of the control, but the question that remains unanswered is whether or not the
increased complexity and the potentially di$cult analyses associated with the nonlinear controller is worth the
increase in performance bene"ts. Moreover, the advantages of nonlinear control strategies are known to be
case-dependent and subjective. Consequently, there is
a clear trade-o! in solving such problems, but the designer often lacks su$cient metrics to make a de"nitive
decision.
The nonlinear behavior under closed-loop operation,
however, can originate from a variety of sources. Con-

* Corresponding author. Tel.: 001-530-752-8774; fax: 001-530-7521031.


E-mail address: anpalazoglu@ucdavis.edu (A. Palazogy lu)
1 Current address: AMD Corp., Sunnyvale, CA 94088.

sider the general formulation of the control problem,


starting with the objective function that re#ects an algebraic criterion to be minimized as a function of the
control law:
min F(r, y, u),

(1)

where the objective may involve the reference r3Rq,


manipulated input u3Rm, and the output y3Rq. This
optimization is subject to system dynamics, often
expressed as
x5 "f (x, u),

y"h(x)

(2)

with x3Rn, and where f (x, u) and h(x) are smooth vectorvalued functions of the appropriate dimensions. These
dynamics establish the equality constraints that de"ne
the hypersurface on which the system states and the
observed outputs should lie. The control problem also
includes operational restrictions in the form of inequality
constraints
p(r, y, u))0

(3)

that describe the feasible region of operation. It should be


clear that the nonlinear behavior of the system under
closed-loop could be due to any of these expressions, or
result from the computational solution of the control
problem. While this makes the assessment of the nonlinearity a more di$cult task, often the designer has some

0009-2509/00/$ - see front matter ( 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 0 9 - 2 5 0 9 ( 9 9 ) 0 0 5 1 4 - X

2394

K.R. Harris et al. / Chemical Engineering Science 55 (2000) 2393}2400

freedom in choosing the objective function, and the operational constraints are typically known. Then, the focus
shifts to the system dynamics that need to be well represented for maximizing performance bene"ts. Consequently, in the sequel we shall focus on the nonlinearities
arising from the system dynamics, in particular, Eq. (2).
In this work, our goal is to quantify the degree of
nonlinearity in process dynamics, thus providing a metric
that can ideally aid in the selection of an appropriate
control law. For example, linear model predictive control
(MPC) strategies (Prett & Garcia, 1988) are well known
for their constraint handling abilities when the process
dynamics is linear over the intended range of operation.
However, if the process exhibits substantial nonlinearities over this operation range, linear MPC may not be
the best choice for the control law. It is this very problem
that this study intends to address. More speci"cally, we
will consider the following design issues:
f Given the range of input moves and the selected
operating point, quantitatively how nonlinear is the
system?
f How does this degree of nonlinearity change as a function of the operating point and other system parameters?
f Although the system is nonlinear, is it possible to select
a linear control law that will achieve satisfactory performance?
The concept of the nonlinearity `sizea was "rst used to
demonstrate the e!ect of linearizing control laws (Desoer
& Wang, 1981). Subsequent research on these types of
measures for system nonlinearities has taken a variety of
approaches. In general, most studies quantify deviations
from linear model behavior or violations of the properties of linear systems. These measures have been computed to assess either open-loop nonlinearities or
control-relevant nonlinearities. While the former is directly related to the input/output map, the latter is closely
tied to the system inverse [8]. Clearly, open-loop nonlinearities would impact the controller design, however, the
control-relevant measures quantify nonlinear behavior
that is directly related to the closed-loop operation.
Open-loop nonlinearity measures have been developed based on a search for the best linear model for
a nonlinear process. How well this model "ts over
a prede"ned range of operation is used to determine the
degree of nonlinearity in the process. This framework has
been applied to analyze the degree of nonlinearity before
and after state feedback compensation, as well as serving
as an objective function to be minimized for approximate
feedback linearization (AllgoK wer, 1995a,b,1996). Others
have utilized Monte Carlo simulations together with an
inner product as a basis for the measure (Nikolaou,
1993). Quantifying violations of properties that are
known to hold for linear systems has also been investi-

gated to assess system nonlinearities (Haber, 1985), while


steady-state nonlinearities have been quanti"ed using
curvature metrics of the steady-state locus (Guay,
McLellan & Bacon, 1995).
Control-law nonlinearity measures have been de"ned
by analyzing the behavior of an optimal controller based
on the original system (Stack & Doyle, 1997). This has
the advantage of incorporating interactions among the
control objective, the control law, and the system dynamics into the computation. Consequently, the measure is
more relevant to the closed-loop operation of the system.
However, the resulting measure depends on the control
scheme used in the analysis. Often the appropriate control law is not known before the application of a nonlinearity measure, as this is the information that the
designer is seeking in the "rst place. Therefore, the outcome of the analysis implicitly includes certain assumptions on the class of applicable control laws.
Common to all these techniques is the reliance on
numerical simulations. The algorithms can be computationally intensive, requiring elaborate optimization
schemes and simulations. Furthermore, these calculations result in a discrete formulation of the nonlinearity
measures. Hence, determining the functional dependence
of the measure on process parameters can be ine$cient,
requiring iterative simulations to explore the relevant
parameter space.
Here, functional expansion (FEx) models are utilized
to yield a continuous formulation of an open-loop nonlinearity measure. Rather than using numerical simulations, FEx models provide a means to determine an
algebraic upper and lower bound on the measure. This
continuous nature of the analysis directly reveals the
dependence of the measure on process parameters in
a very e!ective manner. The information obtained from
this analysis can then be used to provide insight into
the selection of an appropriate operating point, range of
the operation, and ultimately the class of control laws to
be utilized. Note that the more general, open-loop nonlinearity measure will be utilized in this work. This eliminates the dependence of the selected control law used in
the control-law nonlinearity measure from the computation itself.
The paper is organized as follows: First, the FEx models will be reviewed, followed by the introduction of the
nonlinearity measure. The next section will discuss the
computation of the measure based on the FEx models.
Finally, a simulation example will be presented to illustrate the concepts developed.

2. Review of FEx models


Here, a brief overview of the FEx model is presented.
The type of functional expansion utilized in this work
was inspired by the work of Fliess, Lamnabhi and

K.R. Harris et al. / Chemical Engineering Science 55 (2000) 2393}2400

Lamnabhi-Lagarrigue (1983) and rely upon the Laplace}Borel (LB) transform and the shu%e product (P),
de"ned, respectively, as

1 =
F(x )"LB[ f (t)]"
f (t)e~t@x0 dt,
(4)
0
x
0 0~
LB[ f (t)g(t)]"LB[ f (t)]PLB[g(t)]"F(x )PG(x ).
0
0
(5)
Here x is the transformation variable of the LB trans0
form. BatiguK n, Harris and Palazogy lu (1997) outline the
properties and several means of computing the shu%e
product.
We will consider the dynamics of the system described
by Eq. (2), assuming it is a single input/single output
(SISO), fading memory system of the form
x5 "f (x, u)"Ax#bu#m(x, u),
y"h(x),

(6)

where x3Rn, u3R, and the matrices A and b are of


appropriate dimensions. The nonlinearities enter the system in the smooth polynomial vector-valued functions
m(x, u) and h(x). Should the system not be in the form of
Eq. (6), the Carleman linearization or Taylor expansion
can be utilized to approximate non-polynomial with
polynomial-type nonlinearities. Applying Eqs. (4) and (5)
yields a system realization in the LB domain

~1
~1
I
I
b;#
N(X, ;,P),
!A
!A
x
x
0
0
>"H(X,P).

X"

(7)

Note that the nonlinear functions h(x) and m(x, u) have


been replaced by their equivalents in the LB domain,
H(X,P) and N(X, ;,P), with the nonlinearities being expressed as shu%e operations using Eq. (5). Since Eq. (7)
cannot be solved explicitly for X, an expansion solution
is assumed for the state vector,
X"X #X #2#X .
(8)
0
1
q
Note that the "rst term in this series, X , is chosen to be
0
equivalent to the linearization of Eq. (6). After Eq. (8) is
computed recursively and combined with the output
relationship in Eq. (7), the resulting expression can be
converted back to the time domain yielding the output to
the system as a series of operators (Harris & Palazogy lu,
1997)
y[u]"y [u]#y [u]#2#y [u].
(9)
0
1
q
The details of this procedure, although straightforward,
are outlined in the associated references, and have been
omitted for brievity (Harris & Palazogy lu, 1997; Harris,
1998). Eq. (9) is a general FEx model valid for arbitrary
inputs u. In the development of the nonlinearity measure,
it will be convenient to consider, in addition to general

2395

inputs, a single sinusoidal input. This sinusoidal function


is given in the time and LB domains as
u(t)"sin [ut],
ux
0 .
;(x )"
(10)
0
1#u2x2
0
With this choice of the input, the general form of the FEx
model, at steady state, can be shown to be of the form
(Harris & Palazogy lu, 1997)
=
y(t)" + W exp[kujt],
k
k/0
=
"W # + A sin (kut#/ ).
(11)
0
k
k
k/1
Note that the second expression is a Fourier series
describing the steady-state behavior of the system in
response to sinusoidal forcing. For linear systems,
W "0, kO1. Consequently, focusing on the terms
k
W for which kO1 leads to the basis of a nonlinearity
k
measure as developed in the sequel.
A FEx model utilizing q terms is denoted FExq. This
presentation of the basics of FEx models has assumed
that a "rst principles derivation of the dynamics of the
system is available, as identi"cation of FEx models directly from input/output data is an area of current research. The reader is referred to Harris and Palazogy lu
(1997), BatiguK n et al. (1997), and Fliess et al. (1983) for the
details and formalities of the method.

3. Computation of nonlinearity measures


We shall consider systems in the form of Eq. (6), represented by nonlinear input/output operators mapping inputs belonging to a normed input space u3U onto
a
outputs y3Y
y"N(u).

(12)

Eq. (12) can represent static as well as dynamic operators,


but we restrict ourselves to causal, I/O stable systems.
We also assume, without loss of generality, that these are
unbiased, "nite power systems,
lim y(t)"0 for u(t),0 t,
t?=
lim DDuDD (R,
VT
T?=
lim DDyDD (R
VT
T?=
with the norm taken as

1
DDyDD "
:Ty2(t) dt.
VT
0

(13)
(14)
(15)

(16)

Based upon these preliminaries, AllgoK wer (1995a) de"nes


the nonlinear measure of the system as follows.

2396

K.R. Harris et al. / Chemical Engineering Science 55 (2000) 2393}2400

De5nition 1. The nonlinearity measure /U of an I/O


N
stable, causal system N : U PY for input signals
a
u3U-U is de"ned by the nonnegative number
a
DDG[u]!N[u]DD Y
p
/U "inf sup
(17)
N
DDuDD Ua
G
U
p
G| u|
with G: U PY a linear operator, DD ) DD a norm in
a
Y, DD ) DD Ua a norm in U and U the set of inputs.
p
a
Here we have chosen the system norms as
DDyDD Y " lim DDyDD ,
p
VT
T?=
(18)
DDuDD U " lim DDuDD .
VT
p
T?=
The above de"nition sets the size of a nonlinearity of an
input/output system as the largest di!erence between the
output of a nonlinear system N and a linear system G,
measured in a norm DD ) DD Y in Y. Note that the di!erence
p
is taken with respect to the worst case input in U and
with respect to the best linear approximation GH[u] of
the system for that input.
For linear systems, /U is zero. However, a zero nonlinN
earity measure does not imply that the system is linear
because the set U may not contain all admissible inputs.
It is important to remark that in general the value of the
nonlinearity measure can depend on the choice of the
input signals.
De"nition 2 (AllgoK wer, 1996) provides an alternate,
normalized measure that is often easier to interpret than
the measures as calculated using De"nition 1.
De5nition 2. The normalized nonlinear measure /M U of
N
the system described by Eq. (12) is given by
DDG[u]!N[u]DD Y
p ,
/M U "inf sup
(19)
N
DDN[u]DD Y
G
U
p
G| u|
where the signals and norms are as described before.
The de"nition of the normalized nonlinearity measure
provides an absolute measurement of the nonlinearities
within the system. This latter measurement is appealing
because the measure is bounded between zero and one.
Should the input set U encompass inputs that cause the
gain of the nonlinear system N[u] to change sign, then
the best linear approximation of the system is GH[u]"0,
since no linear model can predict this type of behavior.
Consequently, the measure as given by De"nition 2 is
unity in this case. Should G[u]"N[u] for all inputs in
U, then the measure is zero, indicating linearity in the
input set U.
We will now discuss the computation of the measure.
In particular, we will avoid the computational expense of
an exact calculation and rather focus on deriving lower

and upper bounds on the measure by utilizing the FEx


models reviewed above.
3.1. Lower bound computation
By taking the input set U to be a series of sinusoids
LB
parameterized by A and X
U "Mu D u(t)"A sin(ut), A3A,u3XN,
LB
A"MA3R` D A )A)A N,
.*/
.!9
X"Mu3R` D u )u)u N,
(20)
.*/
.!9
the steady-state output can be written as in Eq. (11),
=
y(t)"A # + A sin(kut#/ ).
(21)
0
k
k
k/1
An approximate lower bound for the nonlinear measure
of De"nition 1 can then be computed by (AllgoK wer,
1995a)

= A2(u, A)
1
/ULB * sup
A2 (u, A)# + k
.
0
N
2
(22)
A
X A
k/2
A| ,u|
This measure uses the quasi-linearization for the best
linear plant GH[u] and is derived using the principle of
harmonic balance (Vidyasagar, 1993). The lower bound
arises through the interchanging of the supremum and
in"mum operations in the derivation of Eq. (22).
Similarly, a lower bound on the normalized measure of
De"nition 2 can be formulated as (AllgoK wer, 1996)

A2 (u, A)
1
/M ULB * sup
1!
.
N
A2(u, A)
(u,
A)#+=
2A2
(23)
A
X
k
k/1
0
A| ,u|
In practice, the Fourier series is truncated at some
order, ideally high enough to capture the nonlinear behavior of the system. The coe$cients can be computed
numerically (AllgoK wer, 1995a), however, this leads to
a discrete formulation of the measure. As outlined in the
previous section, the FEx model framework provides
a means of computing the Fourier coe$cients of Eq. (21)
analytically (Harris & Palazogy lu, 1997), leading to a continuous and algebraic representation of the measure. The
analytical nature of the nonlinearity measure then reveals
the functional dependence of the measure on system
parameters in a continuous fashion. This will be illustrated further in the case study.
3.2. Upper bound computation
The nonlinearity measure can also be computed in
a more direct manner by using optimization techniques.
AllgoK wer has shown that the nonlinearity measure given
by De"nition 1 is a convex programming problem (AllgoK wer, 1995b). In this case, no restrictions have been
placed on the input set U chosen for the measure.
The solution to this problem has been addressed by
"rst simulating the nonlinear system to each input in the

K.R. Harris et al. / Chemical Engineering Science 55 (2000) 2393}2400

chosen input set. A model parameterization for the best


linear approximation GH[u] is then chosen, usually of the
form
n
G[u]" + h G [u],
i i
i/1

(24)

where G [u] is a set of stable, linear, continuous or


i
discrete systems. Using some appropriate programming
algorithm, the optimum linear system parameterized by
the set of h 's is selected while maximizing with respect to
i
the input set U.
Although the convexity of this optimization problem
eases the computational demand, for large systems this
demand can still be high. Additionally, this measure also
results in a discrete measure of the nonlinearity. Furthermore, the parameterization of GH[u], chosen somewhat
arbitrarily, must be able to accurately represent the set of
best linear approximations to the nonlinear system.
Here we avoid these di$culties by computing an upper
bound using the FEx models. This upper bound is
achieved by selecting the best linear model as a "xed
linear plant, rather than a parameterization of linear
plants. Here we choose the Jacobian linearization, which
in our experience leads to rather tight bounds on the
nonlinearity measure. When integrated with the FEx
models of Eq. (9), the upper bound on the measure as
calculated by a FExq model is simply
/UUB ) sup
N
u|UUB

DD+q y [u]DD Y
i/1 i
p
DDuDD U
p

(25)

and for the normalized measure the bound is


DD+q y [u]DD Y
i/1 i
p .
/M UUB ) sup
N
DD+q y [u]DD Y
U
UB
i/0
i
p
u|

(26)

Here, Eq. (9) has been substituted into De"nitions 1 and


2, so that the FEx model is used as the nonlinear operator N[u]. Through this use of the FEx model, the
resulting measure is continuous and algebraic in nature.
Additionally, the optimization problem itself is simpli"ed
as only algebraic relations are involved rather than sets
of di!erential equations.
Note that we have allowed for the possibility that the
input sets used in the calculation of the lower and upper
bounds, U and U , to be di!erent in general. Ideally,
LB
UB
the bounds on the measure should be computed with
respect to the same input set. Because a sinusoidal parameterization is required for the lower bound computation, this would then "x the input set used for the
calculation of the upper bound. Clearly, from De"nitions
1 and 2, the measure depends on the input set selected.
The input set must span the input range of interest
adequately, at least in a steady-state sense. The selection
of the input set will be explored further in the following
case study.

2397

Several assumptions are implicit in the above development of the computation of the measures. First, the FEx
model is truncated in practice. This truncation must be
carried out with enough terms to capture the nonlinearities in the system over the chosen input set U. The
convergence of these terms in the FEx model relies on the
assumption of a fading memory system. The theory of
fading memory systems has been developed in the context of the Volterra model (Boyd & Chua, 1985) and
its extension to the FEx model, while straightfoward
(Harris, 1998), is beyond the scope of this work.
Additionally, depending upon the input set chosen and
the complexity of the system itself, the computation of the
Fourier coe$cients for the lower bound and the FEx
models for the upper bound can become quite involved.
This complexity can be alleviated to some extent by the
use of a symbolic programming package.

4. Case study
We will illustrate the developed concepts on a continuous stirred tank reactor (CSTR). Consider an isothermal,
perfectly mixed, nonlinear CSTR, governed by the van de
Vusse reactions
k1 B P
k2 B ,
B P
1
2
3
k3 B ,
B P
1
4

(27)

where the "rst and second reactions have "rst-order


kinetics, while the third reaction is second order. The
following dimensionless deviation state equations then
apply
F
c Q 1 "!k c 1 !k c21 # (c 1 !c 1 ),
B
1 B
3 B
B
< B ,f
F
c Q 2 "k c 1 !k c 2 ! c 2 .
1 B
2 B
B
< B

(28)

The input into the system is considered to be the inlet


#owrate F, while the output variable of interest is the
concentration of B , C 2 . Relevant physical parameters
2 B
are provided in Table 1, while the steady-state values are
provided in Table 2. We will assume the input to be

Table 1
Parameters used in CSTR simulation
k
1
k
2
k
3
c 1
B ,f
<

Kinetic parameter
Kinetic parameter
Kinetic parameter
Inlet concentration
Reactor volume

50 1/h
100 1/h
10 l/mol/h
10 mol/l
1l

2398

K.R. Harris et al. / Chemical Engineering Science 55 (2000) 2393}2400

Table 2
Nominal steady state of CSTR
F
c 1
B
c 2
B

Inlet #owrate
Steady-state concentration of B
1
Steady-state concentration of B
2

34.3 l/h
3.0 mol/l
1.1 mol/l

Fig. 2. System response to step inputs u(t)"H l/h.


Table 3
Nonlinearity measures for CSTR example

Fig. 1. (*) numerical, (- ) -) linear, (- - -) FEx1, and ( ) ) ) ) FEx2


steady-state loci.

constrained by $15 l/h, so that at nominal operation


we have
19.3 l/h)F)49.3 l/h.

(29)

Work focusing on the application of FEx models for the


control of a similar reaction system, as well as other
interesting systems has recently appeared in the literature
(Harris & Palazogy lu, 1998).
The fact that the system is not linear is demonstrated
by Figs. 1 and 2. First, Fig. 1 shows the steady-state
operating locus. Note that the gain of the system has
a critical point at approximately 79 l/h where the gain
changes sign. To the left of this point, the system displays
nonminimum phase behavior, while on the right of this
point the system is minimum phase. Indicated by the
circle is the nominal steady-state at F"34.3 l/h, while
the vertical bars denote the constrained region of operation. The FEx2 model captures the steady-state of the
system well, while the FEx1 has a slight error at the
bounds of the constraint. Fig. 2 demonstrates the dynamic behavior of the system in response to a set of step
inputs at the nominal operating point. Notice the asymmetrical response characteristic of nonlinear systems, as
well as the nonminimum phase behavior since the nominal steady state is located to the left of the critical point
on the system locus.
Figs. 1 and 2 clearly show that the system is nonlinear,
and in what follows the measures developed above will
be applied to determine how nonlinear the system is.

FEx1
FEx2
Numerical on U
c1
Numerical on U
c2

/U
N

/M U
N

[0.0024,0.0133]
[0.0024,0.0168]
0.0039
0.0035

[0.3258,0.6592]
[0.3067,0.4220]
0.3794
0.3461

First, we will compute the lower bounds of the nonnormalized and normalized nonlinear measures through
the use of the Fourier coe$cients. In this case, the input
set U is chosen, following Eq. (20), as
LB
A3A "MA3R` D 0.05)A)15N,
f
u3X "Mu3R` D 0.01)u)10N.
(30)
f
The lower bounds on the measures were computed using
Eqs. (22) and (23) with the Fourier coe$cients derived
from the FEx1 and FEx2 models. These results are provided in Table 3.
To compute the upper bounds on the measure using
Eqs. (25) and (26), we will choose the input set U
UB
to be
U "Mu D u(t)"hS(t),
UB
h3H"Mh3R D !15)h)15NN,

(31)

where S(t) is the unit step function. These results are also
given in Table 3. Note that the bounds are rather tight on
the measure for the more accurate FEx2 model, while the
bounds obtained from the FEx1 model are more conservative. This overestimation of the measure by the upper
bound is due to the error in the FEx1 model, which can
be observed, at least in the steady-state sense, in Fig. 1.
By examining the magnitude of the nonnormalized
measure only, one is led to believe that the system is
rather linear. However, since the gain of the system is

K.R. Harris et al. / Chemical Engineering Science 55 (2000) 2393}2400

2399

low, the normalized measure provides more insight into


the salient nonlinearities in the system under study.
The measure was also computed using the convex
optimization outlined by AllgoK wer (1995b). In this case,
two di!erent discrete sets of inputs were used to provide
insight into the dependence of the measure on the input
set. The "rst set U was given by
c1
U "Mu D u(t)"hS(t),
c1
h3H"Mh3!15,!13,2,13,15NN,
(32)
while for the second set, U , Eq. (20) was parameterized
c2
by
A3A "M1,2,2,14,15N,
f
u3X "M0.01,0.1,2,10N.
(33)
f
A "nite impulse response model was used as the parameterization of the best linear model, given by
n
GH[u]" + x u(k!i).
(34)
i
i/1
Twenty model coe$cients were used, with a process
sampling time of 0.01 h and a simulation horizon of 0.3 h.
The results of this optimization are also provided in
Table 3. In all cases, the numerical measure falls within
the bounds as derived from the FEx models. Additionally, the measure as computed from the two sets agree
within 12%, con"rming only a slight dependence on the
chosen input set.
Now, we will exploit the algebraic nature of the FEx
models to compute the measure bounds as a function of
the operating point in the system. Fig. 3 illustrates the
lower and upper bounds on the normalized measure,
derived from both the FEx1 and FEx2 models. The upper
bounds of the measure indicate that the input set

Fig. 4. Upper and lower bounds of normalized measure as a function of


the inlet #owrate F of B . ( ) ) ) ) FEx1, (- - -) FEx2.
1

U
can cause the system gain to change sign over
UB
a rather wide range of operation. Conversely, the lower
bound veri"es the #owrate at which the gain switches
sign at 79 l/h. Again the FEx1 upper bound over-estimates the nonlinearity in the system due to the inaccuracy of the FEx1 model. Also shown is the results of the
numerical optimization using the same parameters as
above, based on the set U .
c1
Finally, the dependence on the input sets U and
LB
U can be investigated. If the inlet #owrate is further
UB
constrained by $5 l/h,
29.3 l/h)F)39.3 l/h,

(35)

the measure bounds decrease as expected. Fig. 4 shows


these bounds, again as a function of the nominal #owrate.
Note that not only is the measure lower over the range of
operation, but the region at which the gain of the system
changes sign is reduced as well. This last example illustrates the importance of correctly selecting an input set to
span the range of desired operation for the system at
hand.

5. Summary

Fig. 3. Upper and lower bounds of normalized measure as a function of


the inlet #owrate F of B . ( ) ) ) ) FEx1, (- - -) FEx2.
1

Typically, all systems display some form of nonlinearity. A more important issue involves the e!ect of this
nonlinearity on the intended operation of the process. In
this work a systematic measure to quantify this e!ect was
reviewed in the context of a general closed-loop, controller design problem. Using this framework, the engineer
can assess the nonlinear contributions of the objective
function, system dynamics, and system constraints, and
their e!ect on the design of the appropriate control
law. The measures, as presented in this work, focus on

2400

K.R. Harris et al. / Chemical Engineering Science 55 (2000) 2393}2400

the nonlinear contributions of the system dynamics in


particular.
FEx models were used to derive upper and lower
bounds on the measure. Because these bounds are both
algebraic and continuous in nature, the method avoids
the numerical complexity and discrete parameterizations
that exists with other methods. Additionally, the algebraic nature of the metric allows operating steady-states
and system parameters to be e$ciently explored, determining the degree of nonlinearity present in various
operating regimes.
The concepts developed in this work were illustrated
on a benchmark nonlinear CSTR example.
Acknowledgements
The support of the National Science Foundation
(CTS-9400304, CTS-9800073) is gratefully acknowledged.
References
AllgoK wer, F. (1995a). De"nition and computation of a nonlinear
measure. IFAC NOLCOS, 1, 279}284.
AllgoK wer, F. (1995b). Dexnition and computation of a nonlinearity
measure and application to appoximate I/O-linearization. Technical
Report No. 95-1, UniversitaK t Stuttgart.
AllgoK wer, F. (1996). Na( herungsweise Ein-/Ausgangs-Linearisierung Nichtlinearer Systeme. DuK sseldorf: VDI Verlag.
BatiguK n, A., Harris, K. R., & Palazogy lu, A. (1997). Studies on the
dynamics of nonlinear processes via functional expansions: I. Solu-

tion to nonlinear ODEs. Chemical Engineering Science, 52(18),


3183}3195.
Boyd, S., & Chua, L. (1985). Fading memory and the problem of
approximating nonlinear operators with Volterra series. IEEE
Transactions on Circuits and Systems, CAS-32, 1150}1161.
Desoer, C., & Wang, Y. (1981). Foundations of feedback theory for
nonlinear dynamical systems. IEEE Transactions on Circuits and
Systems, CAS-27(2), 104}123.
Fliess, M., Lamnabhi, M., & Lamnabhi-Lagarrigue, F. (1983). An
algebraic approach to nonlinear functional expansions. IEEE Transactions on Ciruits and Systems, CAS-30, 554}567.
Guay, M., McLellan, P. J., & Bacon, D. W. (1995). Measurement of
nonlinearity in chemical process control systems: The steady state
map. Canadian Journal of Chemical Engineering, 73, 868}882.
Haber, R. (1985). Nonlinearity test for dynamic processes. In Proceedings of the IFAC identixcation and system parameter estimation symposium.
Harris, K. R. (1998). Control and analysis of nonlinear chemical processes
using functional expansions. Ph.D. thesis, University of California at
Davis.
Harris, K. R., & Palazogy lu, A. (1997). Studies on the dynamics of
nonlinear processes via functional expansions: II. Forced dynamic
systems. Chemical Engineering Science, 52(18), 3197}3207.
Harris, K. R., & Palazogy lu, A. (1998). Studies on the dynamics of
nonlinear processes via functional expansions: III. Controller design. Chemical Engineering Science, 53(23), 4005}4022.
Nikolaou, M. (1993). When is nonlinear dynamic modeling necessary?
In Proceedings of the ACC (pp. 910}914).
Prett, D. M., & Garcia, C. E. (1988). Fundamental process control.
Boston:Butterworth.
Stack, A. J., & Doyle, F. J. (1997). Application of a control-law nonlinearity measure to the chemical reactor analysis. A.I.Ch.E. Journal,
43(2), 425}439.
Vidyasagar, M. (1993). Nonlinear systems analysis. Englewood Cli!s, NJ:
Prentice-Hall.

You might also like