System identification as an application of optimization

Ana S. R. Brasio, Andrey Romanenko, and Natercia C. P. Fernandes

Citation: AIP Conf. Proc. 1479, 822 (2012); doi: 10.1063/1.4756264
View online: http://dx.doi.org/10.1063/1.4756264
View Table of Contents: http://proceedings.aip.org/dbt/dbt.jsp?KEY=APCPCS&Volume=1479&Issue=1
Published by the American Institute of Physics.

Additional information on AIP Conf. Proc.
Journal Homepage: http://proceedings.aip.org/
Journal Information: http://proceedings.aip.org/about/about_the_proceedings
Top downloads: http://proceedings.aip.org/dbt/most_downloaded.jsp?KEY=APCPCS
Information for Authors: http://proceedings.aip.org/authors/information_for_authors
Downloaded 11 Oct 2012 to 2.83.66.196. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions
System Identication as an Application of Optimization
Ana S. R. Brasio
,
, Andrey Romanenko

and Natercia C. P. Fernandes

1

CIEPQPF, Department of Chemical Engineering, Faculty of Sciences and Technology, University of Coimbra,
3030-790 Coimbra, Portugal

Ciengis, SA, 3030-199 Coimbra, Portugal

Abstract. The work concerns the system identication of industrial processes via the Sequential Quadratic Programming
algorithm. The proposed approach, testing scenarios, and the system identication results are discussed. The tool is tested with
two datasets, the rst one collected in loco from an industrial process and the second one generated with a plant simulator of a
continuous stirred tank reactor, a system widely used in industry. In both cases, the resulting models capture well the process
dynamics.
Keywords: optimization application, system identication, mathematical models
PACS: 07.05.Kf,07.05.Dz, 02.30.Hq,02.30.Yy,02.60.Ed,02.60.Pn
INTRODUCTION
System identication deals with the construction of mathematical models based on the experimental observation of the
system answer to some stimuli. Its use is widespread across all engineering elds, being applied to completely different
systems in various industries (from aerospace, to civil or health industries). It is also used for control purposes in the
manufacturing industry. The operation of the productive lines has to be kept under control for economical and/or
legal reasons. The Proportional-Integral-Derivative (PID) controller is unquestionably the most commonly used in
industry. In fact, more than 95% of the controllers used in industrial processes are PID algorithm or its advanced
versions [1, 2, 3]. The tuning of these controllers require a mathematical model of the process. To achieve a good
control of the process, it is therefore essential to have mathematical models that are able to predict the behaviour of
the units that constitute the production line.
Often it is preferable to simplify as much as possible the process model as long as it still captures the most important
steady-state and dynamic characteristics of the process [4]. The rst order (FO), rst order with time delay, and second
order models have been widely used to design and implement process controllers. In this work, we will exemplify our
methodology using FO models which adapt better to the systems under study.
The FO process is dened as
x

(t) +x(t) = K
p
u

(t), (1)
where x

and u

are the state variable and the input variable, respectively, both expressed as deviation variables. The
static gain, K
p
, represents how much the process output changes, from a steady-state to another, for a unitary variation
of the process input while the time constant, , represents how fast the process responds to a change in the process
input. A generic deviation variable z

relates to the original variable z through a simple translation of the initial steady-
state z, ie, z

= z z.
Dynamic models derived from physical principles typically consist of one or more ordinary differential equations.
State-space models provide a compact and useful representation of a set of ODE equations and can be generally written
as

X(t) = AX(t) +BU(t) +ED(t)

Y(t) =CX(t)
, (2)
where X(t) is the state vector, U(t) is the input vector of manipulated variables, D(t) is the disturbance vector and
Y(t) is the output vector of measured variables (typically a subset of X(t) [5]). Parameters A, B, C and E are constant
matrices.
1
Corresponding author. Email: natercia@eq.uc.pt
Numerical Analysis and Applied Mathematics ICNAAM 2012
AIP Conf. Proc. 1479, 822-825 (2012); doi: 10.1063/1.4756264
2012 American Institute of Physics 978-0-7354-1091-6/$30.00
822
Downloaded 11 Oct 2012 to 2.83.66.196. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions
In order to determine the parameters and to identify the system, the rst step consists on the so-called process
activation procedure that requires a special care. If the process is activated too aggressively, the products quality
may not be acceptable or the process safety may not be guaranteed. On the other hand, if the activation is not enough,
an accurate process model cannot be obtained because the information included in the activated data is limited and the
uncertainties (due for example to measurement noise and disturbances) may become dominant [6, 7]. In the present
work, the process activation is performed through the use of stepwise input signals in open loop scenarios.
PROBLEM FORMULATION
Once process data with sufcient information content is collected, the model parameters are determined such that the
model response reproduces the observed response of the actual process to the set of identication disturbances.
The most frequently used curve tting criterion is the least squares criterion which minimizes the standard deviation
of the model from the dataset. Another criterion is to minimize the sum of the absolute deviation. However, this
becomes not continuous complicating the resolution of the optimization problem. The Chebyshev approximation
criterion minimizes the largest absolute deviation over the entire set. But this criterion is often difcult to apply in
practice since the resulting optimization problem may require advanced mathematical procedures [8].
The nonlinear constrained optimization problem is dened as
minimize
P
J(P) (3a)
subject to

X = f (X,U, P) (3b)
X
L
X X
U
, U
L
U U
U
(3c)
P
L
P P
U
(3d)
g(P) 0, (3e)
where J denotes the objective function, P is the model parameters vector to be estimated, X and U are the vector of
controlled and manipulated variables (respectively), and the subscripts
L
and
U
stand for lower and upper (respectively).
The set of equations (3b) to which the optimization problem is subject are the process model.
Inequalities (3e) may enforce additional identication criteria.
Given a model x = f () and a set of m data points (
i
, x
exp,i
), the objective function J is written, according to the
minimum least squares criterion, as
J =
m

i=1

x
exp,i
f (
i
)

2
. (4)
It is noteworthy that depending on the model and on the constraints set, the optimization problem may become
non-convex. The problem was solved using Sequential Quadratic Programming (SQP), which is a local minimum
algorithm.
RESULTS AND DISCUSSION
In process systems engineering, one may encounter processes in which each output variable depends essentially on
a single input variable. These may be modelled as the so called single-input single-output (SISO) systems. However,
a large class of processes exhibit interaction among variables, i.e., the output variables are dependent on a subset of
input variables. These latter processes are regarded as multiple-input multiple-output (MIMO) systems. The common
industrial practice is to assume that there is no interaction or to design controllers in a way that the interaction is
weakened. However, such approaches may result in suboptimal plant performance. Therefore, multivariable controller
tuning, and thus, the multivariable system identication which is the subject of this work, have a big practical
importance.
Because of the existence of both SISO and MIMO systems in industrial practice, the performance of the tool is
tested in both congurations and the results are presented below.
823
Downloaded 11 Oct 2012 to 2.83.66.196. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions
0.8
0.9
1.0
1.1
1.2
0 12 50 69 86 100
0.90
0.95
1.00
1.05
1.10
x
u
t, min
xexp
x
init
x
t
u
(a) SISO system.
0
5
10
15
20
0 90 180 270 520 610 770 860 1000
10
20
30
40
50
x
u
t, min
C
Ai
Tc
T
CA
xexp
x
t
u
(b) MIMO system.
FIGURE 1. System identication.
TABLE 1. Identication results for the SISO system.
P Initial LB UB Fit Indicators
K
p
0.100 0.0001 5 2.154 J 0.2856
100.000 1 1000 88.868 R
2
0.9727
x 0.800 -10 10 1.004
u 0.100 -10 10 1.005 SNR 11.0
SISO systems identication
A real industrial process that can be regarded as a SISO system was stimulated with a sequence of input steps and
the proles of the input and the output variables were registered. The obtained dataset contains 1200 points covering an
interval of 100 minutes with a sampling period of 5 seconds. For condentiality reasons the data was later normalized.
Figure 1a presents both the stimuli, u, and the system answer, x
exp
, obtained during the process activation stage. The
success of system identication strongly depends on the quality of the data and, therefore, also depends on its signal
to noise ratio (SNR). The collected industrial dataset is characterized by an SNR of 11.0.
The optimization procedure described above applied over these 1200 data points allowed to identify the system.
The set of optimization related conditions and the suggested model parameters as well as some quantitative indicators
to assess the identication quality are presented in Table 1. The dynamic response of the resulting model is drawn in
Fig. 1a (dashed line) for easy comparison with the real system response (thin solid line).
It is noteworthy that the SNR of the data is relatively signicant and that the initial guess for the parameters is poor
(as it is shown by the dotted line representing the rst iteration model prediction). Although these two factors make
the identication process more difcult the resulting model is able to capture well the process dynamics, as proven by
the high correlation factor, R
2
.
MIMO systems identication
A continuous stirred tank reactor (CSTR) equipped with a heating coil is a good example of a MIMO system
commonly used in industry. This system has two input variables (the inlet ow concentration of reactant A, C
A,i
,
and the temperature of the heating uid in the coil, T
c
) and two output variables (the concentration of reactant in
the reactor, C
A
, and the temperature in the reactor, T). In order to collect data for the identication of the system,
a plant simulation was carried out according to Fig. 1b. The interaction among the variables is clear: for instance, a
disturbance in input variable T
c
results in a dynamic response not only of T but also of the second output variable, C
A
.
Similarly, by activating the input variable C
A,i
both output variables are direct or indirectly affected.
824
Downloaded 11 Oct 2012 to 2.83.66.196. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions
TABLE 2. Identication results for the MIMO system.
P Initial LB UB Fit Indicators
A

10
2
10
3
10
3
10
2

10
2
10
2
10
2
10
2

10
5
10
5
10
5
10
5

2.810
3
1.010
3
1.110
3
1.910
3

J 99.215
B

10
3
0
0 10
3

10
5
10
5
10
5
10
5

10
2
10
2
10
2
10
2

1.7110
3
0
0 3.410
4

R
2
0.99713

X

15 5

0 0

45 45

13.247 5.353

U

20 20

1 1

100 100

14.714 28.416

SNR

8.6 3.8

The data collection procedure was carried out for 1000 minutes with a sampling interval of 1 min. This data set
presents SNR of 8.6 and 3.8 for T
exp
and C
A exp
, respectively.
For the MIMO system, simultaneous accounting of both output curves is needed during the optimization process
(ie, the objective function is the sum of 21000 square errors between original and predicted values). The parameter
values determined by the optimization technique for this system are displayed in Table 2. The correlation factor
achieved was even higher than in the case of the SISO system, revealing an excellent t quality. It is worth mentioning
that this was achieved in spite of the high level of noise specially in the second variable (SNR = 3.8).
CONCLUSIONS
A system identication tool for SISO and MIMO models was developed using the Sequential Quadratic Programming
algorithm. The performance of the approach was tested using a dataset from a real industrial process and a simulated
process unit. The resulting models reproduce well the experimental data and, thus, may be used for process simulation
and/or control system design.
ACKNOWLEDGMENTS
The authors gratefully acknowledge the funding received from Fundacao para a Ciencia e a Tecnologia through
CIEPQPF Research Center.
REFERENCES
1. S. Yamamoto, and I. Hashimoto, Present Status and Future Needs: The View from Japanese Industry, in Chemical Process
Control- CPCIV, Proc. 4th Inter. Conf. on Chemical Process Control, 1991, vol. TX.
2. K. strm, and T. Hgglund, Journal of Process Control 14, 635 650 (2004), ISSN 0959-1524, URL http:
//www.sciencedirect.com/science/article/pii/S0959152404000034.
3. L. Eriksson, and H. Koivo, Tuning of discrete-time PID controllers in sensor network based control systems, in Computational
Intelligence in Robotics and Automation, 2005. CIRA 2005. Proceedings. 2005 IEEE International Symposium on, 2005, pp.
359 364.
4. T. Liu, and F. Gao, Industrial Process Identication and Control Design: Step-Test and Relay-Experiment-Based Methods,
Advances in Industrial Control, Springer, 2011, ISBN 9780857299765, URL http://books.google.pt/books?id=
\_5KzBxUnIdwC.
5. D. Seborg, T. Edgar, D. Mellichamp, and I. Francis J. Doyle, Process Dynamics and Control, John Wiley & Sons, 2010, ISBN
9780470128671, URL http://books.google.pt/books?id=\_PQ42kOvtfwC.
6. S. Sung, J. Lee, and I. Lee, Process Identication and PID Control, John Wiley & Sons, 2009, ISBN 9780470824115, URL
http://books.google.pt/books?id=xJsdOXB\_szgC.
7. L. Ljung, System identication: theory for the user, Prentice-Hall information and system sciences series, Prentice-Hall, 1999,
ISBN 9780138816407, URL http://books.google.pt/books?id=gpVRAAAAMAAJ.
8. D. D. Leon, Model tting, Tech. rep., California State University, Fresno (2012), URL zimmer.csufresno.edu/
~doreendl/232.12s/handouts/modelfitting.pdf, consulted in June 2012.
825
Downloaded 11 Oct 2012 to 2.83.66.196. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions