You are on page 1of 5

Linear Approach to Performance assessment of

a 4-tank system
Yousef alipouri, Javad poshtan
Faculty of Electrical Engineering
Iran University of Science and Technology
Tehran, Iran
E-mail: yalipouri@iust.ac.ir, jposhtan@iust.ac.ir
Majid Poshtan
EECE Department
American University in Dubai
Dubai, UAE
E-mail
:
mposhtan@aud.edu
AbstractMonitoring performance of nonlinear system is
an important task for many real world applications
especially for industrial environments; however its
realization is usually difficult. Many monitoring
performance methods have been introduced but they all
rely on loops with accurate linear models of the system.
The methods that are capable to identify MIMO nonlinear
systems are scarce and linear models are not so accurate in
modeling nonlinear systems. In this paper, a combined
series of two capable algorithms is used for identifying the
system. The method also identifies the existence of any
possible disturbance simultaneously in order to reduce the
complexity of the model. The logic behind the proposed
methods is to utilize and approximate the interaction
between the loops in the system. Moreover, the model
explicitly defines the relationship between the outputs and
the inputs of the system such that the performance indexes
can be easily calculated. A well designed control strategy is
introduced to determine the optimal values that are
required to calculate the performance index. A reasonable
decision on optimal situation is necessary to evaluate the
performance index of a system. A minimum variances
control strategy (MVC) is considered here as the most
suitable feedback controller because it achieves the
smallest possible closed-loop output variance. The above
method has been utilized for a 4-tank system and then a
minimum variance controller is designed for determining
the optimal output such that the minimum possible
variance, and consequently the performance index have
been calculated. The achieved Index can be then used for
monitor performance of systems in online and offline
applications.
Keywords- minimum variance controller, minimum
variance index, performance monitoring, Four-tank
benchmark system, maximum likelihood and CMA-ES
algorithms
I. INTRODUCTION
Control engineering deals with the theory, design and
application of control systems. The primary objective of
control systems is to maximize profits by transforming raw
materials into products while satisfying criteria such as
product-quality specifications, operational constraints, safety
and environmental regulations [1]. The design, tuning and
implementation of control strategies and controllers are
undertaken within the main phase in the solution of control
problems.
The most widespread (stochastic) criterion considered for
performance assessment in the process control is variance (or,
equivalently, the standard deviation), particularly for regulatory
control. The widespread use of variance as a performance
criterion is due to the fact that it typically represents the
product-quality consistency. The reduction of variances of
many quality variables not only implies improved product
quality but also makes it possible to operate near the
constraints for increasing throughput, reducing energy
consumption and saving raw materials [2].
The main question in dealing with output variance is how
much the variance can be decreased and how this aim can be
performed. The minimum possible variance is used for
evaluating an index, namely minimum variance index, which
decides on the efficiency of loop performance. Minimum
variance index is one of the common forms of control
performance index (CPI) used in industrial environments.
Bialkowski [3] declared that almost 60 percent of control
performance index used in industrial applications is minimum
variance index. The control performance index is a single
scalar which is usually scaled to lie within [0, 1], where values
close to 0 indicate poor performance and values close to 1
mean better/tighter control [4]. This indeed holds when perfect
control is considered as benchmark. Perfect control determines
the optimal value required for calculating control performance
index. The minimum variances control (MVC) (also referred to
as optimal H2 control and first derived in [5]) is the best
possible feedback control for linear systems in the sense that it
achieves the smallest possible closed-loop output variance.
More specifically, the MVC task is formulated as minimization
of the variance of the error between the set point and the actual
output. The output variance is defined as follows:
( )
_ _
= =
=

=
N
k
N
k
y
k y
N
y y k y
N
1 1
2
2
) (
1
, ) (
1
1
o
MVC can be regarded as optimal solution to both questions
[6]: how much and how output variance can be reduced.
Since 1998, a new approach (Interactor Matrix Method)
was introduced for estimating minimum variance with high
(1)
___________________________________
978-1-4673-1332-2/12/$31.00 2012 IEEE

accuracy [8,9]. It has been shown theoretically that this method


is accurate in estimating the minimum variance for MIMO
linear systems [4]. However, this method has some deficiencies,
1) it needs an exact linear model of system and disturbance, 2)
complex calculation is needed for finding Interactor Matrix and
some methods have been introduced for calculating interactor
matrix [10-12], and 3) it is defined for linear systems. Since
2000, the trend has been going toward decrease of complexity
in calculating Interactor Matrix [13-16]. The used method
needed no information on Intarctor Matrix. In this paper, MVC
controller is used for determining the optimal output variance
(without needing to calculate Intractor Matrix).
Four-tank benchmark system is benchmark system that is
prototype of many real world industrial systems. In this model
waters in tanks must be stabilized in predefined level without
any perturbation even after applying stochastic disturbances.
This system is nonlinear model with high reaction between
loops which can be switch between minimum phase and non-
minimum phase status.
For determining minimum variance controller there is need
to accurate MIMO linear model. Separately analyzing each
loop of MIMO system is not accurate because of the
interactions among loops, therefore for analyzing an MIMO all
loops must be regarded simultaneously. Existence of noise on
data makes this task hard. A lot of research has gone into the
development of ways of analyzing multivariate system data in
both the statistical and artificial intelligence communities.
Statistical MTS modeling methods include the Vector Auto-
Regressive process, the Vector Auto-Regressive Moving
Average process, and other non-linear and Bayesian
approaches, while various AI methods have been developed for
different purposes. In this paper, a well-known method CMA-
ES (Covariance Matrix Adaptation Evolutionary Programming)
[17] will be used for determining parameters of model. By
using this model, the minimum variance controller can be
designed for reaching to optimal variance.
The organization of this paper is as follows: Section II
presents physical equations for four-tank benchmark. Section
III studies combination of CMA-ES with Maximum likelihood.
In Section IV, designing of Minimum Variance Controller
(MVC) is explained. Section V presents the empirical results
on the linear system and nonlinear four-tank benchmark system.
Last section summarize benefits and shortcomings of introduce
method.
II. FOUR-TANK BENCHMARK SYSTEM
The four-tank system laboratory model can be viewed as a
prototype of many industrial applications in process industry
such as chemical and petrochemical plants, oil and gas systems.
The typical control issue involved in the system is how to keep
the desired liquid level in each tank. The principle scheme of
the model is shown in Fig. 1. The basic apparatus consists of
four plexiglass tanks with the cross sectional area of A. These
are connected to each other by cylindrical pipes with the cross
sectional area of a
i
. Each tank is equipped with a static pressure
sensor, which gives a voltage output proportional to the level of
liquid in the tank [18,19].
max
h
denotes the highest possible
liquid level in tanks. In case the liquid levels of 1 and 2 exceed
this value, the corresponding pump will be switched off
automatically.
1
v and
2
v are the voltages applied to pumps 1
and 2. Two variable speed pumps driven by a DC motor are
used in this apparatus. These pumps are designed to give an
accurate flow per rotation. Thus, the flow rate provided by each
pump is proportional to the voltage applied to its DC motor.
Figure 1 : The principle scheme of the model for a four-tank benchmark
process [18]
The physical equation for the four-tank system can be seen
in Eq. (2) [18].
( )
( )
1
4
1 1
4
4
4 4
2
3
2 2
3
3
3 3
2
2
2 2
4
2
4
2
2
2 2
1
1
1 1
3
1
3
1
1
1 1
1
2
1
2
2 2
2 2
u

A
k
gh
A
a
dt
dh
A
k
gh
A
a
dt
dh
A
k
gh
A
a
gh
A
a
dt
dh
A
k
gh
A
a
gh
A
a
dt
dh

+ =

+ =
+ + =
+ + =
where A
i
is the cross sectional area of tank i, a
i
is the cross
sectional area of valve i, h
i
is the height of liquid inside the
tank I and v
i
is voltage applied to pomp i. Constant coefficient
of the system is in the following way:
2
2
2 2 2
981 , 5 . 0 , 057 . 0 ,
, 071 . 0 , , 32 , ; 28 ,
4 2
3 1 4 2 3 1
S
cm
cm
V
c
cm
cm cm cm
g k a a
a a A A A A
= = =
= = =
The initial value of the four-tank system is as follows:
6 . 0 , 7 . 0
35 . 3 , 33 . 3
3 , 3
4 . 1 , 8 . 1 , 7 . 12 , 4 . 12
2 1
2 1
0
2
0
1
0
4
0
3
0
2
0
1
3 3
= =
= =
= =
= = = =

Vs
cm
Vs
cm
V V
cm cm cm cm
k k
v v
h h h h
Disturbance on system affects water level in tanks 3 and 4,
which is modeled by Eq. (3) [19].
(2)

( )
( )
2
4
2
1
4
1 1
4
4
4 4
1
3
1
2
3
2 2
3
3
3 3
1
2
1
2
d
A
k
A
k
gh
A
a
dt
dh
d
A
k
A
k
gh
A
a
dt
dh
d
d

+ =

+ =
u

where k
d1
=1,k
d1
=2,d
1
and d
2
are disturbance inputs.
The manipulated valuables are input voltages of pumps.
Outputs of the model are the height of liquid in tanks 1 and 2.
2 2 1 1
, h y h y = =
The outputs-inputs model is as follows:
2
2
2 2
4
2
4
2
2
2 2
1
1
1 1
3
1
3
1
1
1 1
2 2
2 2
u

A
k
gh
A
a
gy
A
a
dt
dy
A
k
gh
A
a
gy
A
a
dt
dy
+ + =
+ + =
III. IDENTIFYING BY USING CMA-ES ALGORITHM
Equation (6) is a linear model is using for identifying SISO
systems.
( ) ( ) ( ) t e m t u b t u b n t y a t y a t y
m n
+ + + = + + + ) ( 1 ) ( ) 1 (
1 1
" "
where e(t) is white noise which is considered modeling
mismatch.
This model has unknown parameters (7) which must be
estimated.
| |
T
m n
b b a a a " "
1 2 1
= u
By replacing the unknown parameters a
i
in Eq. 2 with A
i
matrixes and replacing the scalar y(t) with vector Y(t), (6) turns
to (8), which is a MIMO model which can model interaction
among loops.
( )
( ) ( ) ( )
( )
( )
( )
( )
( ) ( )
( ) ( )
, ,
, , ,
... 1 ) (
) 1 ( ) ( ) ( ) (
1
1 11 1
1
1 11 1
1
1 11
1
1
0 1
(
(
(

=
(
(
(

=
(
(
(

=
(
(
(

=
(
(
(

=
+ + + + +
+ = + + + =
_ _
= =
i
nm
i
n
i
m
i
i
n
nn n
n
n
i
nn
i
n
i
n
i
i
d p
d
i
i
p
i
i
b b
b b
B
a
a
a
t e t e
t e t e
t
t y
t y
t Y
a a
a a
A
t a d t U B t U B p t Y A
t Y A t a i t U B i t Y A t Y
"
# % #
"
#
"
# % #
"
#
"
# % #
"
"
c
c
c
where p is model order, n is number of outputs, m is number of
inputs, a is average value of disturbance, Y(t) is vector of
output data and
( ) t c
is model mismatch at sample time t.
j i
B a A , ,
are unknown constant parameters, which must be
estimated by a proper algorithm.
For reaching higher accuracy, higher order (p) and capable
parameter estimation methods are needed. Consequently,
higher order of model leads to higher number of unknown
parameters. Estimating the value for high number of
parameters (high dimensional problems) is hard or even
impossible for some methods. Most algorithms cannot be used
in such problems, as they are usually trapped in local
minimums. AI methods practically are used for finding best
parameters for this kind of models [20-22].
In this paper, CMA-ES which is state of art algorithm in
evolutionary algorithms, introduced in [17], will be used. For
implementing this method, at first Maximum likelihood (ML)
algorithm begun to search possible values for model
parameters, but it is so likely that this algorithm trapped in
local minima, and then CMA-ES algorithm is running and
searching for optimal values for parameters. Fig. 2 shows
approach of search.
Fig. 2. Shows approach of using CMA-ES and ML for estimating optimal
model
Fig. 3 shows an example of raw data, primary model and
optimized model.
0 200 400 600 800 1000 1200
-1
-0.5
0
0.5
1
1.5
2
samples
p
system output
0 200 400 600 800 1000 1200
-1
-0.5
0
0.5
1
1.5
2
samples
o
u
tp
u
ts
model output
system output
0 200 400 600 800 1000 1200
-1
-0.5
0
0.5
1
1.5
2
Raw data Primiray Model output Optimized model output
Fig. 3 shows example of reaching to optimal model by approach
introduced in Fig 2.
Interested readers can be referred to [20-22] for extensive
explanation about procedure of model optimization by EAs.
IV. DESIGNING MINIMUM VARIANCE CONTROLLER
In Minimum Variance Controller (MVC), the main objective is
to decrease the control signal variance.
( ) Y Y E J
T
=
By substituting (8) in (9) we have:
{ }
( ) ( ) | | { }
2
1 1
) (
2
) ( ) (
) ( ) 1 ( ) 1 ( min
) ( min ) ( min
t z t U q B t Y q E
t Y E t J
t u
t u t u
c | + + + =
=

Where
p i A
i i
s s = 1 |
( ) ( )
( ) ( ) ( ) ( ) z t t U q B t Y q t Y
t t U q B z p t Y t Y t Y
p
+ + + =
+ + + + + =

c |
c | |
) 1 ( 1 ) (
) ( ) ( ) 1 ( ) (
1 1
1
1
"
( )
1
1
1
1
1
1
) (
+
+
+ + =
+ + =
d
d
p
p
q B B q B
q q
"
" | | |
Eq. (10) is minimized by Eq. (11):
( ) ( ) | | { }
2
) (
1
1
1
) ( ) ( min ) ( ) ( t E t J z t Y q q B t U
t u
c | = + =

Control signal in Eq. (38) minimizes the objective function


defined in (36)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)

V. SIMULATION RESULTS
1) Empirical Tests-A Nonlinear Case
The most important part of designing MVC controller for a
nonlinear system is to estimate an accurate model, which was
the subject of Section III. The sampling time for the four-tank
system is 1 second and time constant is about 30 minutes [18].
The open loop response of system and estimated model
response are shown in Fig. 4.
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
0
2
4
6
8
10
12
Samples
System Output1
Model Output 1
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Samples
System Output 2
Model output 2
Fig. 4 : The output of system Eq. 2 (blue curve) and identified model (red
curve)
The accumulated residuals for 10000 samples is
10
10 6 . 1

and the residual for each sample is in the order of


14
10

,
indicating the ability of model in identifying the property of the
nonlinear four-tank system. The residuals between the sampled
data of model and system are shown in Fig. 5. Table 2 shows
the parameter of estimated model for the simulation data of the
four-tank system.
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
-8
-6
-4
-2
0
2
4
6
x 10
-13
output1
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
-16
-14
-12
-10
-8
-6
-4
-2
0
2
4
x 10
-14
output2
Fig. 5 : Residuals between the output of system (2) and optimal model
(Table 1)
Table 1 : Estimated parameters of optimal model for system (2)
value value value
(



6 2 8 3
8 3 6 2
E E
E E
Q
(

29 . 1
0
0
065 . 1
B (

32 . 10
52 . 8
a
(


0289 . 0 00801 . 0
0077 . 0 0039 . 0
A
3
(


3649 . 0 0055 . 0
0084 . 0 433 . 0
A
2
(

700 . 0 005 . 0
004 . 0 907 . 0
A
1
(


00541 . 0 0046 . 0
0099 . 0 0020 . 0
A
5
(

0054 . 0 0132 . 0
02164 . 0 0051 . 0
A
4
In Table 1, Q is covariance matrix of residuals and other
parameters were introduced in Eq. (8).
Table 2 lists simulation results after applying the designed
MVC to the nonlinear four-tank system.
Table 2: The output and control signal variances for the nonlinear four-
tank benchmark system
Var(y
1
)+
Var(y
2
)
Var(y
2
) Var(y
1
)
Var(u
1
)+
Var (u
2
)
Var(u
2
) Var(u
1
)
1.4943 0.7541 0.7422 1.4092 0.6841 0.7251
Fig. 6 shows the disturbance samples which is applied on
system.
0 5000 10000 15000
-20
-15
-10
-5
0
5
10
15
20
25
samples
D
is
tu
rb
a
n
c
e
C
h
a
n
e
l 2
0 5000 10000 15000
-20
-15
-10
-5
0
5
10
15
20
sample
d
is
tu
rb
a
n
c
e
c
h
a
n
e
l 1
Fig. 6: shows samples of disturbance
Fig. 7 shows the output of system by feedback (without
controller). The system is going to be unstable (from samples
10000) after increasing disturbance variance.
0 2000 4000 6000 8000 10000 12000 14000 16000
0
2
4
6
8
10
12
Samples
D
a
ta
Output1
0 2000 4000 6000 8000 10000 12000 14000 16000
0
1
2
3
4
5
6
7
8
9
10
Samples
D
a
ta
Output2
Fig. 7: shows output of feedback controller

Fig. 8 shows output of system after applying minimum
variance controller. It can be seen that system is stable and the
variance is decreased so considerably.

0 2000 4000 6000 8000 10000 12000 14000 16000
0
2
4
6
8
10
12
14
Samples
D
a
ta
Output1
0 2000 4000 6000 8000 10000 12000 14000 16000
0
2
4
6
8
10
12
Samples
D
a
ta
Output 2
Fig. 7: shows output of MV controller
2) Performance index
After designing optimal controller in (11), the minimum
variance index can be determined by (12).
act
opt
MV
Index
2
2
o
o
=
(12)

where
opt
2
o is optimal variance which is output of minimum
variance controller and
act
2
o is actual variance of system output
(without minimum variance controller). For example,
performance index for feedback controller in Fig. can be
calculated as follow:
From Fig. 7 (variance calculated from samples 1 - 10000)
23 . 6
2
= act o and from Table 2 4943 . 1
2
= opt o , Then:
23 . 0
23 . 6
4943 . 1
= =
MV
Index
In the other word, the feedback controller is working in
20% of optimal situation.
VI. CONCLUSION
Designing minimum variance controller (MVC) is an
option to reach the optimum variance in output which causes to
determine the minimum variance index. It needs an exact linear
model of system and disturbance for determining the minimum
possible variance. Considering that there are no accurate
nonlinear or linear models for most real-world systems,
designing minimum variance controller is very challenging. In
this paper, a new method was proposed for modeling nonlinear
four tank system by using heuristic algorithm CMA-ES. After
modeling the system minimum variance controller has been
designed and applied on nonlinear system. It is shown that the
controller can stabilize the system by decreasing the variance in
output (feedback couldnt did it without MVC controller).
Variance of output while controlled by minimum variance
controller can be regarded as optimal variance, so actual output
variance can be compared with optimal variance to determine
the performance of system. It is shown that the feedback
control has just 20% accuracy of MVC in reaching to minimum
variance. This statement declares that performance of feedback
controller can be enhanced 80% by using suitable controller
(such as MVC).
REFERENCES
[1] Seborg DE, Edgar TF, Mellichamp DA (2004) Process Dynamics and
Control. John Wiley & Sons.
[2] Shinskey FG (1996) Process-Control Systems: Application, Design, and
Tuning. McGraw Hill.
[3] W.L. Bialkowski, Dreams vs. reality: a view from both sides of the
gap, Pulp & Paper Canada 94:1927, 1993.
[4] M, Jelali, (2010), Control System Performance Monitoring Assessment,
Diagnosis and Improvement of Control Loop Performance in Industrial
Automation, Springer.
[5] K J Astrom, (1970), Introduction to Stochastic Control Theory. New
York: Academic Press.
[6] Martensson J., Rojas C. R. and Hjalmarsson H., 2011, Conditions when
minimum variance control is the optimal experiment for identifying a
minimum variance controller, Automatica 47, 578583.
[7] W. DeVries, S Wu, Evaluation of process control effectiveness and
diagnosis of variation in paper basis weight via multivariate time-series
analysis, IEEE Trans Automat Control 23:702708, 1978.
[8] B. Huang, S.L. Shah, Practical issues in multivariable feedback control
performance assessment, Proc IFAC ADCHEM, Banff, Canada, pp
429434, 1997.
[9] B. Huang, S.L. Shah, Performance Assessment of Control Loops,
Springer, 1999.
[10] Rogozinski M., Paplinski A. and Gibbard M., 1987 An algorithm for
calculation of nilpotent interactor matrix for linear multivariable
systems, IEEE Trans Automat Control 32:234237.
[11] Huang B., Shah S., Badmus L. and Vishnubhotla A., 1999, Control
Performance Assessment: An Enterprise Asset Management Solution,
www.matrikon.com/download/ products /lit /processdoctor _pa_eam.
pdf.
[12] Huang B., 1997, Multivariate Statistical Methods for Control Loop
Performance Assessment, PHD thesis, University of Alberta, Canada.
[13] Huang B., Ding S. X. and Thornhill N., 2005, Practical solutions to
multivariate feedback control performance assessment problem: reduced
a priori knowledge of interactor matrices, Journal of Process Control 15,
573583.
[14] Kadali R. and Huang B., 2007, Multivariate controller performance
assessment without interactor matrixa subspace approach, IEEE
Trans on Control Systems Technology, Vol. 15, No. 1.
[15] Shah S. L., Mohtadi C. and Clarke D., 1987, Multivariable adaptive
control without a priori knowledge of the delay matrix, Systems &
Control Letters 9:295306.
[16] Xia H., Majecki P., Ordys A. and Grimble M., 2004, Performance
assessment of MIMO systems under partial information, Proceeding of
the 2004 American Control Conference Boston, Massachusetts.
[17] 10. Hansen, Nikolaus, The CMA Evolution Strategy: A
Tutorial.(www.bionik.tu-berlin.de/user/niko/cmatutorial.pdf), November
24, 2010.
[18] Karl Henrik Johansson (2000), The Quadruple-Tank Process: A
Multivariable Laboratory Process with an Adjustable Zero, IEEE Trans.
on Control Systems Technology, Vol. 8, No. 3, May.
[19] E. P. Gatzke, E. S. Meadows, C. Wang, F. J. Doyle, (2000) Model
based control of a four-tank system, Computers & Chemical
Engineering, Volume 24, Issues 2-7, 15 July, Pages 1503-1509.
[20] I. Rojas, F. Rojas, H. Pomares, L.J. Herrera, J. Gonzalez, O.Valenzuela,
The synergy between classical and soft-computing techniques for time
series prediction, in: Advances in Artificial Intelligence, Lecture Notes
in Computer Science, Vol. 2972, 2004, pp. 3039
[21] C.-S. Ong, J.-J. Huang, G.-H. Tzeng, Model identification of ARIMA
family using genetic algorithms, Appl. Math. Comput. 164 (3) (2005)
885912.
[22] Valds, Julio; Barton, Alan, Multivariate Time Series Model Discovery
with Similarity Based Neuro-Fuzzy Networks and Genetic Algorithms
April 2003, published in Proceedings of the IEEE, INNS, IJCNN 2003
International Joint Conference on Neural Networks. IEEE Catalog
Number: 03CH37464C,ISBN: 0-7803-7899-7. Oregon, Portand, USA.
(13)

You might also like