See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/228494651

An optimization framework to computer-aided

design of reliable measurement systems

Article in Computer Aided Chemical Engineering · December 2006

DOI: 10.1016/S1570-7946(06)80225-7

CITATIONS READS

0 22

4 authors, including:

Carlos A Méndez Luis Puigjaner

Universidad Nacional del Litoral Universitat Politècnica de Catalunya
120 PUBLICATIONS 1,983 CITATIONS 522 PUBLICATIONS 6,600 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Energy Systems Engineering View project

Optimal integration of production, inventorying and distribution of chemical fluids. View project

All content following this page was uploaded by Luis Puigjaner on 05 September 2017.

The user has requested enhancement of the downloaded file.

An optimization framework to computer-aided design and upgrade of measurement systems
A.B. Editor et al. (Editors)
© 2005 Elsevier B.V./Ltd. All rights reserved. 1

An optimization framework to computer-aided

design and upgrade of measurement systems
Raffaele Angelini, Carlos A. Méndez, Estanislao Musulin, Luis Puigjaner*
Chemical Engineering Department-CEPIMA, Universitat Politècnica de Catalunya
ETSEIB, Av.Diagonal 647, E-08028, Barcelona, Spain

Abstract
The optimal design and upgrade of sensor networks have received increasing attention
over the last few years. This work presents a new methodology based on a MINLP
formulation to the optimal design of reliable measurement systems with minimum cost.
The applicability of the MINLP model is illustrated through the resolution of a
benchmark case study taken from literature and its computational performance is
compared with a GA-based approach.

Keywords: Sensor network, design and retrofit, MINLP model.

1. Introduction

A wide-ranging variety of approaches relying on exhaustive enumeration, algorithmic

procedures, rigorous mathematical models and meta-heuristic techniques have been
developed to address the complexity of the sensor network placement problem, usually
assuming a steady-state system at the nominal operating conditions. Within this context,
the sensor placement can be regarded as a highly combinatorial optimization problem
where the main goal is to find the optimal balance between the performance indicators
and the cost of the data acquisition system. The performance of the whole measurement
system as well as of individual process variables is typically determined through
multiple criterions such as precision, flexibility, reliability, robustness and so on, to
which minimum requirements are usually to be satisfied. The concept of observability is
also a necessary condition of any sensor network since that the state of the entire
process must be known, which means that a numerical value needs to be determined for
every process variable by using direct measurement or deduction from other process
variables. On the other hand, the total cost of the measurement system can comprise the
investment cost of new sensors, the installation and maintenance charges and, in the
case of retrofitting, the sensor reallocation cost. The design of a sensor network that
allows for the observation of all process variables was first addressed by Vaclaveck and
Loucka (1976). Later, this problem was solved by Madron and Veverka (1992)
regarding the minimum total cost. However, sensor failure can lead to a reduction of
measurements, thereby seriously affecting the whole process performance. Thus, it is
necessary to assure process observability even after sensor failures. The observation of a
process variable can be expressed mathematically by making non-null the probability of
estimating this variable at a given time t (reliability). The evaluation of this probability
is closely related to the different ways of estimating a process variable given a sensor
failure probability and a specific sensor network. On the basis of these concepts, a
method for optimal sensor location in a pure flow process was developed using graph-

*
Author to whom correspondence should be addressed: [email protected].
2 Angelini et al.

theory by Ali and Narasimhan (1993). Since this preliminary approach did not directly
consider neither the network cost nor the reliability of the individual process variables,
these key topics were later addressed by Bagajewicz and Sánchez (2000). Additionally,
these authors transformed the problem presented by Ali and Narasimhan into a
mathematical programming problem. Later on, Benqlilou et al. (2004) proposed an
approach for evaluating the reliability of process variable estimation taking into account
all the redundancies offered by the system in terms of either functional or hardware. In
this evaluation, both quantitative process knowledge and fault tree analysis are
considered and combined, which leads to a more suitable and practical evaluation of
reliability. The reliability of estimating each one of the key process variables is then
used to determine sensor network reliability, which, in turn, is used as a set of sensor
placement constraints in the design and retrofitting procedure. Thus, a general sensor
placement formulation is first proposed that considers the number (hardware
redundancy or multiplicity) of sensors of a given type (reliability) that are to be assigned
to a given process variable while satisfying the reliability requirements at the minimum
total cost. This proposal was applied for network design as well as for retrofitting.
Although the formulated sensor placement optimization problem was successfully
solved using genetic algorithms, it showed some typical limitations of stochastic
optimization methods, such as the impossibility to assure a global optimum and the high
difficulty for dealing with hard-constrained problems.
The optimization problem as posed involves a very large number of decisions to
determine the number and type of sensors to measure each process variable. The main
challenge is that the problem size scales exponentially in the number of process
variables and sensors. An additional complexity is the lack of direct algebraic equations
to compute the reliability of process variable estimation which was typically calculated
by algorithmic procedures. This work is focused on the optimal design and upgrade of
reliable sensor networks. A novel MINLP-based approach that takes into account sensor
redundancy as well as multiple sensor types is introduced. Different objectives
functions besides cost can be easily implemented. The successful applicability of the
proposed method is illustrated through a challenging case study taken from literature.

2. Reliability of process variable estimation

The probability, rj(t), of estimating a process variable j at time t must simultaneously

consider sensor, hardware and functional reliabilities. For the sake of simplicity, the
time reference is omitted from here on. As an example, consider a simple process unit
with one feed stream 1 and two product streams 2 and 3, as shown in Figure 1. For this
process unit, the three flow variables (F1, F2 and F3) are related to each other through
the pure mass balance expressed by the equality F1 = F2 + F3. Let us assume that all
sensor failures occur randomly and independently and that the mass flow of all streams
can be measured using sensors k = S1, S2, S3, S4, S5, S6, whose sensor reliability rsSi are
all equal to 0.80 at time t. Therefore, using the property that, if P1 and P2 are the
probabilities of two dependent events then (P1 or P2) = P1 +P2 –P1.P2, the reliability
of estimating variable F1 at time t is summarized in Table 1. For these three illustrating
cases (sensor reliability, hardware reliability and functional reliability), rF1 > 0 and
variable F1 is said to be observable. Moreover, in the case of hardware reliability, the
reliability of estimating variable F1 is greater than the reliability of each one of the
sensors measuring it, showing the existence of more than one way to estimate F1. Thus,
rF1 is directly related to the degree of redundancy and can still be estimated even if one
sensor fails. However, in cases of sensor reliability and functional reliability, no sensor
An optimization framework to computer-aided design and upgrade of measurement … 3

failure is permitted. These results illustrate the effect of sensor placement on the
reliability of estimating process variables.

Figure 1. Process variable estimation reliability

Table 1. Calculation of reliabilities depending on sensor placement

Reliability equation Reliability rF1(t) Reliability type
rsS1(t) 0.80 Sensor
s s s s s s
r S3(t) or r S4(t)= r S3(t)+ r S4(t) – r S3(t). r S4(t) 0.96 Hardware
s s s s
r S5(t) and r S6(t)= r S5(t). r S6(t) 0.64 Functional

The lowest reliability associated to the estimation of a process variables is used to

determine the entire sensor network reliability, which in turn, is used for the design and
retrofit of the network. Thus, a reliable sensor network design and retrofit problem can
be formulated and solved through a rigorous mathematical programming optimization
problem. Data, variables and constraints involved in the mathematical representation of
this problem are introduced in next section.

3. The proposed MINLP-based approach

Sets: J (process variables); Jk (process variables that can be measured through sensor k);
Je (process variables involved in the estimation e); K (sensor types); Kj (sensor types that
can be used to measure process variable j); N (sensor multiplicities); Njk (allowable
sensor multiplicities for measuring process variable j through sensor k); E (alternative
ways of process variable estimation); Ej (alternative ways for estimating process
variable j)
Parameters: ckacq (acquisition cost of sensor k); ckins (installation cost of sensor k); ckdeins
(deinstallation cost of sensor k); njmin (minimum number of sensors for measuring
variable j); njmax (maximum number of sensors for measuring variable j); nj,kins (number
of sensors k already installed in variable j); rk (reliability of sensor k); rj min (minimum
reliability requirement for variable j); cmax (maximum total cost of the sensor network)
Continuous variables: Fsnsj,k,n (hardware failure probability for measuring process
variable j through n sensors k∈Kj); Fhrdj (hardware failure probability for measuring
variable j); Ffncj (functional failure probability for estimating variable j); Cinsj,k (sensor
installation cost); Cdeinsj,k (sensor de-installation cost); Cost (total sensor network cost);
Rmin (minimum reliability of the sensor network)
Binary variable: Yj,k,n (binary variable denoting that n sensors k∈Kj are used to measure
process variable j)
4 Angelini et al.

3.1. Model constraints

• Sensors already installed in the network (Retrofit).

∑ ninsj ,k ≤ ∑ ∑ n Y j ,k ,n ∀k (1)
j j∈ J k n∈ N j ,k

• Minimum and maximum allowable sensor redundancy.

n min
j ≤ ∑ ∑ n Y j ,k ,n ≤ nmax
j ∀j (2)
k ∈K j n∈N j ,k

• Hardware failure probability considering sensor multiplicity.

,k ,n = 1 − Y j ,k ,n + (1 − rk ) Y j ,k ,n ∀j , k ∈ K j , n ∈ N j ,k
n
F jsns (3)

• Hardware failure probability considering sensor redundancy and multiple sensor

types.

F jhrd = ∏ ∏ F jsns,k ,n ∀j (4)

k ∈K j n∈ N j ,k

• Functional failure probability considering direct measurement and estimation.

⎡ ⎤
F jfnc = ∏ ⎢⎢1 − ∏ (1 − F jhrd
' )⎥
⎥
∀j (5)
e∈E j ⎣ j' ∈ J e ⎦
• Minimum reliability requirement.

1 − F jfnc ≥ r jmin ∀j (6)

• Total sensor acquisition cost.

⎛ ⎞
⎜ ins ⎟
⎜∑ ∑ n Y j ,k ,n − ∑ j ,k ⎟
Ckacq ≥ ckacq n ∀k (7)
⎝ j∈ J k n∈N j ,k j∈ J k ⎠
• Total sensor installation cost.
⎛ ⎞
⎜ ins ⎟
j ,k ≥ ck ⎜ ∑ j ,k ,n
C ins ins
n Y − n j ,k ⎟ ∀j , k ∈ K j (8)
⎝ n∈N j ,k ⎠
• Total sensor de-installation cost.
⎛ ⎞
⎜ nins − n Y j ,k ,n ⎟
j ,k ≥ ck
C deins deins
⎜ j ,k ∑ ⎟
∀j , k ∈ K j (9)
⎝ n∈N j ,k ⎠
• Total sensor network cost.
An optimization framework to computer-aided design and upgrade of measurement … 5

⎛ ⎞
Cost = ∑ ⎜⎜ Ckacq + ∑ C insj ,k + C deins
j ,k
⎟
⎟
(10)
k ⎝ j∈ J k ⎠
• Reliability of the sensor network

1 − F jfnc ≥ R min ∀j (11)

4. The ammonia plant case study

The dynamic case of the ammonia plant initially introduced in Benqlilou et al. (2004) is
addressed in this section to illustrate the performance of the proposed MINLP model.
This problem comprises eight flow and five level process variables which can be
measured through a catalogue of 5 different flow-meters and 3 level meters,
respectively. An schematic representation of this plant is depicted in Figure 2. The
available flow-meters have reliabilities of 0.70, 0.75, 0.80, 0.85 and 0.90 and their
associated costs are 1500, 1700, 2000, 2300 and 2800 euros, respectively. For the level
meters, reliabilities are 0.60, 0.70 and 0.80 having a cost of 2000, 3500 and 5000 euros.
A maximum hardware redundancy of 2 sensors was considered. A large spectrum of
reliability requirements ranging from 0.6 to 0.99 were evaluated in order to test the
performance of the proposed determinist model under different scenarios. The design
and retrofit problems of the sensor network of the ammonia plant were solved through
the MINLP formulation using the optimizers codes BARON and DICOPT. For each
scenario, maximum computational times of 1 minute, 10 minutes and 1 hour were
enforced for BARON whereas no time restriction was needed to DICOP because of the
low CPU time required to prove local optimality. Figure 3 shows the sensor network
cost as a function of the minimum reliability requirements. Solutions obtained through
the different solvers with different time restrictions are compared in this Figure.
For comparison purposes, an enhanced version of the proposed GA-based approach
introduced in Benqlilou et al. (2004) was also utilized. Figure 4 presents the results
obtained by means of the determinist and the stochastic approaches. In order to perform
a fair comparison, a time limit of 10 minutes was enforced for BARON and GA. The
results put on evidence the better performance of the MILP approach, which was able
satisfy the minimum reliability requirements with lower cost in most of the cases.

Figure 2. Schematic representation of the ammonia plant.

 6 Angelini et al.

MINLP-BASED APPROACH
30000
AMMONIA PLANT
(Design and Retrofit)

25000 1 min - 0.01 step - BARON

10 min - 0.03 step - BARON
1 hour - 0.13 step - BARON
0.01 step - DICOPT
20000
Cost

design
15000

10000
retrofit
5000
0,6 0,65 0,7 0,75 0,8 0,85 0,9 0,95 1

Reliability

Figure 3. Comparison of solutions obtained with different time restrictions

GA-MINLP COMPARISON
30000 AMMONIA PLANT
(Design and Retrofit)

25000 10 min - 0.03 step - BARON

10 min - 0.03 step - 5000 ind - GA

0.03 step - DICOPT

20000
Cost

design
15000

10000 retrofit

5000
0,6 0,65 0,7 0,75 0,8 0,85 0,9 0,95 1

Reliability

Figure 4. Comparison of solutions obtained through MINLP and GA-based approaches

5. Conclusions

An efficient MINLP-based approach for the optimal design and retrofit of reliable
sensor networks with minimum cost has been presented. The formulation takes into
account most of the problem features such as sensor redundancy, different sensor types
and sensor network cost. The reliability of estimating individual process variables is
explicitly computed through a set of constraints that considers sensor, hardware and
functional reliabilities. The performance of the model was successfully compared with a
GA-based approach. Results showed a more stable and robust behavior in the MILP
method which was able to find better solutions than the stochastic approach in most of
the cases. Global optimality could also be proved through global optimizers such as
BARON in a reasonable time for a design problem.

Acknowledgments
Financial support received from the European Community (PRISM project MRTN-CT-2004-
512233) is fully appreciated.

Literature
Ali, Y. & Narasimahan, S. (1993). AIChE J. 39, 820 – 826.
Bagajewicz, M. J.& Sánchez, M. C. (2000). Computers and Chemical Engineering. 23, 1757 –
1762.
Benquilou, C.; Graells, M.; Musulin, E. & Puigjaner, L. (2004). Industrial & Engineering
Chemistry Research, 43, 8026 – 8036.
Madron, F. & Veverka, V. (1992). AIChE J. 38, 227 – 236.
Vaclaveck, V. & Loucka, M. (1976). Chemical Engineering Science. 1199 – 1205.

View publication stats