Computers and Chemical Engineering 32 (2008) 2030–2049

Contents lists available at ScienceDirect

Computers and Chemical Engineering

journal homepage: www.elsevier.com/locate/compchemeng

Design of robust, reliable sensor networks using constraint programming

Prakash R. Kotecha, Mani Bhushan, Ravindra D. Gudi ∗
Department of Chemical Engineering, Indian Institute of Technology Bombay, Powai, Mumbai 400076, Maharashtra, India

a r t i c l e i n f o a b s t r a c t

Article history: Various criteria have been considered in the literature for selection of optimal sensor networks. Amongst
Received 24 October 2007 these, maximization of network reliability is an important criterion. While there are several approaches
Received in revised form 1 February 2008 for designing maximum reliability networks, uncertainty in the available sensor reliability data has not
Accepted 3 March 2008
been considered in these designs. In this article we present two novel formulations that incorporate
Available online 26 March 2008
robustness to uncertainties in the reliability data. Towards this end the sensor network design problem
for maximizing reliability is formulated as explicit-optimization (MINLP) problem using failure rates of
Keywords:
sensors which have better scaling properties instead of sensor reliabilities. Constraint programming (CP)
Constraint programming
Reliability
has been used for solving the resulting optimization problems. Use of CP also enables easy generation of
Multi-objective optimization pareto front characterizing trade-offs between performance, cost and robustness for various uncertainty
Robust sensor network design scenarios. The utility of the proposed approach is demonstrated on a case study taken from the literature.
© 2008 Elsevier Ltd. All rights reserved.

1. Introduction

The optimal placement of sensors in a chemical plant has received considerable attention in the literature due to its importance for plant
safety, fault detection and diagnosis, optimal control, and plant economics. In a recent review on smart plant operations (Panagiotis et al.,
2007), the deployment of real time sensor networks (including sensor network design issues) has been identified as a key research area.
The problem of sensor network design is to select a suitable subset of variables to be measured from the complete set of plant variables so
as to satisfy/optimize criteria such as cost, observability, reliability, estimation accuracy, and flexibility, etc. Initial attempts for the design of
optimal sensor networks involved the use of a graph-based branch and bound type algorithm (Meyer, Le Lann, Koehret, & Enjalbert, 1994)
to determine the minimal cost sensor network. In the sensor network design literature, the concepts of reliability and network reliability
were formalized in Ali and Narasimhan (1993). The concept of network reliability (Ali & Narasimhan, 1993, 1995, 1996) accounted for
sensor failures and also indirectly incorporated the observability criterion. This work (Ali & Narasimhan, 1993, 1995, 1996) used a graph
theory-based greedy search algorithm to obtain the most reliable sensor network. A tree type enumeration procedure for the design of
minimal cost network with constraints on precision, availability, resilience and error detectability was reported in Bagajewicz (1997). Later,
this problem was posed as an explicit mixed integer non-linear programming (MINLP) optimization problem in Bagajewicz and Cabrera
(2002). This work also showed the conversion of MINLP to MILP thereby guaranteeing optimality but at the cost of increasing the problem
size. A duality between the reliability and the variance framework for non-redundant, linear sensor network design problem was shown
in Kotecha, Bhushan, and Gudi (2008). This work can enable the solution of the precision problems by greedy search algorithms and the
reliability problems by explicit-optimization formulation. An integration of genetic algorithms (GA) with graph-theoretic concepts to solve
the problem of optimal design of sensor network for linear processes based on optimizing cost, estimation accuracy and network reliability
was proposed in Sen, Narasimhan, and Deb (1998). The advantage of performing corrective maintenance was shown in Sanchez and
Bagajewicz (2000). The design of sensor networks from a fault diagnosis perspective was considered in Bhushan and Rengaswamy (2002a,
2002b). Increasingly, control researchers are also considering sensor related issues such as sensor failures while developing fault tolerant
control strategies (Nael, El-Farra, & Christofides, 2005; Prakash, Narasimhan, & Patwardhan, 2005; Mhaskar, Gani, McFall, Christofides, &
Davis, 2007). Other recent developments include a rigorous methodology based on the use of cutsets (Gala & Bagajewicz, 2006a, 2006b)
and use of Constraint programming (CP) to solve some of the sensor network problems (Kotecha, Bhushan, & Gudi, 2007a).

∗ Corresponding author. Tel.: +91 22 2576 7204; fax: +91 22 2572 6895.
E-mail address: [email protected] (R.D. Gudi).

0098-1354/$ – see front matter © 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2008.03.005
 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2031

Nomenclature

aij parameter relating the ith unmeasured variable to the jth measured variable
ci , C* cost of ith sensor, total monetary allowance
CT set of all the cutsets
e uncertainty tolerated by the network for unknown levels of uncertainty
G graph with V nodes and E edges
K set of key variables
L upper bound on the uncertainty of a network
n, m number of edges/variables/streams, number of nodes/units excluding the environment node
N, M set of variables, set of measured variables
Ri , i reliability, failure rate of sensor measuring variable i
R,  network reliability, network failure rate
Ri exponentially scaled reliability
ˆi
R̂i ,  failure rate, reliability of the ith variable
S, E set of nodes S ⊂ V, edges in S
U set of uncertain sensors
V, E set of all nodes and edges in graph G
xi binary variable to indicate the placement of a sensor on variable i
i ratio of the maximum failure rate to the nominal failure rate for the ith uncertain sensor
max max(l )
l∈U
small least failure rate among the uncertain sensors
ˆc
 critical/threshold network failure rate
i linearly scaled failure rate for ith sensor
ˆ ∗, 
 ˆ U∗ optimal network failure rates for the nominal and maximum failure rate scenario
U
i
maximum failure rate for the ith uncertain sensor
i parameter relating the failure rate of ith variable to the failure rate of its sensor

In this article, we focus on maximum reliability designs. While these have been considered in the literature mainly via enumeration or
greedy search-based approaches (Ali & Narasimhan, 1993, 1995,1996; Bagajewicz, 1997), we present an explicit constrained optimization
formulation using the notion of cutsets. This results in a MINLP problem. This formulation is posed using the failure rates of the sensors
and not the sensor reliabilities. While theoretically equivalent, the use of failure rates enables better numerical-scaling of the data in the
problem. In the literature, the reliability data for the sensors is assumed to be known accurately. However, often exact reliabilities for various
sensors may not be known; instead only approximate values may be available. To design sensor networks that incorporate robustness to
such uncertainty in the reliability data, we present two novel formulations. These formulations also result in MINLP problems. We use
CP, a domain reduction-based technique, to solve these combinatorial optimization problems to global optimality. This is in contrast to
the greedy search-based algorithms in the literature which do not guarantee global optimality. Additionally, typically the sensor network
design problems have multiple global optimal solutions. However, most of the work in sensor network design literature focuses on the
determination of only one optimal solution and neglects the presence of the multiple solutions. These multiple solutions are important
to the designer for it may enable the designer to select a network that satisfies some of the objectives that were not considered during
the optimization procedure. In this work, we also show the utility of CP in determining all multiple solutions of the sensor network
design problem without imposing any integer cuts. Using this multiple-solution generation capability of CP, we also generate trade-off
(pareto optimal) solutions characterizing trade-offs between objectives of: minimizing sensor network cost, maximizing performance as
characterized by network reliability (corresponding to nominal values of sensor reliability data), and maximizing network robustness
to uncertain reliability data. These trade-offs enable the designer a wider choice in deciding the most appropriate sensor network for
a given process. We restrict the discussion in this article to linear, steady state processes (processes involving overall mass flows) and
non-redundant sensor network designs.
The rest of the article is structured as follows: we first begin with the preliminary definitions that are used in this article. We then
develop an explicit MINLP formulation for the determination of the network with minimum failure rate followed by two formulations that
help in the determination of robust networks in the presence of uncertainties in the failure rates. These formulations consider both the
possible scenarios of known and unknown levels of uncertainty. This is followed by a brief introduction to constraint programming. We
then demonstrate the utility of the proposed formulations along with the determination of multiple solutions and pareto fronts generation
for a flowsheet taken from the literature. The article is then concluded by summarizing the proposed work and presenting possible future
extensions.

2. Preliminary definitions

In this section, we provide brief definitions of the relevant terms. A more detailed description can be obtained in Ali and Narasimhan
(1993) and Sanchez and Bagajewicz (2000).
Non-redundant sensor network design for a linear process. A sensor network that employs minimum number of sensors (num-
ber of variables n − number of equations/units m) for a process characterized by overall mass flow balances is called as
2032 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049

non-redundant sensor network design for a linear process. The following discussion is for the non-redundant, linear process
scenario.
Observability. The steady state mass flow balances over each unit of a process can in general be represented by

Ax = 0 (1)

where A is the material balance matrix of the process and x represents the variables (mass flows) of the process. The elements of each row
of the A matrix are either +1, −1 or 0, depending on whether the corresponding flow variable occurs as an input, output or is not associated
with the unit, respectively, for which the flow balance is written. Moreover, the matrix A can be partitioned corresponding to the measured
variables xm and the unmeasured variables xu as Am and Au , respectively, as:

Au xu + Am xm = 0 (2)

Hence the unmeasured variables can be estimated using

xu = −A−1
u Am xm (3)

If the matrix A−1

is not invertible, then the system is unobservable. This implies that not all unmeasured variables can be estimated. All
u
variables which can be estimated either by direct measurement, or their relationships with other measured variables are referred to as
observable variables, and a process where each variable is observable is defined as an observable system. Unlike the traditional control
literature (where measurements of combinations of variables are allowed), in the current work (and in the traditional sensor network
design literature) it is assumed that each variable can be directly measured. Given that we are working with overall flowrates, this is a
reasonable assumption.

2.1. Reliability

Sensor reliability. Reliability R(t) of a sensor at time t is the probability that the sensor has not failed between time 0 and t. In general,
reliability is a function of time and can be experimentally determined by testing same type of sensors and finding the ratio of the number
of sensors working at time t to the total number of sensors at time t = 0.
Variable reliability. Reliability of a variable is the probability of being able to estimate the variable between time 0 and t. If a variable
is measured, its reliability is equal to the reliability of the sensor measuring that variable. However, if a variable is not measured but is
observable, its reliability is given by the product of reliabilities of the various sensors that are used in estimating that variable. Thus the
reliability of an observable variable can be expressed as
 ai
R̂i (t) = Rj (t) j (4)
j∈M

In the above equation, if the ith unmeasured variable is independent of the jth measured variable, then aij is zero else aij is one, and M is
the set of selected sensors.

2.2. Instantaneous failure rate (or hazard rate)

Sensor failure rate. The instantaneous failure rate of a component (Dhillon, 1983) is time dependent and is defined as

1 dRi (t)
i (t) = − (5)
Ri (t) dt

Based on the failure rates, most components can be associated with three phases in their life cycle (Dhillon, 1983). The first phase is
the infant mortality phase where the components are prone to early failures. The failure rate in this phase keeps decreasing with respect
to time. This phase is followed by the constant failure rate phase wherein the failures are attributed to random failures. The second phase
forms a major part of the component’s life time. The third and the final stage is the increasing failure rate period with the failures being
attributed to wear-outs and thereby warranting the removal of the component from its usage. These three phases form the life time of
a component and this behavior is commonly known as the bath tub curve for its resemblance to a bath tub. Unlike the dimensionless
reliability, the unit of failure rate is time−1 . In this article, we will consider the constant failure rate scenario, since this period is typically
the “useful life” of the component (Dhillon, 1983). For the constant failure rate case, it is seen from Eq. (5) that:

Ri (t) = exp(−i t) (6)

ln Ri (t)
⇒ i = − (7)
t
Variable failure rate. Similar to reliability of a variable, we now define the failure rate of a variable and relate it to the failure rates of the
sensors required to estimate that variable (assuming it is observable). By substituting Eq. (6) into Eq. (4) and taking natural log on both
sides we get the desired relationship as:

ˆ i = − ln(R̂i (t)) =
 aij j (8)
t
j∈M
 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2033

3. Sensor network design: base case optimization formulations

3.1. Sensor network design for maximizing network reliability

Using the concept of “a chain being no stronger than its weakest link”, Ali and Narasimhan (1993) have defined the network reliability
to be the minimum reliability amongst all variables (measured or unmeasured):

R(t) = minR̂i (t) (9)

∀i ∈ N

Even though reliabilities are functions of time, in the typical sensor network design literature (Ali & Narasimhan, 1993, 1995, 1996;
Bagajewicz & Sanchez, 2000), constant values of reliabilities have been considered. The sensor network design problem then is: given
(constant) sensor reliabilities, choose variables to be measured so as to maximize R. The given sensor reliabilities can be thought of as
corresponding to an arbitrary time t.
Scaling issues in reliability framework. The reliability of any variable (measured or unmeasured) is less than or equal to one. This can be a
source of computational issue while solving the maximum network reliability problem, since the reliability of an unmeasured variable can
be a very small value as it involves the product of sensor reliabilities as seen from Eq. (4). Such a scenario can typically arise either when
the sensor reliabilities are very small or in very large flowsheets wherein the estimation of a variable can depend on a substantial number
of sensors. For example, consider a scenario wherein an unmeasured variable is to be estimated by 10 sensors with their reliabilities of
the order of 10−2 . Thus, the reliability of the unmeasured variable is of the order of 10−20 and this can lead to computational problems
while designing sensor networks for maximizing network reliability. Ideally, in such scenarios, one would like to scale the reliability to
avoid working with smaller values. However, the scaling of reliability is not possible in a linear fashion, i.e., it is not possible to multiply
the reliabilities of all the sensors by a constant value and yet preserve the equivalence between the sensor network design problem with
the original and the scaled reliability values. Instead, the reliabilities can be scaled exponentially, i.e., if Ri is the given reliability of sensor
i, then instead of Ri the following value can also be used for reliability of sensor i for the sensor network design problem:

Ri (t) = (Ri (t))˛ , ∀ i∈N (10)

where ˛ ∈ (0, ∞) is the chosen scaling coefficient. It can be seen from Eq. (6) that this scaling essentially corresponds to specifying sensor
reliabilities at different times t since if t = ˛t, then
˛
(Ri (t))˛ = (exp(−t)) = exp(−t  ) = Ri (t  ) (11)

However, despite the possibility of this exponential scaling, numerical issues with very small unmeasured variable reliabilities can still
occur since each sensor reliability cannot be greater than 1 irrespective of the choice of ˛ in Eq. (10).
To overcome this drawback, in this article, we propose working with the failure rates of sensors instead of their reliabilities since the
failure rates can be linearly scaled. This issue and the resulting sensor network design problem are discussed next.

3.2. Sensor network design for minimizing network failure rate

Similar to the concept of network reliability defined earlier (Ali & Narasimhan, 1993), we define the network failure rate as the maximum
failure rate amongst all variables.

ˆi
 = max  (12)
∀ i∈N

The sensor network design problem can then be equivalently stated as: given the failure rates of all the sensors, choose the variables to
be measured in a process so as to minimize the network failure rate. Unlike the sensor reliabilities which are functions of time, the sensor
failure rates are constants (in the “useful life” of the component). In Eq. (9), by taking natural log on both sides (natural log is a monotonically
increasing function), then multiplying by −1 (converting the “min” on RHS to “max”), and dividing both sides by the time t = 0 at which
the constant values of sensor reliabilities have been specified, we get:
 
ln R(t) ln R̂i (t)
− = max − (13)
t ∀i ∈ N t

which on using the relation between failure rate and reliability of a given variable (Eq. (8)) gives rise to Eq. (12) where  = −(ln R)/t is the
overall network failure rate.
Hence, it is seen that performing sensor network design to maximize network reliability is exactly equivalent to performing sensor
network design to minimize network failure rate. For the sake of brevity, only a short proof of this equivalence was sketched in the
discussion above. For more details, an interested reader can refer to (Kotecha et al., 2008) where the proof of duality between reliability
and equivalent variance framework is analogous to the proof of equivalence between reliability and equivalent failure rate framework.
As mentioned earlier, working in the failure rate framework avoids the numerical-scaling problems encountered in reliability, since as
seen from Eq. (8), the calculation of failure rate of a variable involves the summation of failure rates of the sensors used for estimating the
variable. This situation is in contrast to the reliability case where products of sensor reliabilities were involved. Due to this summation,
the sensor failure rates can be multiplied by an arbitrary scaling constant ˛ ∈ (0, ∞) while ensuring the equivalence between the sensor
network design problems with the original and the scaled failure rates. This can be easily seen by multiplying Eq. (12) by ˛ leading to

 = ˛ = max˛ ˆ
ˆ i = max
i
(14)
i∈N i∈N
2034 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049

where from Eq. (8) it is seen that:

 
ˆ  = ˛
 ˆi = aij (˛j ) = aij j (15)
i
j∈M j∈M

where j = ˛j is the scaled failure rate of the jth sensor and  ˆ  ,  are the corresponding scaled failure rates of the ith variable and the
i
overall network. It is this linear scaling in the failure rate framework that enables avoidance of numerical issues that can occur during
reliability calculation. It can be further noticed that the exponential scaling discussed in reliability framework (Eqs. (10) and (11)) actually
corresponds to the linear scaling in the failure rate framework.
In the rest of the article, to increase the readability, we will be interchangeably using the terminologies: reliability and failure rate, in
the sense that minimization of failure rate corresponds to the maximization of reliability and vice versa. Additionally, for easy notation
we drop the time dependence of all reliabilities since they are specified at an arbitrary time. We next present an explicit-optimization
formulation to determine the minimum failure rate network.

3.2.1. Explicit-optimization formulation

In this section, we present an explicit-optimization formulation for the minimization of the network failure rate based on concepts
from graph theory. While Eq. (8) can be used to calculate the failure rate of a variable given the measurements, to be able to pose the
sensor network design problem as an explicit-optimization problem, an expression for calculating the failure rate independent of the
measurements is required. Using well-known concepts from graph theory, we propose such an explicit-optimization formulation. We
restrict ourselves to a brief discussion of graph-theory related concepts and the interested reader is referred to Deo (1974) for more details.
The sensor network design problem can be stated as
Given: (i) The process equations (ii) the failure rates of all the sensors available for measuring the process variables.
Determine: The variables to be measured subject to the set of constraints if any.
Variables: We introduce a binary variable xi , i ∈ N such that xi = 0 will indicate that the ith variable is measured while 1 will indicate it is
unmeasured. It can be seen that the number of binary variables will be equal to the total number of variables in the process.
Constraints: The set of constraints should ensure that the sensor network is observable. This is discussed next.
Observability: As discussed in Ali and Narasimhan (1993), a linear (overall mass flow balance) process can be represented as a graph
G = (V, E) where V denotes the set of nodes (units (equations) corresponding to flow balances) and E denotes the set of edges (streams
(variables) incident on the units). The graph also contains a hypothetical node, called the environment node which corresponds to overall
flow balance for the entire flowsheet. With the inclusion of the environment node, the graph becomes cyclic. In this graph, for a non-
redundant observable sensor network design, the set of unmeasured edges should form a spanning tree (Ali & Narasimhan, 1993) where a
spanning tree is the minimum set of edges such that the graph is connected. All the edges in the graph which are not part of the spanning
tree should be measured. For a graph with m + 1 nodes, a spanning tree will have m edges. This corresponds to the number of independent
equations and hence the number of unmeasured variables in the process. Different spanning trees correspond to different sets of measured
variables. Hence, to ensure observability, constraints enforcing that the unmeasured variables form a spanning tree of graph G will be
implemented. These will be discussed later when the actual formulation is presented.
Additionally, depending on the objective function, one of the following constraints can be specified:

(i) Cost constraint. This constraint ensures that the cost of the selected sensor network is less than the maximum available monetary
resource

ci (1 − xi ) ≤ C ∗ (16)
i∈N

where ci is the cost of locating sensor for measuring variable i and C* is the total available cost. The cost constraint can be specified
when the objective is to minimize network failure rate.
(ii) Constraint on network failure rate. For problems where the objective is to minimize the sensor network cost, the following constraint
that ensures that the network failure rate is below a threshold value can be specified:

ˆ 1, 
max( ˆ2 ...
ˆ n) ≤ 
ˆc (17)

where  ˆ c is the critical network failure rate. This constraint with cost minimization objective ensures that the least cost network
satisfying the desired performance criteria ( ˆ c ) is selected. Additionally, thresholds on failure rates of individual variables can also be
included as constraints. Explicit expression for calculation of  ˆ i is next discussed.

Expression for calculation of failure rate of a variable independent of the sensor network. Given a process, corresponding to different sets
of measurements, there are different ways of estimating a variable. In graph theory literature, all such possibilities of estimating the
variable correspond to cutsets involving that variable. In a graph G, a cutset is the minimal set of edges whose removal disconnects G (Ali
& Narasimhan, 1993). We will denote the set of all cutsets involving variable i by the set Ci , the kth cutest from this set by Cki and the set of
total cutsets of a flowsheet by CT . For an observable sensor network, each unmeasured variable will be involved in exactly one cutset each
such that all other variables in that cutset are measured (Meyer et al., 1994). Thus the general expression for calculating reliability R̂i of a
 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2035

variable i can be given by

⎧ ⎛ ⎞ ⎫
⎪
⎪ ⎪
⎪
⎪
⎪ ⎜ ⎟ ⎪
⎪
⎨ ⎜  ⎟ ⎬
R̂i = max (1 − xi )Ri , ⎜
⎜ Rj (1 − xj )⎟
⎟, ∀ k = 1, 2, . . . , |C i | , ∀ i∈N (18)
⎪
⎪ ⎝ ⎠ ⎪
⎪
⎪
⎪ j ∈ Cki , ⎪
⎪
⎩ ⎭
j =i

where |Ci | denotes the cardinality (number of cutsets involving variable i) of set Ci . The above equation can be rewritten as
⎡ ⎤
|C i | ⎢ ⎥
 ⎢  ⎥
R̂i = (1 − xi )Ri + xi ⎢ Rj (1 − xj )⎥ , ∀ i ∈ N (19)
⎢ ⎥
k=1 ⎣ i j ∈ Ck ,
⎦
j = i

The first term on the RHS corresponds to direct measurement of the ith variable (corresponding to xi = 0), while the second term
(corresponding to xi = 1) incorporates all possible ways of indirectly estimating the ith variable based on other measurements. For the
second term on RHS, since there is only one cutset involving variable i which will have all other variables (other than i) measured, only one
term in the summation will be active. Using Eq. (7) (with time t = 1) in (18), the explicit equation for failure rate can be written as
⎧ ⎛ ⎞ ⎫
⎪
⎪ ⎪
⎪
⎪
⎪ ⎜ ⎟ ⎪
⎪
⎨ ⎜  ⎟ ⎬
ˆ i = min
 − ln((1 − xi )Ri ), − ln ⎜
⎜ Rj (1 − xj )⎟
⎟, ∀ k = 1, 2, ..., |C i | , ∀ i∈N (20)
⎪
⎪ ⎝ ⎠ ⎪
⎪
⎪
⎪ j ∈ Cki , ⎪
⎪
⎩ ⎭
j = i

However, Eq. (20) is equivalent to the following equation in the presence of tree constraints (exactly one term in the RHS of Eq. (20) will
be finite).
⎡⎛ ⎞⎛ ⎞⎤
|C i | ⎢⎜ ⎟⎜ ⎟⎥
 ⎢⎜  ⎟ ⎜  ⎟⎥
⎢⎜ ⎟ ⎜ (1 − xj )⎟⎥
⎟⎥ , ∀ i ∈ N
ˆ i = (1 − xi )i + xi
 j ⎜ (21)
⎢⎜ ⎟
k=1 ⎣⎝ i ⎠⎝ i
j ∈ Ck ,
⎠⎦
j ∈ Ck ,
j = i j = i

It can be seen that Eq. (21) does not involve non-linear operators such as min and ln. Hence, we will be using Eq. (21) for determining the
failure rate of a variable.
Based on the above discussion, we next present the mathematical formulation for the base case optimization based on the concept of
failure rate. We broadly consider three objectives, namely: (i) minimizing cost of the sensor network, (ii) minimizing network failure rate
and (iii) minimizing cost of the sensor network subject to thresholds on failure rates of some of the key variables.
Minimization of cost. The objective is to minimize the overall cost of the sensor network given as

C= ci (1 − xi ) (22)
i∈N

Minimization of network failure rate. If the objective is to design a reliable sensor network, i.e., minimize the network failure rate, the
objective function is given as:
ˆi
 = max  (23)
∀ i∈N

With the above description of the problem, the overall optimization problem formulation is:

• F1:

Min cost or network failure rate (24)


∗
s.t. ci (1 − xi ) ≤ C
i∈N (25)
or
max( ˆ2 ...
ˆ 1,  ˆ n) ≤ 
ˆc

xi ≤ |S| − 1, ∀S ⊂ V, |S| > 2 (26)
i ∈ E

(1 − xi ) = n − m (27)
i∈N
2036 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049

⎡⎛ ⎞⎛ ⎞⎤
|C i | ⎢⎜ ⎟⎜ ⎟⎥
 ⎢⎜  ⎟ ⎜  ⎟⎥
ˆ i = (1 − xi )i + xi
 ⎢⎜  ⎟ ⎜ (1 − x ⎟⎥ ∀ i ∈ N
j ⎟⎥ ,
) (28)
⎢⎜ j⎟ ⎜
k=1 ⎣⎝ i ⎠⎝ i
j ∈ Ck ,
⎠⎦
j ∈ Ck ,
j = i j = i

xi ≥ 1, ∀ k ∈ 1, 2, ..., |C T | (29)
i ∈ CT
k

xi ∈ {0, 1}, ∀ i∈N (30)

In formulation F1, depending on the objective function, the appropriate constraint from Eq. (25) can be chosen. For better readability,
the formulation involving minimization of network failure rate subject to cost constraints will be referred to as F1(a) while the formulation
involving minimization of cost subject to threshold on network failure rate will be referred to as F1(b). Other than constraint (25), the
rest of the constraints are common to both the formulations. Eqs. (26) and (27) are required to ensure observability of all variables and
non-redundant sensor network design. Eq. (26) ensure that the unmeasured variables form a tree (Deo, 1974). S denotes a subset of the set
of nodes V subject to the criterion that the cardinality of S (number of nodes in S) is greater than 2. E denotes the set of edges connecting
the nodes present in S. Eq. (26) is written for all subsets of V with cardinality greater than 2. Eq. (28) are the same as Eq. (21) for failure
rate calculation written for each variable. Constraints (29) ensure that at least one variable in each cutset is not measured. The problem
is defined completely even without the inclusion of these cuts and hence these are redundant cuts. They have been included as they
were observed to substantially reduce the computational burden while solving the above MINLP optimization formulation using CP. The
inclusion of these cuts does not involve any additional effort as the cutsets have to be anyway determined for writing the expressions for
failure rates in Eq. (28). The algorithm for determining all the cutsets in a graph is presented in Appendix A.

Remark 1. It can be noted that formulation F1(a) will be infeasible if the available cost C* is not sufficient to ensure observability of all
ˆ c for a
variables. Formulation F1(b) on the other hand will be infeasible if it is not possible to achieve the threshold network failure rate 
non-redundant sensor network design.

Remark 2. Apart from the above constraints, other process/network specific constraints can also be included in the formulations. However
in our work we have not included any such network specific constraint.

Minimal cost network subject to thresholds on key variables. Another frequently considered practical scenario in the sensor network
design literature is to choose a minimal cost observable network such that the reliabilities of key variables are greater than some specified
threshold values. This problem can be formulated as follows:

• F1(c):

Min ci (1 − xi ) (31)
i∈N

s.t. xi ≤ |S| − 1, ∀ S ⊂ V, |S| > 2 (32)
i ∈ E

(1 − xi ) = n − m (33)
i∈N

⎡⎛ ⎞⎛ ⎞⎤
|C i | ⎢⎜ ⎟⎜ ⎟⎥
 ⎢⎜  ⎟ ⎜  ⎟⎥
(1 − xi )i + xi ⎢⎜ j ⎟ ⎜ (1 − xj )⎟⎥ ≤ i i , ∀ i ∈ K (34)
⎢⎜ ⎟⎜ ⎟⎥
k=1 ⎣⎝ i ⎠⎝ ij ∈ Ck ,
⎠⎦
j ∈ Ck ,
j = i j = i


xi ≥ 1, ∀ k ∈ 1, 2, . . . , |C T | (35)
i ∈ CT
k

xi ∈ {0, 1}, ∀ i∈N (36)

where i i denotes the threshold failure rate on the ith variable and K denotes the set of key variables. The constant i indicates the
allowable deterioration (i > 1) or required improvement (i < 1) in the estimated failure rate of the variable as a factor of its sensor failure
rate i .

4. Sensor network design: optimization formulations to incorporate robustness to uncertainty in data

The explicit-optimization formulation F1 was based on the fact that the sensor failure rates are exactly known. However, this is seldom
the case in practice as failure rates may be either determined experimentally or values obtained for other similar components may be
 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2037

used. Additionally the failure rates will also be a function of the process conditions and it will be difficult to accurately estimate them in
laboratory conditions. In this section, we present two formulations that can accommodate this uncertainty in failure rates and aid in the
determination of robust networks. Both these formulations require the knowledge of the set of uncertain sensors U whose failure rates are
not precisely known, however nominal (most likely) values for the failure rates are available.

4.1. Robustness with unknown level of uncertainty

This is a scenario where the uncertainty levels in some sensors are not known and it is desired to determine a network that is not only
reliable but also robust to the uncertainties. To incorporate robustness to these unknown uncertainties in the chosen design, we formulate
the problem by proposing to choose a sensor network that will maximize the increase it is able to tolerate in the failure rate of the uncertain
sensors. However, this tolerance will be in general different for different uncertain sensors. We then choose the minimum tolerance amongst
all the uncertain sensors as characterizing the robustness of the network, and hence the objective function is to maximize the minimum
tolerance. This is similar to the concept of robustness as used in Kotecha et al. (2007a) for robust sensor network design for fault diagnosis.
The resulting optimization problem F2 is:

• F2:

Max e (37)
xi ,e

s.t. xi ≤ |S| − 1, ∀ S ⊂ V, |S| > 2 (38)
i ∈ E

(1 − xi ) = n − m (39)
i∈N

⎡⎛ ⎞ ⎤
⎛ ⎞
⎢⎜ ⎟ ⎥
⎢⎜ ⎟ ⎥
⎢⎜  |C i | ⎟ ⎜ ⎟⎥
 ⎢⎜  ⎟ ⎜  ⎟⎥
⎢⎜ j ⎟ ⎜ (1 − xj )⎟⎥
⎟⎥ , ∀ i ∈ U
ˆ i = (1 − xi )(i + e) + xi
 (j + e) +
⎢⎜ ⎟ ×⎜ (40)
k=1 ⎢⎜ j ∈ C i , ⎟ ⎝ ⎠⎥
⎢⎜ j ∈ Ck , ⎟
i i
j ∈ Ck , ⎥
⎣⎝ j = i,k j = i, ⎠ j= i ⎦
j∈U j∈
/U

⎡⎛ ⎞ ⎤
⎛ ⎞
⎢⎜ ⎟ ⎥
⎢⎜ ⎟ ⎥
⎢⎜  |C i | ⎟ ⎜ ⎟⎥
 ⎢⎜  ⎟ ⎜  ⎟⎥
⎢⎜ j ⎟ ⎜ (1 − xj )⎟ ⎥
⎟⎥ , ∀ i ∈ N\U
ˆ i = (1 − xi )i + xi
 (j + e) +
⎢⎜ ⎟ ×⎜ (41)
⎢⎜ ⎟ ⎝ ⎠ ⎥
k=1
⎢⎜ j ∈ Cki , j ∈ Cki , ⎟ j ∈ Cki , ⎥
⎣⎝ j = i, j = i, ⎠ j= i ⎦
j∈U j∈
/U

  
xi ≥ 1, ∀ k ∈ 1, 2, . . . , C T  (42)
i ∈ CT
k

max( ˆ2 ...
ˆ 1,  ˆ n) ≤ 
ˆc (43)

ci (1 − xi ) ≤ C ∗ (44)
i∈N

xi ∈ {0, 1}, ∀ i ∈ N; e ∈ [0, L] (45)

The objective function in Eq. (37) is the minimum tolerance amongst uncertain sensors and that has to be maximized. Eqs. (38) and (39)
ensure observable non-redundant design while Eq. (40) are used to determine the failure rates of the variables with uncertain sensors.
Eq. (41) are similar to (28) and are used to determine the failure rates of the variables that can be measured with sensors with precisely
known failure rates. Eq. (42) represents the cuts based on the cutsets. Eqs. (43) and (44) ensure satisfaction of minimum performance
specifications for the network failure rate and network cost. In particular, Eq. (43) ensures that the network failure rate is below a critical
ˆ c while Eq. (44) ensures that the cost of the network is below the allowable monetary resource C* . In formulation F2, L is a
failure rate 
large number corresponding to being able to tolerate practically infinite uncertainty in the failure rate of a sensor and is required to ensure
that the solution to F2 is always bounded (even if no uncertain sensor is selected thereby allowing e to achieve its maximum value L).
Formulation F2 can have three types of solutions namely: (i) Infinite robustness. A value of L for the objective function in F2 will indicate
that the sensor network is infinitely robust to the uncertainties indicating that none of the uncertain sensors have been selected. This is the
ideal situation. (ii) Limited robustness. These networks are characterized by objective function value between 0 and L. (iii) Zero robustness.
2038 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049

This indicates that no uncertainty can be tolerated if the network failure rate is to be maintained at or below the critical failure rate  ˆc
with the available cost C* . Then the only possibility of incorporating some robustness to uncertain failure rates will be to decrease the
performance specification, i.e., increase the threshold  ˆ c for network failure rate value and/or increase the available cost C* . It should be
noted that as the set of uncertain sensors increases the robustness can start decreasing from infinity to zero.

Remark 3. The value of L in formulation F2 has to be chosen carefully. It should be sufficiently large to avoid getting suboptimal solutions
ˆ c − small ,
but not very large as it may affect the computational speed. An upper bound on the value of L that can prevent sub optimality is 
where ␭small is the least nominal failure rate among all the uncertain sensors.

4.2. Robustness with known levels of uncertainty

In formulation F2, the only available information was the set of uncertain sensors but without any knowledge of the level of uncertainty.
We now consider the determination of robust networks when the uncertainty levels are known for the uncertain sensors. This can be
viewed as the determination of performance for the worst case scenario. Apart from the nominal failure rates of sensors, this formulation
also requires the value of maximum failure rate U l
for the lth uncertain sensor. For convenience, this quantity can be expressed relative
to the corresponding nominal failure rate l as U l
= l l . The objective function in this formulation is the minimization of the maximum
worst case network failure rate subject to threshold on the network failure rate for the nominal scenario.

• F3:
   
ˆU ,
Min max  ∀ i∈N (46)
xi i


s.t. xi ≤ |S| − 1, ∀ S ⊂ V, |S| > 2 (47)
i ∈ E

(1 − xi ) = n − m (48)
i∈N

xi ≥ 1, ∀ k ∈ 1, 2, . . . , |C T | (49)
i ∈ CT
k

⎡⎛ ⎞ ⎛ ⎞⎤
|C i | ⎢⎜ ⎟ ⎜ ⎟⎥
 ⎢⎜  ⎟ ⎜  ⎟⎥
ˆ i = (1 − xi )i + xi
 ⎢⎜  ⎟ × ⎜ (1 − x ⎟⎥ ∀ i ∈ N
j ⎟⎥ ,
) (50)
⎢⎜ j⎟ ⎜
k=1 ⎣⎝ i ⎠ ⎝ i ⎠⎦
j ∈ Ck , j ∈ Ck ,
j = i j = i

⎡⎛ ⎞ ⎛ ⎞⎤
|C i | ⎢⎜ ⎟ ⎜ ⎟⎥
 ⎢⎜  U ⎟ ⎜  ⎟⎥
ˆ U = (1 − xi )U + xi ⎢⎜ ⎟ ⎜ (1 − xj )⎟ ⎥
 i i ⎢⎜ j ⎟ × ⎜ ⎟⎥ , ∀ i ∈ N (51)
k=1 ⎣⎝ i ⎠ ⎝
j ∈ Ck ,i ⎠⎦
j ∈ Ck ,
j = i j = i

ˆ 1, 
max( ˆ2 ...
ˆ n) ≤ 
ˆc (52)

ci (1 − xi ) ≤ C ∗ (53)
i∈N

xi ∈ {0, 1}, ∀ i∈N (54)

where for sensors with no uncertainty U

j
= j . Formulation F3 can have the following types of solutions namely: (i) Infinite robustness.
It can be noted that the best robust network for the above formulation will be indicated by  ˆ U∗ = 
ˆ ∗ where 
ˆ U∗ is the optimal solution
for F3 and ˆ ∗ is the optimal, nominal network failure rate (i.e., solution of formulation F1(a)). This implies that despite consideration
of uncertainty the optimal network failure rate does not increase. (ii) Limited robustness. Such scenarios correspond to the objective
function value being larger than  ˆ ∗ . The difference between the objective function and 
ˆ ∗ indicates the increase (deterioration) in the
network failure rate if true failure rates of the uncertain sensors were at their maximum given values. Additionally, the objective function
is bounded between  ˆ ∗ and max 
ˆ ∗ where max = max(l ).
l∈U

Remark 4. If all the sensors were to be uncertain with a failure rate of U

l
= l , then the above formulation will have an optimal failure
ˆ ∗ with the same set of optimal solutions as in F1(a).
rate of 
 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2039

Remark 5. While both F2 and F3 incorporate robustness to the uncertain data, they deal with different scenarios. For unknown level of
uncertainty case (formulation F2), the designer only has the approximate failure rate data of the sensor but has no idea of its correctness.
The known uncertainty level case (formulation F3) on the other hand can be used when the designer has an idea of the worst case value of
failure rate. Typically in designs when dealing with uncertainty, probability distributions of uncertain data are often required (or assumed)
(Diwekar, 2003). If for the design problem considered in this article, probability distribution of uncertain failure rate of sensor i is available,
then in formulation F3, the nominal failure rate can correspond to the mean (or any other measure of centrality) of the distribution, while
Ui
can correspond to a suitable upper limit from the distribution, such as the value corresponding to 95 percentile. Depending on the
situation at hand, the designer can choose either F2 or F3 for incorporating robustness.

Remark 6. All the formulations presented in this article can be solved using an alternate, explicit-optimization-based approach that
does not make use of cutsets. Such an approach will involve the use of duality proposed in Kotecha et al. (2008) using which the failure
rate (or reliability) needs to be converted into an equivalent variance for an explicit MINLP formulation. However, the drawback of such an
approach includes (i) the non-availability of robust solvers to solve these formulations to global optimality as well as difficulties in obtaining
multiple optimal solutions, and (ii) the appropriate choice for the value of the equivalent variance for the fake sensor. As mentioned in
Bagajewicz and Cabrera (2002), a large value is required (to approximate infinity as being the variance of a fake sensor) but this can lead
to computational issues.

In formulations F1–F3 presented so far, we have considered various objectives, such as the nominal network failure rate, sensor network
cost, and robustness in the sensor network as separate single-objective optimization problems. We next discuss some multi-objective
optimization strategies for selecting sensor networks which in some sense simultaneously consider these multiple objectives.

5. Multi-objective optimization formulations

The formulations presented so far consider only one objective at a time. But we may be interested in simultaneously optimizing
multiple objectives, namely performance (in terms of nominal network failure rate), cost and robustness. Such multi-objective optimization
problems are often encountered during design activities. A key characteristic of multi-objective optimization problems, which makes them
different from single-objective optimization problems, is that it is often not possible to simultaneously optimize all the objectives, instead
these problems are usually characterized by a set of optimal solutions. We next briefly discuss some commonly used approaches for
multi-objective optimization problems.

5.1. Pareto optimization

This strategy involves the evaluation of trade-offs between the various objectives. The trade-off solutions are typically represented as
a pareto-optimal front comprising of the set of non-dominated solutions. A solution is said to be non-dominated if it is feasible and there
is no other feasible solution which has better values for all the objectives (Deb, 2001). In the case study section we present results that
denote the trade-offs between various objectives considered in this article for sensor network design.

5.2. Lexicographic optimization

Lexicographic optimization (or preemptive optimization) is a special form of multi-objective optimization in which the various
objectives are given a precedence ordering. The primary objective is given the highest priority followed by decreasing priorities on
subsequent objectives. The basic philosophy in this approach is that even a marginal improvement in a higher precedence objective
is considered more valuable than an arbitrarily large improvement in a lower ranked objective (Deb, 2001; Bhushan, Narasimhan,
& Rengaswamy, 2008; Sherali, 1982). It should be noted that the addition of a subsequent objective makes sense only if the pre-
ceding objective has multiple optimal solutions. In our case, if the minimization of network failure rate is a primary objective with
robustness as a secondary objective then the value of critical failure rate in F2 and F3 should be equated to the value of optimal
ˆ ∗ ). Then the formulation F2 (F3) in conjunction with F1(a) ensures the desired lexicographic
ˆc = 
nominal failure rate of F1(a) (i.e., 
optimization. As this is a lexicographic approach, formulation F2 (F3) is guaranteed to have at least one feasible solution. More-
over, all solutions of F2 (F3) will form a subset of all the multiple solutions of F1(a). Similarly, lexicographic optimal solutions with
respect to objectives of nominal network failure rate and cost can also be generated by appropriately solving formulations F1(a)
and F1(b) in a sequence reflective of the preemptive ordering. Results for such scenarios are also included in the case study sec-
tion.
Many of the other strategies used to solve the multi-objective problems convert the multiple objectives into a single objective. But this
requires some kind of a priori information either by association of weights to the individual objectives or the specification of minimum
performance. The optimal solution to these problems is dependent on this a priori information. Moreover, it may also not be possible to
associate weights or performance targets with all the objectives. Some of the widely used techniques that convert the multi-objective
problems to single-objective optimization problems include goal programming and ε constraint method. However, apart from various
other drawbacks, these algorithms find only one pareto-optimal solution in one simulation run.
We now present a brief overview of CP which has been used for solving the various formulations posed in this article.

6. Constraint programming

Constraint programming was developed in the Computer Science and Artificial Intelligence community and has been widely used to solve
combinatorial feasibility problems. It is an intelligent enumeration-based search technique that uses constraint propagation as its inference
engine and continuously reduces the domain of its variables to reach feasible solutions. It solves an optimization problem by transforming
it into a feasibility problem. The constraint programming technique can be considered to consist of two distinct parts namely constraint
2040 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049

Fig. 1. Schematic diagram of constraint programming.

propagation and a search technique. Constraint propagation is used to reduce the domain of variables till all the variables are assigned a
unique value from their domains. If any of the variable domain becomes empty, it means that there is no feasible solution. However, in the
absence of a unique value for the variables from their domains, a non-deterministic choice is made and constraint propagation is carried
further. If this choice does not yield a solution, backtracking is done to make another choice and the procedure is repeated. The process of
making a non-deterministic choice is called as choice point selection and reaching infeasibility for a particular choice point is known as
failure. Fig. 1 shows a generic schematic of the solution procedure using constraint programming. The important merits of CP include: (i)
easier modeling because of its expressive power, for example, conditions such as = or implication constraints can be directly implemented
(without any transformation), (ii) guarantee of global optimality even for non-convex problems, (iii) ease of determination of multiple
optimal solutions, and (iv) determination of globally optimal pareto solutions. A detailed overview of CP is available in Kotecha et al.
(2007a), Lustig and Puget (2001), Van Hentenryck (2002) and Hooker (2000). We next briefly discuss CP-based strategies for determining
multiple solutions of a single-objective optimization problem and the determination of pareto-optimal solutions for a multi-objective
optimization problem.

6.1. Determination of multiple solutions of a single-objective optimization problem

A very important advantage of CP is the ease with which it can determine multiple optimal solutions of an optimization problem. Such
solutions, if available, provide additional flexibility to the designer in choosing the most appropriate design. Consider an optimization
problem P1 with the maximization of f(x) as the objective function and a set of constraints represented by G(x) as shown below

• P1:

Max f (x)
(55)
s.t. G(x)

where x denotes the decision vector. Since we propose to work in the CP framework, it may be noted that the set of constraints need not
be inequalities but can be any expressive construct supported by the CP programming language. Let the optimal value of P1 be f * . We now
solve a subsequent feasibility problem P2

• P2:

Solve G(x)
(56)
f (x) = f ∗

All feasible solutions of P2 correspond to all the multiple optimal solutions of P1. Thus, all optimal solutions of an optimization problem
can be obtained by solving a two-step problem. Infact, the fathoming condition in CP can also be modified thereby obtaining all solutions
in one step by solving (55). For a lexicographic optimization problem, it is clear that if f(x) is the primary objective then the lexicographic
optimal solutions are subset of the multiple optimal solutions of P2. For example, if h(x) is the secondary objective, then all the multiple
lexicographic optimal solutions can be obtained by simply listing the solutions of P2 with maximum h(x) value (assuming h(x) is to be
maximized). This strategy is followed for generating lexicographic optimal solutions in the case study. A significant advantage of using CP
for determining multiple solutions is that the user has control over the search strategy and can search for multiple solutions that differ
 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2041

only in a certain subset of the decision variables. To understand this, consider the following example

Solve x + y ≤ 10
(57)
(x, y) ∈ {1, 2, .., 5}

It is obvious that there are 25 different solutions. Consider a specific scenario wherein the user is interested in solutions that differ only
in y and not in x. This can be easily handled in CP by not branching on x after obtaining a feasible solution. This search strategy for the
problem in (57) leads to five solutions (x = 1, y = 1, 2, 3, 4, 5). It has to be noted that this set of five solutions is not unique, and depending
on the choice point for x there are four other sets of five solutions each which satisfy the required criteria. This idea of obtaining solutions
differing only in certain variables will be illustrated in the case study.

Remark 7. Another recent important development is the modification of the fathoming rule in the conventional branch and bound
procedure as implemented in BARON (Tawarmalani & Sahinidis, 2002) to obtain the K best solutions. As in BARON, CP can also be used to
determine these K best feasible solutions. Additionally, CP supports several other powerful constructs. For example, a condition that a set
of decision variables should not be equal to one another can be easily handled in CP by either using the all different construct or by directly
including the = operator. Another example would be the ease of handling disjunctive and implication constraints.

6.2. Determination of pareto-optimal solution

The following strategy is used to generate pareto-optimal solutions.

• P3:

Solve G(x) (58)

where G(x) is the set of all constraints. A naı̈ve strategy to generate pareto-optimal solutions would be to obtain all feasible solutions of
P3 and then sort them to identify the non-dominated solutions (Kotecha et al., 2007a). However, determining all feasible solutions can be
quite expensive for combinatorial problems. Hence in this work, we follow a different two-step strategy, as presented in Kotecha, Bhushan,
& Gudi (2007b).

6.2.1. Solution of P3 with Type I cuts

The Type I cut ensures that subsequent feasible solutions are not dominated by the current (ith) feasible solution, and hence for a two
objective problem (with both objectives being minimized) can be written as

(F1 < F1i ) ∨ (F2 < F2i ) (59)

where Fji is the value of the objective function Fj for the ith feasible solution and the symbol ∨ indicates the OR operator. As soon as a
feasible solution is obtained, an additional Type I constraint corresponding to that solution is added to P3, thereby ensuring that future
feasible solutions are non-dominated to all the past feasible solutions. This procedure is carried out till the problem becomes infeasible.
The second step of the strategy is then implemented.

6.2.2. Solution of P3 with Type II cuts

Though the sorting of all the feasible solutions generated in step (i) can give pareto-optimal solutions, the original problem P3 is once
again re-solved after the simultaneous addition of cuts (labeled Type II cuts) listed as

(F1 = F1i & F2 = F2i ) ∨ (F1 < F1i ) ∨ (F2 < F2i ) i = 1 . . . K (60)

where the symbol & indicates the AND operation. Here, K denotes the number of Type I cuts that were added. The solution to this modified
P3 problem is the set of all the pareto solutions along with their multiple realizations (if any). This procedure is summarized in Fig. 2.
More details about the algorithm are discussed in Kotecha et al. (2007b). We adopt a minor change to the above algorithm to reduce the
computational burden for the case study presented later. Instead of adding a Type I cut after the determination of a single solution, we
simultaneously add 1000 Type I cuts after the determination of 1000 feasible solutions. It may happen that some of these 1000 Type I cuts
are redundant. This procedure avoids the frequent initialization of the problem thus resulting in comparatively less computational burden.
This behavior can be attributed to the fact that the subsequent feasible solutions are only marginally better than the current feasible point.
In general, adding simultaneous Type I cuts at a single time need not necessarily benefit but can even lead to high computational times.
However, for the problems presented here, we have observed that there is a significant reduction in the computational time due to this
modification.
We next present a case study to illustrate the ideas proposed in this article.

7. Case study

The proposed formulations are demonstrated on a case study that has been widely used in the sensor network design literature. This
case study was initially proposed by Madron and Veverka (1992) as a retrofit design problem. Meyer et al. (1994) added the cost information
of the various sensors while Bagajewicz and Cabrera (2002) added the flowrates for the various streams. This flowsheet was also considered
by Bagajewicz (Bagajewicz, 1997; Gala & Bagajewicz, 2006a) to demonstrate various ideas for the design and upgrade of sensor networks.
2042 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049

Fig. 2. Algorithm for the determination of pareto points.

7.1. Description of case study

This flowsheet in Fig. 3 was initially proposed as a retrofit problem with 24 variables and 11 equations (overall mass bal-
ances corresponding to 11 units). In the literature (Madron & Veverka, 1992), it was assumed that sensors were already placed
on streams 11–24 and the decision variables were the placement of sensors on streams 1–10. However, we consider grass root
design, and hence will assume that none of the streams are already measured. There are 24 C11 combinations for choosing 13
variables to be measured for a non-redundant sensor network design. However, all these combinations do not form observable net-
works.
The failure rates and the costs of the sensors have been generated randomly as reported in Table 1. The number of tree constraints is
2570. Further details on the number of tree constraints have been discussed in Appendix B. The number of distinct cutsets in this flowsheet
are |CT | = 148 while the number of cutsets for each variable is given in Table 1.

7.2. Results

We now present results of the application of the proposed formulations on this case study. These results were generated with the default
options in ILOG OPL Studio 3.7.1 and CP Solver 6.1 (Solver, 2005).

7.2.1. Determination of the minimum cost and maximum reliability network

The minimum cost (without any threshold on network failure rate) is determined to be 210 with a failure rate of 101 (Table 2) by using
the formulation F1(b). However, there are no multiple solutions. As shown in Table 2, the minimum network failure rate (in the absence of
cost constraint) based on formulation F1(a) is found to be 92 and is characterized by as many as 736 solutions. Of these solutions, lowest
cost of 261 is obtained for one sensor network configuration.
To get additional insights for the minimization of the network failure rate, consider Fig. 4 that shows the number of sensor networks (out
of 736 multiple optimal solutions) in which a particular variable is measured. For example, variables 1 and 2 are measured in 464 sensor
networks whereas variables 6 and 17 are not measured in any of the 736 sensor networks. Additionally, it can be seen that the variables 7,
8, 16, 18, 23 are always measured and hence may be characterized as critical measurements. Such analysis would have not been possible
by considering only one single optimal solution.

Fig. 3. Schematic of the case study.

 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2043

Table 1
Data for case study

Var, i Failure rate, i Sensor cost, Ci Number of cutsets, |Ci |

1 13 32 84
2 19 16 84
3 21 10 30
4 25 7 108
5 10 50 93
6 20 15 120
7 15 27 120
8 19 23 36
9 12 35 90
10 10 46 54
11 11 44 33
12 19 20 58
13 16 25 30
14 17 24 30
15 19 19 33
16 28 5 33
17 14 31 61
18 19 17 50
19 14 29 46
20 11 42 30
21 20 12 33
22 22 8 64
23 11 38 61
24 16 26 46

Fig. 4. Statistics of variables measured in the multiple solutions for the minimization of the nominal network failure rate.

Now, we demonstrate the determination of multiple solutions satisfying some particular conditions as discussed earlier. For this,
consider the multiple optimal solutions having minimum failure rate. Suppose that we are not interested in all these solutions but only
in the solutions where key variables 1, 12 and 20 are measured. Using the once construct in ILOG, it was found that there are 88 solutions
which measure all these three variables and still have the minimum network failure rate of 92.
Thus, we see that CP is able to determine multiple optimal solutions for these combinatorial problems and the resulting choices can
greatly aid the designer in selecting meaningful solutions.

7.2.2. Determination of minimum cost networks with thresholds on key variables

In this section, we present the results for the determination of least cost networks which need to be observable with the failure rates
of some key variables being below a threshold value (formulation F1(c)). The two cases considered are given in Table 3 along with the key
variables and their thresholds. For example, 15(3) denotes that the estimated failure rate of variable 15 should be less than three times the

Table 2
Lexicographic ordering of cost and failure rate

Primary objective Optimal Number of Secondary Optimal Number of

value solutions objective value solutions

Cost 210 1 Failure rate 101 1

Failure rate 92 736 Cost 261 1
2044 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049

Table 3
Minimum cost networks with thresholds on key variables

Case Key variables () Optimal cost Number of optimal solutions Optimal network failure rate Number of optimal solutions

I 1(2), 3(3), 9(4), 15(3), 17(2), 21(3) 231 1 100 1

II 1(2), 16(3), 22(4) 218 1 101 1

Table 4
Minimum cost robust (I) networks

Case Uncertain sensors ˆ c − small )

L ( Optimal robustness Number of optimal Optimal cost Number of optimal
solutions solutions

I 2, 14, 24 76 (92–16) 76 (infinite) 104 281 1

II 1, 2, 4, 9, 12, 14, 24 80 (92–12) 47 72 306 1
III 1, 2, 4, 6, 8, 9, 10, 12, 14, 24 82 (92–10) 0 736 261 1

failure rate of the sensor that can be used for measuring 15. As seen in Table 3, the minimum cost network for Case I is 231 and for Case II
the cost is 218. Both these cases are characterized by single sensor networks. Comparing Case I of Table 3 and the minimum cost solution
of Table 2, it can be seen that the specification of thresholds on key variables leads to an increase in the sensor network cost from 210 to
231.
Thus the proposed formulation F1(c) can be used to determine minimal cost networks which satisfy the thresholds on key variables. In
the presence of multiple solutions, additional criteria can be employed to select the best network.

7.2.3. Determination of robust minimum failure rate network

In this section, we consider the selection of robust sensor networks in the presence of uncertainties in the failure rates. In the results
presented, the critical network failure rate in Eqs. (43) and (52) is assumed to be 92 which is the optimal nominal network failure rate. For
the ease of reference, we will refer to robustness in F2 as robustness (I) and the robustness in F3 as robustness (II). We first assume that
only the uncertain sensors are known but their worst case failure rates are not known. Under these circumstances, we solve formulation
F2.

7.2.3.1. Robustness with unknown levels of uncertainty (formulation F2). In this section, we demonstrate the three categories of solutions
(related to robustness achieved) discussed earlier. The results for various scenarios are listed in Table 4. The value of L is calculated as earlier
and is shown in Table 4. First, we assume that sensors 2, 14 and 24 are uncertain (Case I, Table 4) and determine the robustness to be equal
to L. This indicates that the network failure rate is infinitely tolerant to the failure rate of these sensors indicating that these variables are
not measured. Of the 104 optimal solutions with infinite tolerance, the minimum cost is found to be 281 and is achieved by only 1 solution.
This single solution is then lexicographic optimal solution with network failure rate as the primary, robustness as the secondary and cost
as the tertiary objective. We now assume that in addition to the above sensors, the sensors for measuring variables 1, 4, 9, 12 are also
uncertain (Case II, Table 4). With the inclusion of these uncertain sensors, the robustness level decreases to 47 (see Table 4). Additionally
the minimum cost required to achieve this robustness increases to 306 as costly sensors have been used to maintain the network reliability.
We now increase the set of uncertain sensors to 10 as shown in Case III, Table 4. These set of uncertain sensors bring down the robustness
to 0. This means that all the solutions now obtained for the minimum network failure rate have zero robustness. This could have been
guessed from Fig. 4 since variable 8 was measured in all the 736 optimal solutions. The minimum cost of such network will be 261. Thus,
we see that the increase in the set of uncertain sensors decreases the robustness and that formulation F2 can be used to determine robust
sensor networks for uncertainty in various sets of variables.

7.2.3.2. Robustness with known levels of uncertainty (formulation F3). In this section, we determine minimum cost robust networks and their
realizations when the set of uncertain sensors are known along with their maximum uncertainty levels. Table 5 lists four different cases of
uncertain sensors along with the extent of their uncertainty. For example, 14(3) indicates that the uncertainty in failure rate for sensor of
variable 14 is thrice its nominal value listed in Table 1. Cases I and II have infinite robustness though the number of uncertain sensors is high
in Case II. However, it can be seen that the number of optimal solutions for Case II reduces to 72 whereas the number of optimal solutions
for Case I is 432. This shows that all the optimal solutions of Case I are not optimal solutions of Case II. Cases II, III and IV combined show
that the robustness can deteriorate as the number of uncertain sensors increases without a decrease in the number of multiple solutions
for each case. Table 5 also shows the minimum cost robust network for all these cases along with the number of realizations. As in Section
7.2.3.1, these are lexicographic optimal solutions with nominal network failure rate as the primary objective, robustness as the secondary
objective and cost as the tertiary objective.

Table 5
Minimum cost robust (II) networks

Case Uncertain sensors () Optimal robustness Number of optimal Optimal cost Number of optimal
solutions solutions

I 2 (2), 14 (3), 24(2) 92a 432 272 1

II 1(5), 2(2), 4(2), 9(3), 12(2), 14(3), 24(2) 92a 72 290 1
III 1(5), 2 (2), 4(3), 9(2), 12(4), 14 (3), 16(3), 24(2) 148 352 274 1
IV 1(5), 2 (2), 4(3), 6(4), 8(5), 9(2), 10(7), 12(4), 14 (3), 24(2) 173 64 265 1
a
Infinite robustness.
 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2045

Table 6
Pareto front between cost, network failure rate and failure rates of key variables

Tag Objectives Number of realizations

Network failure Cost Failure rates of key variables

1 16 22

A 92 269 13 28 22 1
B 92 262 75 28 22 1
C 92 261 76 28 22 1
D 95 262 13 28 22 1
E 95 255 75 28 22 1
F 95 254 76 28 22 1
G 96 243 13 28 22 1
H 96 236 75 28 22 1
I 96 235 76 28 22 1
J 97 238 13 28 22 1
K 97 231 75 28 22 1
L 97 230 76 28 22 1
M 98 232 13 28 22 1
N 98 225 75 28 22 1
O 98 224 76 28 22 1
P 100 222 13 28 22 1
Q 100 215 75 28 22 1
R 100 214 76 28 22 1
S 101 218 13 28 22 1
T 101 211 75 28 22 1
U 101 210 76 28 22 1

Thus, the formulations F2 and F3 can be used to determine globally optimal minimum cost robust sensor networks with the help
of CP.

7.2.4. Determination of pareto fronts

In this section, we determine pareto fronts (along with their realizations) between various conflicting objectives using CP. We have
considered three cases: (i) the trade-offs between the failure rates of some key variables with network cost and network failure rate, (ii)
the trade-off between robustness (I) and cost of the sensor network, and (iii) the trade-off between cost, nominal network failure rate and
robustness (II). The procedure for determining pareto front is as discussed in Section 6.2. For the visualization of the pareto fronts, in this
article, we will use the parallel coordinate system (Gala & Bagajewicz, 2006a, 2006b; Kotecha et al., 2007a) for more than two objective
functions. The visualization of pareto fronts with more than two objective including the parallel coordinate systems have been discussed
in Deb (2001).

7.2.4.1. Cost, network failure rate and failure rates of key variables. In this section, we assume that the key variables are 1, 16 and 22. We
study the trade-offs between the failure rates of these three variables along with network cost and network failure rate. There are a total
of 21 distinct pareto points. However, there are no multiple realizations for any of these pareto points. The details of these points are given
in Table 6 along with the realizations. It should be noted that we did not specify any thresholds on key variables to generate the pareto
fronts as these may restrict generation of the full pareto front, but if needed, these can also be included and the pareto points obtained in
a similar manner. Depending on the problem at hand, such scenarios can reduce the computational burden.

Table 7
Pareto front for cost and robustness (I) along with realizations with network failure rate = 92

Tag Objectives Number of realizations

Robustness Cost

Case I
A1 76 281 1
B1 20 272 1
C1 23 277 1
D1 17 265 1
E1 8 264 1
F1 5 261 1

Case II
A2 47 306 1
B2 41 290 1
C2 11 285 1
D2 7 265 1
E2 5 264 1
F2 2 261 1
G2 8 273 1

Case III
A3 0 261 1
2046 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049

Fig. 5. Parallel coordinate representation for pareto points between network failure rate, cost and the failure rates of variables 1, 16 and 22.

An interesting observation from the list of pareto points is that there is no trade-off involved in the variables 16 and 22. In all the pareto
points, they have a failure rate of 28 and 22, respectively, and are always measured. It can also be seen that the Point S corresponds to
the minimum cost network as obtained in Case II of Table 3. Fig. 5 shows the pareto points listed in Table 6 on parallel coordinate system.
The parallel coordinate system has five vertical parallel lines for each of the five objectives. The polygonal lines correspond to the various
pareto points and intersect with each of the vertical line at the corresponding objective value. It can be seen that each of the polygonal line

Fig. 6. Pareto points between robustness (I) and cost.

 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2047

Table 8
Pareto front for cost, failure rates and robustness (II) along with realizations

Tag Objectives Number of

Network failure rate Robustness (II) Cost realizations

A 92 173 265 1
B 92 186 264 1
C 92 191 261 1
D 95 173 258 1
E 95 186 257 1
F 95 191 254 1
G 96 234 235 1
H 97 233 249 1
I 97 234 230 1
J 98 224 229 1
K 98 233 224 1
L 100 236 217 1
M 100 242 214 1
N 101 237 213 1
O 101 242 210 1
P 123 211 253 1
Q 124 196 249 1
R 126 188 248 1
S 126 214 246 1
T 128 206 245 1
U 129 191 241 1
V 131 209 238 1

will also be intersecting every other polygonal line. This is due to the fact that for each pair of pareto-optimal solutions (i, j), solution i is
inferior to solution j in at least one objective and similarly solution j is inferior to solution i in at least one objective.

7.2.4.2. Cost and robustness (I). In this section, we evaluate the trade-offs between robustness as defined in formulation F2 and cost.
However, we include a constraint that the pareto points should have an optimal network failure rate of 92. The pareto fronts for all the
three cases of uncertain sensors listed in Table 4 are determined and the results listed in Table 7 are shown in Fig. 6. The pareto front
for Case I is characterized by six solutions. While solution A1 corresponds to the maximization of robustness, solution F1 corresponds to
minimization of cost. It can be seen that solution A1 corresponds to Case I of Table 4. The pareto front for Case II is characterized by seven
points as shown in Table 7. As expected, point A2 corresponds to Case II of Table 4. By comparing with Case III of Table 4, it can be easily

Fig. 7. Pareto points between network failure rate, robustness (II) and cost.
2048 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049

seen that for Case III of Table 7, there will be only one pareto point as there is zero robustness and this pareto point corresponds to the
minimum cost network as shown in Table 4.

7.2.4.3. Cost, nominal network failure rate and robustness (II). We determine the complete trade-offs between nominal network failure
rate, robustness (II) and cost of the sensor network for the uncertainty scenario of Case IV in Table 5. The pareto front is characterized
by 22 different pareto points as shown in Table 8. All the points are characterized by single realizations. From points A and B, it can be
seen that the network cost increases to provide more robustness to the uncertainty. Points H and Q show the decrease in reliability for
increasing the robustness. Also, it can be observed that the Point A corresponds to Case IV of Table 5 as this point corresponds to the
maximization of robustness (II) in lexicographic sense with the nominal network failure rate being the primary objective. Fig. 7 shows the
parallel coordinates displaying the pareto-optimal solutions.

Remark 9. We have not considered determination of pareto front between network failure rate and robustness (I) (with or without cost
objective) because the robustness can be increased to infinity by increasing the network failure rate and several artificial (corresponding to
the same sensor network) trade-off points between these two objectives can be generated. However, such issues do not arise for robustness
(II) because for a given sensor network configuration, the nominal failure rates and the worst case failure rates have unique values.

Before concluding, we would like to mention certain enhancements which may greatly improve the computational performance of CP
for typical sensor network design problems.
Very often, a problem may be solved more than once (Gala & Bagajewicz, 2006b) with or without minor modifications. For example,
one would like to determine the maximum reliable network with different reliabilities for some sensors than those considered earlier. In
such circumstances, it will be desirable to use information about infeasibilities (choice points leading to failures) that was obtained from
solving the problem at previous instances. In the results presented, we have not used such past information. For example, we solved the
minimization of network failure rate problem followed by maximization of robustness but did not use any information obtained from
solving the minimization problem. Using such information will immensely benefit in reducing the computational burden as failures will
not be repeated for every solution instance. While solving the sensor network design problems, this procedure is bound to lead to significant
benefits because the solution to any problem will be from the list of spanning trees (corresponding to observable networks).
The dynamic addition of constraints require reinitialization of the problem for adding the Type I pareto cuts. This reinitialization leads
to additional computational burden.
However, to the best of our knowledge existing CP algorithms do not provide these features and this can be a potential area for further
development.

8. Conclusion

In this article, an explicit MINLP optimization framework, based on cutsets, for the design of sensor network which minimizes the
network failure rate (equivalently maximizes the network reliability) for the non redundant case is presented. Also, two novel formulations,
corresponding to known and unknown uncertainty levels, to select networks which are robust to the uncertainties in the sensor failure
rates, have been proposed. The use of CP for solving these combinatorial problems to global optimality along with the determination
of multiple solutions is also demonstrated. These multiple solutions offer the flexibility in selecting solutions by considering any other
objective/requirement that was not included in the optimization formulation. Additionally, we have also demonstrated the use of CP to
determine the globally optimal pareto front between conflicting objectives along with their realizations. These pareto-optimal solutions
help the designer in making informed choices whereas the realizations offer additional flexibility after the selection of a pareto point.
Future work will be oriented towards enhancing these formulations by considering various other criteria in the sensor network design
problem.

Appendix A

The following algorithm is used to generate all the possible cutsets of a flowsheet. The interested reader is referred to Deo (1974) for
details about graph-theoretic concepts. This algorithm is based on the following two standard results (Deo, 1974) from graph theory: (i) the
ringsum of any two cutsets in a graph is either a third cutset or an edge-disjoint union of cutsets, and (ii) the minimal number of cutsets
needed to obtain all the cutsets is equal to the fundamental cutsets of any spanning tree.

Step 1: Determine any one spanning tree.

Step 2: Determine all the fundamental cutsets of the spanning tree. The number of cutsets will be equal to the number of branches in the
tree (unmeasured variables).
Step 3: Determine all the cutsets of the flowsheet by the ringsum of all the fundamental cutsets obtained in Step 2 with each other. Ringsum
of two cutsets is defined as the set of edges which are present in either of the two cutsets but not in both.
Step 4: Remove redundant cutsets as well as sets which are edge-disjoint union of cutsets.

Appendix B

For the sake of brevity, the counting of tree constraints is illustrated for a small flowsheet taken from literature (Meyer et al., 1994).
 synthesis plant and is characterized by six (vertices) units (including the environment node) and eight variables
Fig. A2.1 is of an ammonia
m+1 (m+1)
(edges). A total of 42 = k=3
Ck constraints can be constructed for the graph in Fig. A2.1. However, some of these tree constraints
can be redundant. Table A2.1 shows some typical valid and redundant constraints for this graph.
 P.R. Kotecha et al. / Computers and Chemical Engineering 32 (2008) 2030–2049 2049

Fig. A2.1. Simplified ammonia network.

Table A2.1
Tree constraints

E S Tree constraint Validity

4, 5, 7 4, 5, 6 x4 + x5 + x7 ≤ 2 Valid
1, 2, 8 1, 2, 3, 5 x 1 + x2 + x8 ≤ 3 Redundant
1, 2 1, 2, 3 x 1 + x2 ≤ 2 Redundant
1, 2, 3, 4, 8 1, 2, 3, 4, 5 x1 + x2 +x3 +x4 + x8 ≤ 4 Valid

For example, the constraint x1 + x2 + x8 ≤ 3 is redundant since the left hand side will be always less than or equal to 3 as the variables are
binary. The total number of remaining tree constraints after the removal of such redundant constraints for the ammonia flowsheet is 12.
For the case study presented in this article, similar redundant tree constraints removal procedure is adopted leading to 2570 constraints.

References

Ali, Y., & Narasimhan, S. (1993). Sensor network design for maximizing reliability of linear processes. AIChE Journal, 39, 820–828.
Ali, Y., & Narasimhan, S. (1995). Redundant sensor network design for linear processes. AIChE Journal, 41, 2237–2249.
Ali, Y., & Narasimhan, S. (1996). Sensor network design for maximizing reliability of bilinear processes. AIChE Journal, 42, 2563–2575.
Bagajewicz, M. J. (1997). Design and retrofit of sensor networks in process plants. AIChE Journal, 43, 2300–2306.
Bagajewicz, M., & Cabrera, E. (2002). New MILP formulation for instrumentation network design and upgrade. AIChE Journal, 48, 2271–2282.
Bagajewicz, M., & Sanchez, M. (2000). Cost-optimal design of reliable sensor networks. Computers & Chemical Engineering, 23, 1757–1762.
Bhushan, M., Narasimhan, S., & Rengaswamy, R. (2008). Robust sensor network design for fault diagnosis. Computers & Chemical Engineering, 32, 1067–1084.
Bhushan, M., & Rengaswamy, R. (2002a). Comprehensive design of a sensor network for chemical plants based on various diagnosability and reliability criteria. 1. Framework.
Industrial and Engineering Chemistry Research, 41, 1826–1839.
Bhushan, M., & Rengaswamy, R. (2002b). Comprehensive design of a sensor network for chemical plants based on various diagnosability and reliability criteria. 2. Applications.
Industrial and Engineering Chemistry Research, 41, 1840–1860.
Deb, K. (2001). Multi-objective optimization using evolutionary algorithms. New York, NY, USA: John Wiley & Sons, Inc.
Deo, N. (1974). Graph theory with applications to engineering and computer science. Upper Saddle River, NJ, USA: Prentice-Hall, Inc.
Dhillon, B. S. (1983). Reliability engineering in systems design and operation. New York: Van Nostrand Reinhold.
Diwekar, U. M. (2003). Introduction to applied optimization. Massachusetts, USA: Kluwer Academic Publisher.
Gala, M., & Bagajewicz, M. J. (2006a). Rigorous methodology for the design and upgrade of sensor networks using cutsets. Industrial and Engineering Chemistry Research, 45,
6679–6686.
Gala, M., & Bagajewicz, M. J. (2006b). Efficient procedure for the design and upgrade of sensor networks using cutsets and rigorous decomposition. Industrial and Engineering
Chemistry Research, 45, 6687–6697.
Hooker, J. (2000). Logic-based methods for optimization: Combining optimization and constraint satisfaction. New York: Wiley.
Kotecha, P. R., Bhushan, M., & Gudi, R. D. (2007a). Constraint programming based robust sensor network design. Industrial and Engineering Chemistry Research, 46, 5985–5999.
Kotecha, P. R., Bhushan, M., & Gudi, R. D. (2007b). Efficient generation of globally optimal pareto fronts using constraint programming. In 2007 AIChE Annual Meeting.
Kotecha, P. R., Bhushan, M., Gudi, R. D., & Keshari, K. M. (2008). A duality based framework for integrating reliability and precision for sensor network design. Journal of Process
Control, 18, 189–201.
Lustig, I. J., & Puget, J.-F. (2001). Program does not equal program: Constraint programming and its relationship to mathematical programming. Interfaces, 31, 9–53.
Madron, F., & Veverka, V. (1992). Optimal selection of measuring points in complex plants by linear models. AIChE Journal, 38, 227–236.
Meyer, M., Le Lann, J. M., Koehret, B., & Enjalbert, M. (1994). Optimal selection of sensor location on a complex plant, using a graph oriented approach. Computers & Chemical
Engineering, 18, 535–540.
Mhaskar, P., Gani, A., McFall, C., Christofides, P. D, & Davis, J. F. (2007). Fault-tolerant control of nonlinear process systems subject to sensor faults. AIChE Journal, 53, 654–668.
Nael, A. G., El-Farra, H., & Christofides, P. D. (2005). Fault-tolerant control of process systems using communication networks. AIChE Journal, 51, 1665–1682.
Panagiotis, J. F. D., Christofides, D., Nael, H., El-Farra, Don Clark, Kevin, R. D., et al. (2007). Smart plant operations: Vision, progress and challenges. AIChE Journal, 53, 2734–2741.
Prakash, J., Narasimhan, S., & Patwardhan, S. C. (2005). Integrating model based fault diagnosis with model predictive control. Industrial and Engineering Chemistry Research,
44, 4344–4360.
Sanchez, M. C., & Bagajewicz, M. J. (2000). On the impact of corrective maintenance in the design of sensor networks. Industrial and Engineering Chemistry Research, 39,
977–981.
Sen, S., Narasimhan, S., & Deb, K. (1998). Sensor network design of linear processes using genetic algorithms. Computers & Chemical Engineering, 22, 385–390.
Sherali, H. D. (1982). Equivalent weights for lexicographic multi-objective programs: Characterizations and computations. European Journal of Operational Research, 11,
367–379.
Solver. (2005). ILOG OPL Studio 3.5 User’s manual. Mountain View, CA: ILOG, Inc.
Tawarmalani, M., & Sahinidis, N. V. (2002). Convexification and global optimization in continuous and mixed-integer nonlinear programming: Theory, algorithms, software and
applications. Norwell, MA, USA: Kluwer Academic Publishers.
Van Hentenryck, P. (2002). Constraint and integer programming in OPL. INFORMS Journal on Computing, 14, 345–372.