Accepted Manuscript

Reallocation index based sensor network design

Suryanarayana Kolluri, Mani Bhushan

PII: S0950-4230(17)30491-6
DOI: 10.1016/j.jlp.2017.05.026
Reference: JLPP 3525

To appear in: Journal of Loss Prevention in the Process Industries

Received Date: 1 January 2017

Revised Date: 28 April 2017
Accepted Date: 29 May 2017

Please cite this article as: Kolluri, S., Bhushan, M., Reallocation index based sensor network design,
Journal of Loss Prevention in the Process Industries (2017), doi: 10.1016/j.jlp.2017.05.026.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Reallocation Index Based Sensor Network Design

Suryanarayana Kolluri, Mani Bhushan∗

PT
Department of Chemical Engineering, Indian Institute of Technology Bombay, Mumbai,
India- 400076

RI
Abstract

SC
Recently, the concept of reallocation index has been proposed in literature for
sensor network design. In this work, we extend this concept to enable base case

U
design of sensor networks for various criteria. Reallocation index indicates the
possibilities of future reallocation of sensors while retrofitting an existing design.
AN
Higher value of reallocation index indicates more possibilities of reallocation of
sensors that can be more cost effective than adding (upgrade) new sensors.
M
Reallocation index is maximized as an additional objective while performing
base case design of a sensor network. The resulting design is then optimal not
only for the base case, but also ensures that any future upgrade and reallocation
D

can be carried out effectively with less cost. The proposed sensor network design
TE

approach is illustrated on three types of sensor network design problems, namely

selecting variables to be measured to ensure: (i) robust design for reliable fault
diagnosis, (ii) observability of process variables after a change in the process
EP

layout, and (iii) estimation of important process variables with low uncertainty.
For each case, simulation based experiments are performed to demonstrate the
utility of base case designs obtained after maximization of reallocation index.
C

Keywords: Reallocation and upgrade; optimization; robust fault diagnosis;

AC

observability; estimation variance;

∗ Corresponding Author, Tel: +91-22-25767214, Fax: +91-22-25726895

Email addresses: suryanarayana.k@iitb.ac.in (Suryanarayana Kolluri),
mbhushan@iitb.ac.in (Mani Bhushan )

Preprint submitted to Journal of Loss Prevention in the Process Industries May 31, 2017
 ACCEPTED MANUSCRIPT

1. Introduction

Safety and optimality of any industrial process is critically dependent on the

PT
sensors measuring key variables in that process. Selection of variables to be
measured in a process to satisfy various technical and economic criteria is called
the sensor network design problem. Given the increased complexity of chemical

RI
processes due to tighter heat and mass integrations, systematic methods are
needed to solve the sensor network design problem. This area has thus received

SC
increasing attention from researchers in the last couple of decades.
In literature, Vaclavek and Loucka [1] first proposed algorithms for sensor
network design problem as selection of necessary measurements to ensure observ-

U
ability required for multi-component mass balance. Later Madron and Veverka

AN
[2] designed minimum cost based sensor networks while ensuring observability
of key variables using Gauss Jordan elimination method. Meyer et al. [3] pro-
posed sensor network design formulation for minimum cost networks to ensure
M
observability of key variables with minimum set of measured variables. Later,
other objectives related to reliable estimation [4, 5], and availability, resilience
D

and error detectability while performing data reconciliation [6] have also been
proposed. There is significant amount of literature in the area of sensor network
TE

design for ensuring safe process operations by enabling efficient fault diagnosis
as well. Lambert [7] used fault trees to analyze the location of sensors depend-
ing on the effect of faults on process variables. Methods for designing sensor
EP

networks for efficient fault diagnosis have also been presented when the process
is represented as a directed graph [8] and signed directed graph [9, 10]. Chang
C

et al. [11] presented a sensor network design method based on diagnostic perfor-
mance of state space model based filtering approach. Commault et al. [12] and
AC

Krysander and Frisk [13] have investigated sensor network design problem for
fault detection and diagnosis based on structural analysis of state space mod-
els. Bagajewicz [14] has presented a review of work done in the area of sensor
network design for chemical processes.
Most of the work in the area of sensor network design is for base case design

2
 ACCEPTED MANUSCRIPT

scenarios when the process does not have any existing sensors. However, over
a period of time, the process usually goes through significant changes since the
operating conditions of the process may change. Additionally new process util-

PT
ities may be added or existing ones replaced or removed from the process. Due
to these changes, a sensor network design which was optimal to begin with, may

RI
no longer be optimal or even adequate for ensuring safe operations for the exist-
ing process. Hence, for existing processes a relevant problem is sensor network
retrofit, i.e. modifying an existing sensor network to meet specific requirements

SC
or improving the quality of information available about the process. Given the
advances in instrumentation hardware and software technologies including wire-

U
less sensing, adding new sensors in the process and reallocating existing sensors
is increasingly possible. Recognizing these possibilities, Bagajewicz and Sánchez
AN
[15] proposed an optimization formulation for maximizing precision in process
plants by reallocating existing sensors and then implementing upgrade strategy
by adding new sensors at minimum cost to achieve maximum precision. In their
M
formulation, new sensors can be added (termed upgrade) or existing sensors can
be transferred to other variables (termed reallocation). Subsequently, Bhushan
D

et al. [16] used the upgrade and reallocation idea of Bagajewicz and Sánchez
TE

[15] to retrofit an existing sensor network from a fault diagnostic perspective.

Recently, Kolluri and Bhushan [17] extended these ideas and proposed a generic
formulation for audit of sensor networks which performs retrofit of an existing
EP

design as well as identifies non-contributory sensors i.e. sensors that are non-
informative for the problem at hand. In contrast to the above approaches that
involve modifying an existing network to account for process changes, Bo and
C

Bing-Zhen [18] took a different approach to the problem of retrofit. In particu-

AC

lar, they proposed a base case design methodology that ensures observability of
all variables for a linear mass balance process, even when the process undergoes
topological change (such as edge deletion or addition). This was ensured by
installing additional sensors in the base case design based on consideration of
all possible changes in the process topology in the future. However, the num-
ber of possible changes can be combinatorially large and thus it is not easy to

3
 ACCEPTED MANUSCRIPT

extend this approach to larger systems or other types of sensor network design
problems.
From the above literature survey, it can be seen that while most of the

PT
work is in the area of base case design of sensor networks, there is some work
for retrofitting existing sensor networks as well. However, there are very few

RI
approaches that integrate the two problems. In particular, Kolluri and Bhushan
[19] presented a design approach wherein base case sensor networks that ensured
future retrofit with low cost were obtained. This was based on a novel concept of

SC
reallocation index. The main idea in the approach was to maximize possibilities
of reallocation of sensors in future while obtaining base case design, without

U
compromising on the optimality of the base case design. However, their work was
preliminary in nature and was applied only to sensor network design for reliable
AN
fault diagnosis. In this work, we elaborate the notion of reallocation index and
apply it to several criteria, namely related to robust design for reliable fault
diagnosis, observability of variables in presence of changes in process topology
M
in a steady state flow system, and estimation of key variables for steady state
linear systems using data reconciliation. The proposed approach is generic in
D

nature and can be used for designing sensor networks for various other criteria
TE

as well.
The rest of the paper is structured as follows: in section 2, we briefly summa-
rize the previous relevant work in the area of sensor network design and retrofit
EP

for fault diagnosis, observability of variables for steady state flow systems, and
data reconciliation based accurate estimation of key variables for steady state,
linear systems. We also summarize the generic optimization formulation for
C

retrofitting an existing sensor network as this formulation is later used while

AC

demonstrating the utility of our proposed base case design approach. In sec-
tion 3 we present the concept of reallocation index and discuss its computation
for non-redundant and redundant sensor networks. In section 4 we present the
specific optimization formulations that enable reallocation index to be incor-
porated in base case design of sensor networks for the criteria summarized in
section 2. In section 5 we apply our proposed sensor network design approach

4
 ACCEPTED MANUSCRIPT

to standard case studies taken from literature to design sensor networks for
these three criteria. For each of these criteria, we compare the performance of
the network obtained using our proposed approach with the performance of a

PT
network obtained from literature. The paper is then concluded in section 6.

RI
2. Previous Relevant Work

In this section, we first briefly summarize three criteria for sensor network

SC
design. The proposed reallocation index based sensor network design approach
will be illustrated for these criteria. In section 2.4, we summarize a generic
constrained optimization based retrofit formulation available in literature, that

U
is used in later sections.

AN
2.1. Reliable Fault Diagnosis Related Criteria

Raghuraj et al. [8] presented a sensor network design approach which en-
M
sured that the selected sensors could detect faults as well as identify the root
cause of the fault, thereby reducing the risk of untoward incidents remaining
undetected in the process. Towards this end, they modeled the process as a di-
D

rected graph (DG). By propagation from causal root nodes (equivalently referred
TE

to as faults) in the DG, set of variables affected by these faults was identified.
The optimal sensor network was then obtained by ensuring that at least one
variable affected by each fault is measured and also the pattern of effects on the
EP

measured variables are different for different faults. Bhushan and Rengaswamy
[9] extended the work to the case when process is modeled as a signed digraph
C

(SDG). Bhushan and Rengaswamy [20] extended the design methodology and
presented optimization formulations for sensor network design based on the con-
AC

cept of unreliability of detection of a fault which is defined as the probability

of that fault occurring and remaining undetected due to simultaneous failure of
sensors affected by that fault. For fault i, the unreliability of detection is [20]:

m
Y
Ui = fi (sj )(Bij xj ) , i = 1, 2, ..., p (1)
j=1

5
 ACCEPTED MANUSCRIPT

where fi is the probability of occurrence of fault i, sj is the failure probability

of sensor placed on variable j, m is the number of variables and p is the number
of faults in the process, xj is the number of sensors used for measuring variable

PT
j, and matrix B is the fault-variable bipartite matrix such that Bij = 1 if fault
i affects variable j and is 0 otherwise. A fault with low Ui has a low probability

RI
of remaining undetected and hence a higher likelihood of being contained.
The matrix B can be obtained by analyzing the process DG or SDG [9].
The procedure for obtaining this matrix is illustrated in Appendix A using an

SC
example of a reactor. The following example illustrates the concept of unrelia-
bility of detection based on the reactor example.

U
Example 1 : Consider the fault-variable bipartite graph as shown in Figure 1 for
AN
the CSTR example considered in Appendix A. This is a pictorial representation
of the fault-variable bipartite matrix B whose construction is illustrated in Ap-
pendix A. The probabilities of occurrences of the six faults and the failures of
M
sensors available to measure the four variables are also listed in figure 1 next to
the nodes representing these variables and faults. Now consider that variables
D

S1 , S4 are measured while S2 , S3 are not measured. Then the unreliabilities of

TE

detection of various faults are computed as follows:

Fault F1 : U1 = f1 × s1×1
1 × s1×0
2 × s1×0
3 × s1×1
4 = 10−7
EP

Fault F2 : U2 = f2 × s1×1
1 × s1×0
2 × s1×0
3 × s1×1
4 = 10−6

Fault F3 : U3 = f3 × s1×1
1 × s1×0
2 × s1×0
3 × s1×1
4 = 10−8

Fault F4 : U4 = f4 × s1×1 × s1×0 × s0×0 × s0×1 = 10−5

C

1 2 3 4

Fault F5 : U5 = f5 × s1×1
1 × s0×0
2 × s0×0
3 × s0×1
4 = 10−6
AC

Fault F6 : U6 = f6 × s0×1
1 × s1×0
2 × s0×0
3 × s0×1
4 = 10−3

2.1.1. Design of Sensor Networks for Reliable Fault Diagnosis

Bhushan and Rengaswamy [20] proposed sensor network design for reliable

6
 ACCEPTED MANUSCRIPT

(0.001) (0.001) (0.001) (0.01)

S1 S2 S3 S4

PT
RI
F1 F2 F3 F4 F5 F6
[0.01] [0.1] [0.001] [0.01] [0.001] [0.001]

SC
Figure 1: Bipartite graph of CSTR example in Appendix A with probabilities

fault diagnosis as a mixed-integer linear programming (MILP) optimization

U
problem that involved minimization of the system unreliability subject to con-
AN
straints (such as related to cost) on the sensor network. The system unreliability
was defined as the maximum unreliability of detection amongst all faults. This
is based on the philosophy that a safe process should have a low probability of
M
faults occurring and remaining undetected. Narasimhan and Rengaswamy [21]
have proposed methods to quantify economic benefit of fault detection and diag-
D

nosis. Subsequently, Bhushan et al. [22] extended the optimization formulation

to deal with cases where some of the data (various probabilities) and the process
TE

model (signed digraph) may not be accurately known. The optimization formu-
lation they proposed ensured selection of a design that was robust to changes in
the uncertain data. The optimization formulation given in section 4 later is an
EP

extension of the work of Bhushan et al. [22] that considers robustness to some
sensor failure probabilities as well as some fault occurrence probabilities.
C

2.2. Observability via Flexibility

AC

A variable is said to be observable if it can be estimated by using the mea-

surements and steady state process model [25]. Thus, a measured variable is
trivially observable, while an unmeasured variable is observable if it can be es-
timated using the measured variables and its relationships with the measured
variables. Bo and Bing-Zhen [18] introduced the concept of flexibility of sensor

7
 ACCEPTED MANUSCRIPT

networks for linear flow processes (processes involving only overall mass flow
balances) using ideas from graph theory. A sensor network with higher flexibil-
ity has a higher likelihood of ensuring observability of all variables when new

PT
edges are added in the process graph. These new edges correspond to new pipe-
connections between various units in the process. The notion of flexibility is

RI
explained using Example 2 next and a formal definition is provided after the
example.

SC
Example 2 : Consider the process graph of an ammonia process [18] as shown
in Figure 2.

U
5

8
AN 7 4
M
6 5
1 6 4
D

1 3
TE

2 3
2
EP

Figure 2: Process graph of ammonia process

In the figure, the nodes represent process units while the edges represent flows
C

between the process units. Under the steady state assumption, mass balances
AC

involving the flows can be written across every unit and can be used to estimate
unmeasured flow rates. An edge (flow rate) is said to be observable if it can be
estimated based on the available measurements and the process model. In this
process, there are a = 6 nodes and m = 8 edges. Consider the following two
different sensor networks for illustration:

8
 ACCEPTED MANUSCRIPT

1. Measured variables are flow-rates of streams [5, 6, 7, 8]: This network can
ensure observability for new edges added between any of the node pairs
(1, 6), (2, 6), (3, 6), (4, 6) and (5, 6). Note that for some of these node

PT
pairs, such as (1, 6) an edge already exists and an additional edge will
lead to parallel edges between these nodes.

RI
2. Measured variables are flow-rates of streams [3, 5, 6, 8]: This network can
ensure observability for new edges added between any of the node pairs

SC
(1, 4), (1, 5), (1, 6), (2, 4), (2, 5), (2, 6), (3, 4), (3, 5) and (3, 6).

Thus, the second sensor network has more flexibility than the first network.
It should also be noted that both the networks ensure observability of all the

U
variables in the current process graph. Formally, for a given sensor network, the

AN
flexibility, labeled Γ, is defined as [18]:

q
Γ=1− a

(2)
2
M
a


where q is number of unmeasured paths in the process and 2 is the total
number of node pair candidates. The flexibilities of the two sensor networks
D

considered for Figure 2 are 0.333 and 0.6, respectively.

TE

Bo and Bingzhen [18] proposed to select a sensor network that maximized

flexibility while ensuring observability of all variables.
EP

2.3. Data Reconciliation based Accurate Estimation

Bagajewicz and Cabrera [23] presented a sensor network design approach

for linear systems that ensured low variance estimation of key variables since
C

the variance in estimates have a bearing on the economic performance of the

AC

process [24]. In particular, Bagajewicz and Cabrera [23] proposed a mixed

integer linear programming (MILP) optimization formulation for selecting a
sensor network that minimized total cost of the selected sensors while ensuring
that the variances in estimates of the variables of interest would be less than user
specified threshold variances. The variances of estimates of process variables
were computed using data reconciliation approach. In the current work, we

9
 ACCEPTED MANUSCRIPT

apply the reallocation index based sensor network design for the problem of
precise estimation of key variables in section 4.3.

PT
2.4. Retrofit of Sensor Networks

Bagajewicz and Sánchez [15] proposed optimization formulations for reallo-

RI
cation and upgrade of sensor networks for improving precision of the estimated
parameters. This idea was implemented by Bhushan et al. [16] for retrofitting
sensors from a fault diagnostic perspective. Kolluri and Bhushan [17] extended

SC
this idea and proposed generic formulation for audit of sensor networks which
performs retrofit as well as finds non-contributory sensors for the problem at

U
hand. The generic formulation for reallocation and upgrade of sensor networks
is (adapted from Formulation I, Kolluri and Bhushan [17]) :
Formulation I : AN
min f (x) (3)
M
subject to:

gL ≤ g(x) ≤ gU (4)
D

X XX
cj qj + ht,r ut,r ≤ C ∗ (5)
j∈SS t∈T r∈Rt
TE

X
xj = qj + x∗j − uj,r ; j ∈ ((R ∩ SS) \ T ) (6)
r∈Rj
X
xj = qj + x∗j + ut,j ; j ∈ ((R ∩ SS) \ T ) (7)
EP

t∈Tj
X X
xj = qj + x∗j + ut,j − uj,r ; j ∈ (T ∩ R ∩ SS) (8)
t∈Tj r∈Rj
C

xj = qj + x∗j ; j ∈ (SS \ (T ∪ R)) (9)

AC

X
uj,r ≤ x∗j ; j ∈ T (10)
r∈Tj

xj , qj , ut,r ∈ Z+ (11)

In the above formulation, the sensor network retrofit relevant decision variables
are:

10
 ACCEPTED MANUSCRIPT

qj : number of new sensors to be placed on variable j,

ut,r : number of existing sensors reallocated from variable t to r.
Further, x∗j is the number of existing sensors measuring variable j, and xj is the

PT
net number of sensors measuring variable j as a result of retrofit. Equations
6 to 9 relate xj to the existing, new and reallocated sensors on variable j. In

RI
these equations, T is the set of variables whose sensors can be reallocated to
other variables, R is the set of variables which can receive sensors from other
variables, and SS is set of variables where new sensors can be installed. Further,

SC
Rj is the set of variables which can receive sensors from variable j, and Tj is the
set of variables which can transfer sensors to variable j. These sets are required

U
since transfer of sensors only between compatible pairs of variables is allowed.
The sets R, T, and SS are thus user defined.
AN
In Formulation I, f (x) is the objective function that is to be minimized while
g(x) denotes the performance related constraints with upper and lower bounds
being gU and gL . The quantities f (x), g(x), gU , gL , depend on the specific
M
criteria being considered for sensor network retrofit. Further, cj is the cost of a
new sensor for measuring variable j, ht,r is the cost of reallocating a sensor from
D

variable t to r and C ∗ is the total cost available for performing retrofit. Thus,
TE

constraint (5) ensures that the cost incurred in retrofitting the sensor network
does not exceed the total available cost.
Formulation I is used later in our work for demonstrating the utility of the
EP

proposed reallocation index based sensor network design approach. We next

present the concept of reallocation index.
C

3. Reallocation Index
AC

For a given sensor network, reallocation index (RI ) is defined as the total
number of variables to which sensors can be transferred (reallocated). Under the
assumption that it is less expensive to reallocate existing sensors than to install
new sensors, RI is a measure of options that the designer would have in future
for retrofitting the sensor network. Thus, we propose that RI can be maximized

11
 ACCEPTED MANUSCRIPT

as an additional objective in the sensor network design problem solved at the

design stage. It may be possible to maximize this additional objective without
affecting the primary design objective as often the primary design problem has

PT
multiple optimal solutions. Amongst these multiple solutions, the solution that
maximizes RI can be selected.

RI
To compute the reallocation index, it is assumed that information about
compatible pairs of variables between which sensors can be reallocated, is avail-
able. Let this information be stored in a binary matrix, labeled reallocation

SC
compatibility matrix (RCM) M of size m × m with m being the number of
variables in the process, such that

U

 1 if sensor can be transferred from variable j to k
Mjk = (12)
 0 otherwise AN
The RCM will in general be symmetric but this will be process specific. In
M
this work, symmetry is not assumed. Depending on whether hardware redun-
dancy (more than one sensor measuring a variable) is allowed or not, RI can be
computed in the following ways:
D

1. Reallocation index when hardware redundancy is allowed:

TE

Reallocation index can be computed as [19]:

m m
!
X X
RI = nj Mjk (13)
EP

j=1 k=1

where the binary variable nj = 1 if j th variable is measured (irrespective

C

of the number of sensors placed on variable j) and is 0 if j th variable is

not measured. From Equation 13 it can be seen that a transfer of a sensor
AC

from variable j to variable k is considered possible, and hence contributes

to RI , irrespective of the number of existing sensors on variable k. The
calculation of RI is illustrated with an example next.

Example 3 : Let us consider four variables [S1, S2, S3, S4] and correspond-

12
 ACCEPTED MANUSCRIPT

ing reallocation compatibility matrix (RCM) given in Equation 14.

PT
S1 S2 S3 S4
 
S1  0 1 0 1 
 

RI
S2 
 
1 0 1 1 
M= 


 (14)
S3 
 0 1 0 0 


SC
 
S4 1 1 0 0

Assuming [S2, S4] to be measured with a sensor each, we get n2 = n4 = 1

U
while n1 = n3 = 0. The value of RI can then be obtained by noting that
X4 X 4
M2k = 3 and M4k = 2. Thus, RI = 3 + 2 = 5.
k=1 k=1 AN
2. Reallocation index when hardware redundancy is not allowed:
In some situations, due to logistical or economic reasons, multiple sensors
M
to measure a single variable are not feasible. We propose a modified
definition of reallocation index for cases when hardware redundancy is
D

not allowed as:

m m
!
TE

X X
RI = nj Mjk (1 − nk ) (15)
j=1 k=1

In the above equation, if nj = nk = 1, then both variables j and k are mea-

EP

sured and reallocation is not possible from variable j to k as that would

lead to more than one sensor on variable k and hardware redundancy is
not allowed. The term (1−nk ) inside the bracket on the right hand side of
C

Equation 15 ensures that possibility of reallocation to a measured variable

AC

is not considered while computing the reallocation index.

Example 4 : Consider same scenario as Example 3 but with hardware

redundancy not allowed. The value of RI then turns out to be 3.

In the next section, we propose sensor network design formulations that

13
 ACCEPTED MANUSCRIPT

incorporate maximization of reallocation index in the objective function.

4. Reallocation Index based Sensor Networks

PT
In this section we propose sensor network design formulations incorporating
RI for the following three types of criteria:

RI
1. Robust design for reliable fault diagnosis,
2. Ensuring observability of all variables in a linear, steady state, mass flow

SC
process, and
3. Accurate estimation of variables using data reconciliation in a linear,
steady state process.

U
4.1. Reallocation Index based Robust Sensor Networks for Reliable Fault Diag-
nosis
AN
In the current section we incorporate the idea of reallocation index in the
M
base case design of robust sensor networks for reliable fault diagnosis. Bhushan
et al. [22] proposed robust sensor network design formulation for fault diagnosis
assuming the presence of uncertainty in some of the available probability data.
D

The resulting design was robust in the sense that even if the uncertain probabil-
TE

ity values increased, the overall system unreliability would not change. We now
propose to incorporate reallocation index in the design formulation presented
by [22]. This results in Formulation II presented next:
EP

Formulation II :

min [β1 U − β2 Φ − β3 RI − cs ] (16)

C

subject to,
AC

U ≥ log(Ui ), ∀i ∈ I \ (If ∪ Is ) (17)

U = log(Ui ) + φi , ∀i ∈ (If ∪ Is ) (18)

m m
!
X X
RI = nj Mjk (19)
j=1 k=1

nj ≤ xj , j = 1, 2, ..., m (20)

14
 ACCEPTED MANUSCRIPT

φ∗i = φ∗f i , ∀i ∈ (If \ Is ) (21)

φ∗i = φ∗si , ∀i ∈ (Is \ If ) (22)

PT
φ∗i = φ∗f i + φ∗si , ∀i ∈ (If ∩ Is ) (23)
X
where φ∗si = − Bij (log sj )xj , ∀i ∈ Is (24)
j∈Js

RI
P zi ≥ φi − φ∗i , ∀i ∈ (If ∪ Is ) (25)

P (zi − 1) ≤ φi − φ∗i , ∀i ∈ (If ∪ Is ) (26)

SC
Φ ≤ φi + φ∗ zi , ∀i ∈ (If ∪ Is ) (27)

Φ ≤ φ∗ (28)
m

U
X
cj xj + cs = C ∗ (29)
j=1

nj ∈ {0, 1} ; xj ∈ Z+ ; U ∈ R−
AN (30)

(cs , Φ, φi , φ∗si , φ∗i ) ∈ R+ ; zi ∈ {0, 1} (31)

M
In the above formulation, xj , j = 1, 2, ..., m are the main decision variables
indicating the number of sensors used to measure variable j with m being the
D

total number of variables in the process. Further, the following sets of variables
TE

and faults are defined:

1. Js : set of indices of variables that can be measured by sensors with un-

certain failure probabilities.
EP

2. I: set of indices of all the faults under consideration. For a process with
p faults, I will be the set of indices of these p faults if only fault observ-
C

ability is being considered. However, if single fault resolution is also being

considered, then I will consist of indices of up to an additional p(p − 1)/2
AC

pseudo faults that correspond to pairwise resolution between the original

p faults [22].
3. If : set of indices of faults whose occurrence probabilities are uncertain.
This may again include pseudo faults apart from the original faults [22].
4. Is : set of indices of faults that affect variables belonging to set Js .

15
 ACCEPTED MANUSCRIPT

In Formulation II, the objective function (Equation 16) consists of four

terms: the first, second and fourth terms are taken from Bhushan et al. [22]
while the third term corresponds to RI proposed in this article. The various

PT
terms in the objective function are briefly described next:

1. The first term U is the primary objective, known as system unreliabil-

RI
ity of detection. It is defined as the maximum unreliability of detection
amongst all faults. This is captured by constraints 17 and 18. Ui in these

SC
constraints is the unreliability of detection of fault i (Equation 1). Con-
straint 17 is for the set of faults without any uncertainty, while constraint
18 is for the set of uncertain faults. The non-negative slacks φi in these

U
constraints ensure robustness of the system unreliability with respect to

AN
the unreliabilities of the uncertain faults. Note that the use of log(Ui )
ensures that the formulation is linear in terms of the decision variables xj
[22].
M
2. The second term in the objective function (Equation 16) corresponds to
the maximization of Φ that is defined as the system robustness. It is
D

defined as the minimum value amongst all the slacks φi (constraint 27)
that are below their maximum meaningful values φ∗i . In the worst case
TE

uncertain probabilities can increase up to 1 but not beyond that. Thus,

there is an upper limit on the meaningful value of slack for every uncertain
fault. Maximum meaningful slack φ∗i is computed for uncertain fault i
EP

using constraints 21 - 24. If φi is greater than φ∗i , then overall system

robustness Φ will not be affected by φi . Constraints 25 - 27 with P a large
C

positive number, ensure this by introducing a binary variable zi that takes

a value 1 if φi ≥ φ∗i and is 0 otherwise. Constraint 27 provides an upper
AC

bound φ∗ on overall system robustness Φ and will be active in the unlikely

case of all the individual slacks being above their maximum meaningful
values. Note that the quantities φ∗ and φ∗f i can be precomputed based on
the given probability data. The latter is used to compute the maximum
meaningful value of slack for uncertain fault i. Details on computation of

16
 ACCEPTED MANUSCRIPT

these numbers are available in [22].

3. The third term in the objective function (Equation 16) corresponds to
the maximization of reallocation index RI as proposed in this work in

PT
section 3. Equations 19 and 20 enable computation of RI as a function of
the sensor network. Note that in this formulation hardware redundancy

RI
is permitted. In Constraint 20, binary variable nj can take a value 1 if
variable j is measured and is 0 if variable j is not measured. In the former
case, nj ends up assuming the value 1 since increasing nj leads to increase

SC
in RI which is being maximized in the objective function.
4. The fourth term in the objective function corresponds to the maximization

U
of slack variable cs which is the cost saved and is defined as the total
available cost (C ∗ ) minus cost used for installing sensors for satisfying
AN
primary objective U . This is captured by constraint 29.

The constants β1 , β2 and β3 in the objective function are non-negative weighting

M
coefficients giving relative weights to various objectives. Similar to [22], we
choose them so as to ensure lexicographic optimization, i.e. consideration of
D

the 4 objectives in decreasing order of priority. The procedure used to calculate

these weighting coefficients is similar to the one presented in detail in [22].
TE

Formulation II is a mixed integer linear programming (MILP) problem and is

applied to the Tennessee Eastman case study in section 5.1 to demonstrate its
applicability.
EP

4.2. Reallocation Index based Sensor Networks for Observability of all Variables

In this section we propose a reallocation index based sensor network design

C

formulation for ensuring observability of all process variables in a steady state,

AC

overall mass balance process with m variables (streams or edges) and a nodes
(process units). The proposed optimization formulation that maximizes reallo-
cation index while ensuring observability of all variables is given next.

17
 ACCEPTED MANUSCRIPT

Formulation III:

max RI (32)

PT
such that,
Y
(1 − xj ) = 0 , i = 1, 2, · · · , l (33)

RI
j∈Li
Xm
(1 − xj ) = m − k (34)
j=1

SC
m m
!
X X
RI = xj Mjk (1 − xk ) (35)
j=1 k=1

xj ∈ {0, 1} , j = 1, 2, · · · , m

U
(36)

AN
Formulation III is an integer, nonlinear programming problem with maximiza-
tion of reallocation index being the objective function. The binary decision
variables xj indicate absence or presence of a sensor on variable j as:
M

 1 if variable j is measured
xj = (37)
if variable j is unmeasured
D

 0

The various constraints are defined next:

TE

• Constraint 33 ensures that there are no cycles in the process involving only
unmeasured variables. Here l is the total number of cycles in the process
EP

with length not more than m − k where k is the number of sensors to be

placed in the process. Li is the set of variables in the ith cycle. Ensuring
that the unmeasured variables don’t form a cycle ensures observability of
C

all the variables [4].

AC

• Constraint 34 specifies the number of unmeasured variables in the process

or equivalently specifies the number of measured variables in the process
to be k.

• Equation 35 is the definition of reallocation index. Note that hardware

redundancy is not allowed in this particular design problem.

18
 ACCEPTED MANUSCRIPT

The utility of Formulation III in selecting sensor networks that can be retrofitted
in a cost effective manner is demonstrated in section 5.2 on a crude distillation
unit case study.

PT
4.3. Reallocation Index based Sensor Networks for Accurate Estimation of Key
Variables

RI
The observability formulation presented in 4.2 treats observability as a yes-
no condition, i.e. a variable is either observable or not. However, due to the

SC
inevitable noise present in sensor measurements, the true values of variables
can only be estimated with some uncertainty (or variance). Bagajewicz and

U
Cabrera [23] proposed a minimum cost sensor network design approach that
minimizes the cost of the selected sensors while satisfying precision requirements
AN
on key variables. In particular, the precision of a variable is quantified by
the variance of its estimate. Larger variance implies lower precision and vice-
versa. For a linear, steady process, we extend the formulation of Bagajewicz
M
and Cabrera [23] to incorporate reallocation index as an additional objective in
the optimization formulation. The resulting formulation is presented next:
D

Formulation IV:
TE

 
Xm
min λ  cj xj  − RI (38)
j=1

subject to
EP

∗
Ŝjj ≤ Sjj , j ∈ Mp (39)

Ŝ = S − SAT P AS (40)
C

P (ASAT ) = I (41)
AC

xj + yj = 1, j = 1, 2, ..., m (42)
m m
!
X X
RI = xj Mjk (1 − xk ) (43)
j=1 k=1

xj , yj ∈ {0, 1}, j = 1, 2, ..., m (44)

19
 ACCEPTED MANUSCRIPT

In the above formulation, m is the number of variables in the process, Ŝ is the

covariance matrix of reconciled estimates, P is the inverse of ASAT and has
been included to avoid symbolic inversion of ASAT . Further, S is the given

PT
sensor noise variance matrix. It is a diagonal matrix with the diagonal entry j
computed as: Sjj = σj2 xj + Lyj . Here σj2 is the variance of sensor that can be

RI
used for measuring variable j while L is a large positive number approximating
the variance when no real sensor is placed on variable j. Accordingly, xj cor-
responds to choice of a real sensor on variable j while yj corresponds to a fake

SC
sensor (with a large variance). Fake sensor is used in the formulation to be able
to work with constant size S and A matrices irrespective of the resulting sensor

U
network [23]. Further, A is the adjacency matrix of the underlying mass balance
process [23]. The various constraints and objective functions in Formulation IV
are briefly discussed next:
AN
• Equation 38 is the objective function consisting of two terms. The first
M
term in the objective function is the total cost of selected sensors while
second term is the reallocation index. The positive coefficient λ ensures
D

lexicographic optimization with cost being the primary objective. Proce-

dure for calculating λ is the same as that given in [22].
TE

• In constraint 39, Ŝjj is the reconciled variance of variable j that should be

∗
less than the specified threshold variance Sjj . Mp is the set of variables
EP

with such precision specifications.

• Equations 40 and 41 relate the covariance matrix Ŝ of reconciled estimates

to the sensor variance matrix S and the adjacency matrix A. Here I is
C

the identity matrix.

AC

• Constraint 42 ensure that exactly one of the real or fake sensor on every
variable is selected.

• Equation 43 is the definition of RI when hardware redundancy is not

allowed, as in this problem.

20
 ACCEPTED MANUSCRIPT

• Constraint 44 specifies that variables xj , yj are binary variables.

Formulation IV is a mixed integer nonlinear programming (MINLP) op-

PT
timization problem. It can be noted that letting Mp include all the process
variables and specifying the threshold variances to be less than L for all vari-
ables, will ensure observability of all variables in the process. Formulation IV is

RI
used to design a sensor network for a topping unit in section 5.3 to demonstrate
its applicability.

SC
5. Case Studies

U
In this section, we now present three case studies each demonstrating the
utility of the reallocation index based proposed design methodology on the three
AN
types of sensor network design problems illustrated in this article.

5.1. Reliable Fault Diagnosis: Tennessee Eastman Process

M
The proposed optimization Formulation II incorporating reallocation index
for sensor network design for reliable fault diagnosis is applied to the Tennessee
D

Eastman (TE) process [26]. This case study has been widely used by the process
systems engineering community in the area of fault diagnosis. Its flowsheet is
TE

presented in Figure 3. The process involves five major unit operations, namely:
an exothermic two phase reactor, product condenser, flash separator, reboiled
EP

stripper and a recycle compressor. The process produces products G and H

from reactants A, C, D and E. Additionally, a byproduct F is also produced
and an inert species B is also present [26]. The list of streams (Table B.15) and
C

indicators (Table B.16) of the TE process in Figure 3 is given in Appendix B.

Maurya et al. [27] presented and analyzed a signed digraph model of the TE
AC

process. In particular, they considered 33 faults and 50 variables and generated

an initial response table that lists the signs of the responses (+,-,0 or indetermi-
nate) on these variables corresponding to the 33 faults. Bhushan et al. [22] used
this information to generate the effect of pseudo-faults corresponding to reso-
lution between the original faults. The fault-variable bipartite matrix B used

21
 ACCEPTED MANUSCRIPT

by them thus involved the original as well as the pseudo faults. We use this
matrix B in the current work as well. The data for sensor failure probabilities,
fault occurrence probabilities, sensor costs is also taken from Bhushan et al.

PT
[22] and presented in Tables 1, 2 and 3 respectively. Further, Table 4 lists the
allowed reallocations in this process and the corresponding reallocation costs

RI
[16]. The allowed reallocations shown in this Table are the non-zero elements
of the reallocation compatibility matrix M (Equation 12) that will be used in
the computation of RI .

SC
Sensor networks resulting from the proposed formulation (Formulation II)
for various values of available cost C ∗ are tabulated in Table 5. TOMLAB
CPLEX operating in Matlab

U
R R
version 7.4 is used as the MILP solver. For
a given C ∗ , Table 5 lists sensor selected by Formulation II and also U, Φ, RI
AN
and cs values. In the last column in this table, the number in bracket indicates
hardware redundancy, i.e. the number of sensors selected for placement on a
M
FI

FI
•
8 9 Purge
FI
D

A
1 JI
Compressor
CWS XA

XB
FI
PI A
D
2 TI
LI
N
XC
TE

A XD
Condenser CWR
7 L
FI
SC
13 Y
XE

E
3 Z XF
PI E
XG
TI R
LI
XH
5
EP

PI
FI
CWS
10
•

XA A
N
XB
A
XC L
6 TI
Stripper Vap / liq
Separator
Y CWR
C

XD FI
Z 12 TI
XE E
R TI
XF
A XD
AC

Reactor N
FI XE
A
LI
Stm L XF
Y
XG
Cond Z
E XH
FI R

FI Product
•
4 11
C

Figure 3: Tennessee Eastman process

22
 ACCEPTED MANUSCRIPT

Table 1: Faults for TE process

Tag Fault var log10 fi Tag Fault var log10 fi
F1 F1+ -2 F18 V LrV+Pbias -2
F2+ V Ls+

PT
F2 -2 F19 Xmbias -2
F3 F3+ -2 F20 V Ls+Xmset -2
F4 F4+ -2 F21 V Ls+V Pbias -2
F5 F8+ -2 F22 V Lp+Xmbias -2

RI
F6 F9+ -2 F23 V Lp+Xmset -2
F7 Tr+ -1 F24 V Lp+V Pbias -2
−
F8 Cd− -2 F25 V LrXm bias
-2

SC
F9 F1− -2 F26 −
V LrXm set
-2
F10 F2− -2 F27 V LrV−Pbias -2
F11 F3− -2 F28 V Ls−Xmbias -2
F12 F4− -2 F29 V Ls−Xmset -2

U
F13 F8− -2 F30 V Ls−V Pbias -2
F14 F9− -2 F31 V Lp−Xmbias -2
F15
F16
F17
Tr−
+
V LrXm
+
V LrXm
bias
AN
-1
-1
-2
F32
F33
V Lp−
V Lp−
Xmset
V Pbias
-2
-2
set
M
Table 2: Sensor data for TE process
Tag Var log10 sj Tag Var log10 sj Tag Var log10 sj Tag Var log10 sj
1 Pr -3 14 YA,7 -3 27 YF,8 -3 40 XG,r -3
D

2 Ps -3 15 YB,7 -3 28 YG,8 -3 41 XH,r -3

3 Pm -3 16 YC,7 -3 29 YH,8 -3 42 F10 -3
4 F6 -3 17 YD,7 -3 30 XD,10 -3 43 F11 -3
TE

5 YA,6 -3 18 YE,7 -3 31 XE,10 -3 44 Ts -2

6 YB,6 -3 19 YF,7 -3 32 XF,10 -3 45 V LrXm -2
7 YC,6 -3 20 YG,7 -3 33 XG,10 -3 46 V Lre -4
8 YD,6 -3 21 YH,7 -3 34 XH,10 -3 47 V LsXm -2
EP

9 YE,6 -3 22 YA,8 -3 35 XG,11 -3 48 V Lse -4

10 YF,6 -3 23 YB,8 -3 36 XH,11 -3 49 V LpXm -2
11 YG,6 -3 24 YC,8 -3 37 XD,r -3 50 V Lpe -4
12 YH,6 -3 25 YD,8 -3 38 XE,r -3
C

13 F7 -3 26 YE,8 -3 39 XF,r -3
AC

particular variable. For example, 45(2) indicates 2 sensors are placed on variable
45. As expected, as the available cost C ∗ increases, the primary objective
namely the system unreliability of detection U decreases or remains the same.
Due to the lexicographic nature of the formulation, for an optimal U the best
possible value of Φ is obtained. Similarly, the other two objectives RI and cs are

23
 ACCEPTED MANUSCRIPT

Table 3: Sensor cost for TE process

Cost Variables
100 1,2,3,46,48,50

PT
150 45,47,49
200 42,43
300 4,13
500 44

RI
700 30,31,32,33,34,35,36,37,38,39,40,41
800 5,6,7,8,9,10,11,12,14,15,16,17,
18,19,20,21,22,23,24,25,26,27,28,29

SC
Table 4: Cost for sensor reallocation for TE process
Sensor reallocation cost Sensor reallocation cost
From To From To

U
1 2 10 8 25 160
1 3 10 9 26 160
2
2
3
1
3
1
AN10
10
10
10
13
22
27
4
5
160
60
160
3 2 10 23 6 160
M
4 13 60 24 7 160
5 22 160 25 8 160
6 23 160 26 9 160
7 24 160 27 10 160
D

also solution of the lexicographic optimization problem in decreasing order of

TE

priority. Thus, for an optimal U, Φ combination, the best possible value of RI is

obtained and for an optimal U, RI , Φ combination, the best possible value of cs
EP

is obtained. Consider the designs corresponding to C ∗ = 2000 and C ∗ = 3000.

In both these designs, the U values are the same but RI is higher for the latter
design thereby providing more reallocation options while retrofitting in future.
C

Comparison of Base Case Sensor Network Designs Obtained with and without
AC

Consideration of Reallocation Index

To demonstrate the utility of the proposed RI based base case design ap-
proach, we compare the performance after retrofit of two sensor networks: (i)
the sensor network, labeled SN1 , obtained for C ∗ = 6000 in Table 5 in this work,
and (ii) the sensor network, labeled SN2 , obtained for C ∗ = 6000 in Bhushan

24
 ACCEPTED MANUSCRIPT

Table 5: Results for reallocation index based robust sensor networks in the presence of uncer-
tain probability data

C∗ U Φ RI cs Sensors selected

PT
1000 -2 6 7 0 1, 2, 3(2), 4(2)
2000 -2 15 3 0 3(5), 4(5)
3000 -2 21 7 0 1, 2, 3(7), 4(7)
4000 -2 30 3 0 3(10), 4(10)

RI
5000 -5 0 11 100 1, 2, 3, 4, 5, 8, 9, 13, 42, 43, 45(2),
46, 47(2), 48, 49(2), 50
6000 -5 9 9 0 2, 3(4), 4(4), 5, 8, 9, 13, 42, 43, 45(2),

SC
46, 47(2), 48, 49(2), 50
7000 -5 15 11 100 1, 2, 3(6), 4(6), 5, 8, 9, 13, 42, 43, 45(2),
46, 47(2), 48, 49(2), 50
8000 -5 24 9 0 2, 3(9), 4(9), 5, 8, 9, 13, 42, 43, 45(2),

U
46, 47(2), 48, 49(2), 50
9000 -8 0 7 50 3(2), 4(2), 5(2), 8(2), 9(2), 13(2), 42(2), 43(2), 45(3),
AN
46(2), 47(3), 48(2), 49(3), 50(2)

et al. [22] (problem IV in that paper) without consideration of RI . These two

M
sensor networks and their reallocation indices are:
SN1 = [2, 3(4), 4(4), 5, 8, 9, 13, 42, 43, 45(2), 46, 47(2), 48, 49(2), 50]; RI,1 = 9
SN2 = [3(4), 4(4), 5, 8, 9, 13, 42, 43, 45(2), 46, 47(2), 48, 49(2), 50]; RI,2 = 7
D

To compare their performances, the original initial response matrix, labeled A

TE

was randomly perturbed. The matrix A contains effects of original faults on

variables as +1 (positive effect), -1 (negative effect), ? (effect with unknown
sign) and 0 (no effect) and is randomly perturbed. This perturbation simulates
EP

unknown changes in the process. In particular, randomly chosen +1 entries are

changed to -1 or ? and similarly randomly chosen -1 entries are changed to +1
or ?. The new fault-variable bipartite matrix B is generated again using single
C

fault resolution procedure [20]. With this bipartite matrix, the sensor network
AC

obtained using base case design (SN1 or SN2 ) is retrofitted using Formulation I
presented in section 2.4. Towards this end, the objective f (x) in Formulation I
is taken to be [U, Φ, cs ] in decreasing order of priority with C ∗ = 6000 and g(x)
is replaced by constraints relevant to robust fault diagnosis as in Formulation
II. This is done for both SN1 and SN2 as the existing sensor network. This

25
 ACCEPTED MANUSCRIPT

process is carried out for 1000 perturbations of the A matrix and retrofit results
analyzed. It was found that for about 65% of the cases, lower cost was used to
retrofit network SN1 compared to the cost used for retrofitting SN2 while at

PT
the same time obtaining same or lower system unreliability. This demonstrates
that incorporating RI in the design stage enables cost-effective retrofit.

RI
5.2. Observability of All Variables: Crude Oil Distillation Unit

In this section, a simplified crude oil atmospheric distillation unit shown in

SC
Figure 4 [18] is considered as a case study to show the applicability of reallo-
cation index based sensor network design. As shown in Figure 4, there are 5

U
units in the process (representing 5 mass balances) and 13 edges (representing
mass flows). It can be seen that edges 10, 11, 12 and 13 are parallel edges.
AN
Thus, to ensure observability of all these four edges, at least three will have to
be measured. Therefore, without any loss of generality it is assumed that flow
rates corresponding to edges 11, 12 and 13 are measured and eliminated from
M
further analysis. The reallocation compatibility matrix is given in Table 6 which
lists the indices of the non-zero columns in each row.
D

Table 6: Reallocation compatibility matrix (RCM) for crude oil distillation unit
TE

Row Columns
1 2, 7, 9
2 1, 3, 8
3 2, 4, 5, 7, 8, 9
EP

4 3, 5, 6, 7, 8, 10
5 3, 4, 6, 8
6 4, 5, 7, 10
7 1, 3, 4, 6, 8, 10
C

8 2, 3, 4, 5, 7
9 1, 3, 10
10 4, 6, 7, 9
AC

Formulation III is now used to obtain a sensor network that maximizes

the reallocation index. This is done for various number of allowed sensors k.
The resulting optimization problems are solved using TOMLAB MINLP solver
operating in Matlab R
version 7.4. The results are presented in Table 7. Apart

26
 ACCEPTED MANUSCRIPT

2 2 3 7 5 8

4 9

PT
10

RI
1 4 11

SC
12

1 6

U
5 AN3 13

Figure 4: Atmospheric distillation unit [18]

M

from the reallocation index (that is maximized in Formulation III), the flexibility
D

(Equation 2) of the resulting sensor network is also listed in this table. For the
sake of comparison, sensor network designs obtained using Formulation V taken
TE

from literature [18] and presented in Appendix C are also shown in Table 8.
Formulation V minimizes the number of unmeasured paths thereby maximizing
EP

flexibility. The reallocation indices of the resulting sensor networks are also
listed. It can be seen from Tables 7 and 8 that while the reallocation indices
are higher in Table 7 compared to Table 8, the reverse holds for flexibility.
C

Table 7: Sensor networks designed to maximize reallocation index (Formulation IV)

AC

k (Allowed sensors) Sensors selected Flexibility RI

6 2, 4, 5, 7, 9, 10 0.5333 16
7 1, 2, 4, 5, 6, 7, 9 0.6 13
8 1, 2, 5, 6, 7, 8, 9, 10 0.8 10

27
 ACCEPTED MANUSCRIPT

Table 8: Sensor networks designed to maximize flexibility (Formulation V)

k (Allowed sensors) Sensors selected Flexibility RI
6 1, 2, 3, 6, 7, 8 0.6 11

PT
7 1, 2, 4, 6, 7, 8, 10 0.8 9
8 1, 3, 4, 5, 6, 7, 8, 9 0.8667 7

RI
Comparison of sensor network designs based on reallocation index and flexibility
To show the utility of reallocation index based design, the sensor networks
obtained for k = 6 from Tables 7 and 8 are compared. These networks are:

SC
SN1 : [2, 4, 5, 7, 9, 10]
SN2 : [1, 2, 3, 6, 7, 8]

U
To compare these networks, the process graph is perturbed as follows: a maxi-
mum of seven new edges can be added without introducing any further parallel
AN
edges. We consider addition of all possible triplets of edges out of these seven
edges, i.e. 73 = 35 combinations. For each case, networks SN1 and SN2 are

retrofitted using Formulation I. Minimization of retrofit cost is considered to be

M
the objective function in Formulation I while constraint is imposed to ensure ob-
servability of all variables (i.e. no cycles involving only unmeasured variables).
D

The cost of all the sensors including the sensors on the new edges, is taken to
TE

be 100 units. The cost of reallocation between any compatible pair is taken
to be 10 units. Further, it is assumed that sensors cannot be reallocated to
the three new edges and only new sensors can be installed on these edges. It
EP

is found that for the 35 possible perturbations, SN1 was retrofitted using cost
of 200 units each time, while SN2 was retrofitted using cost of 300 units each
time. Further, reallocation was performed once while retrofitting SN1 , while it
C

didn’t take place at all while retrofitting SN2 . This illustrates that the cost of
AC

modifying the base sensor network design obtained by maximizing reallocation

index is lower than the cost of modifying the base case sensor network design
obtained by maximizing flexibility.

28
 ACCEPTED MANUSCRIPT

5.3. Accurate Estimation of Key Variables: Topping Unit of a Refinery

In this section, process graph of a topping unit in refinery as shown in

Figure 5 [28] is considered to demonstrate the applicability of reallocation index

PT
while designing sensor networks for accurate estimation of key variables. The
process graph shown in Figure 5 represents a steady state, overall mass balance

RI
process with 20 edges (overall mass flow rates) and 11 nodes (corresponding to
10 mass balances). The reallocation compatibility matrix is given in Table 9.

SC
In particular, indices of only the non-zero columns in each row are listed in this
Table. The data for sensor variance and sensor cost is given in Table 10. The
reallocation costs for all compatible pairs are taken to be two units.

U
11
AN
1
4
1
M
11
5
2
D

14 17 20 6
3 7 8 9 10
TE

2 3 4 5 6 8
19
18
10 13
12
EP

7 9
15
C

16
AC

Figure 5: Process graph of a topping unit in refinery [28]

Reallocation index based base case design is obtained for this case study
using Formulation IV. The results are presented in Table 11 with constraint im-

29
 ACCEPTED MANUSCRIPT

Table 9: Reallocation compatibility matrix (RCM) for topping unit of a refinery

Rows Columns
1 3, 5, 13

PT
2 4, 5, 10, 13, 14, 15
3 1, 8, 17, 19, 20
4 2, 8, 15, 17,18
5 1, 2, 8, 19

RI
6 9, 10, 17, 19
7 12
8 3, 4, 5, 14, 17
9 6, 13

SC
10 2, 6, 14, 17, 20
11 12
12 7, 11, 17, 19
13 1, 2, 9, 14, 15, 16

U
14 2, 8, 10, 13, 16, 20
15 2, 4, 13, 17, 19
16
17
AN
13, 14, 18
3, 4, 6, 8, 10, 12, 15
18 4, 16
19 3, 5, 6, 12, 15
M
20 3, 10, 14

Table 10: Sensor variance and sensor cost data for topping unit in refinery
D

Variable Sensor Sensor Variable Sensor Sensor

variance cost variance cost
TE

1 100 100 11 200 50

2 100 100 12 500 20
3 100 100 13 400 25
4 200 50 14 400 25
EP

5 200 50 15 500 20
6 100 100 16 400 25
7 100 100 17 400 25
8 200 50 18 500 20
C

9 200 50 19 400 25
10 500 20 20 500 20
AC

posed on required variance on estimate of variable 6. Formulation IV was solved

through enumeration in Matlab R
version 7.4 for various values of threshold
variance.

30
 ACCEPTED MANUSCRIPT

Table 11: Results for reallocation index based sensor network design for data reconciliation
∗
S. No. S6,6 Ŝ6,6 Network cost RI Sensors selected
1 73 72.72 600 20 2, 3, 4, 5, 6, 7, 8, 9

PT
2 75 75 320 17 4, 5, 6, 8, 9, 10
3 90 80 250 14 4, 5, 6, 8

RI
Comparison of Base Case Sensor Network Designs Obtained with and without
Reallocation Index
We compare the retrofit performance of two networks: (i) sensor network,

SC
∗
labeled SN1 , obtained using reallocation index (Formulation IV) with S6,6 =
90, (ii) sensor network, labeled SN2 , obtained without considering reallocation

U
∗
index with S6,6 = 90. SN2 was obtained by solving an optimization problem
similar to Formulation IV but without considering reallocation index in the
AN
objective function and related constraints. These resulting networks and the
corresponding reallocation index are:
SN1 = [4, 5, 6, 8], RI,1 = 14
M
SN2 = [5, 6, 8, 9], RI,2 = 11
To compare their performances, the process graph was randomly perturbed ten
D

times by adding two new edges each time between two pairs of nodes which
TE

are not joint by an existing edge in the graph. The corresponding costs and
variances of sensors that can be used to measure these additional variables are
randomly chosen. Reallocations are not allowed to the newly added edges, but
EP

new sensors can be installed on these. It was found that about 50% cases use
lower cost to estimate the variance of variable 6 using retrofit strategy given in
Formulation I and sensor network SN1 .
C
AC

6. Conclusions

In this work, a novel concept of reallocation index based sensor network de-
sign has been developed. This approach achieves the integration of base case
design and retrofit. In particular, it ensures maximization of reallocation options
in the future without compromising on the primary design objectives. Having

31
 ACCEPTED MANUSCRIPT

these reallocation options ensures that retrofit in future can be carried out at
relatively lower costs. The proposed approach is discussed with three different
criteria related to (i) robust design for reliable fault diagnosis, (ii) observability

PT
of variables for steady state flow processes, and (iii) accurate estimation of key
variables using data reconciliation for steady state linear processes. Application

RI
on case studies demonstrated the utility of reallocation index approach in ob-
taining sensor networks that can be retrofitted easily in future thereby enabling
safe and optimal process operations. Extension of these ideas in selecting vari-

SC
ables to be measured by explicitly incorporating fault severity is currently being
pursued.

U
Nomenclature

A
AN
: Adjacency matrix of mass balance process in Formulation IV
a : Number of process units (nodes) in Formulation III
M
B : Bipartite matrix between faults and measurable variables
Bij : Element if fault i affecting variable j in faultset matrix
D

∗
C : Available cost for sensor location
cj : Cost of sensor used to measure variable j
TE

cs : Cost saved (non-negative slack variable for the cost constraint)

fi : Occurrence probability of fault i
htr : Cost of transferring sensor from variable t to r
EP

I : Set of fault indices considered in a Formulation II

If : Set of indices of faults with uncertain occurrence probability
C

Is : Set of indices of faults which affect uncertain sensors

J : Set of indices of all measurable variables
AC

Js : Set of indices of uncertain sensors

k : Number of measured variables in the process (Formulation III).
L : Large positive number in Formulation IV
M : Reallocation compatibility matrix
m : Number of variables

32
 ACCEPTED MANUSCRIPT

N : Total number of variables measured in the process

n : Number of measurable variables in the process
nj : Binary variable which is 1 if variable j is measured and is 0 otherwise

PT
P : A large positive constant in Formulation II, or
Inverse of ASAT in Formulation IV

RI
p : Number of faults in the process
q : Number of unmeasured paths in Equation (2)

SC
qj : Number of new sensors upgraded at variable j
R : Set of variables which can receive sensors from other variables
Rj : Set of variables which can receive sensors from variable j

U
S : Sensor variance matrix
Sjj : Diagonal element of sensor variance matrix S
Sb
AN
: Covariance matrix of reconciled estimates
Sbjj : Reconciled variance at variable j
M
∗
Sjj : Threshold variance at variable j
SS : Set of variables where new sensors can be installed (upgrade)
sj : Failure probability of sensor used for measuring variable j
D

T : Set of variables whose sensors can be reallocated to other variables

TE

Tj : Set of variables which can transfer sensors can to variable j

U : System unreliability in log form (maximum unreliability of detection amongst all faults)
Ui : Unreliability of detection of fault i
EP

utr : Number of sensors transferred from variable t to r

xj : Number of sensors used for measuring variable j
Binary decision variable in Formulation III
C

x∗j : Number of existing sensors at variable j

AC

yj : Fake sensor at variable j in Formulation IV

zi : Binary variable corresponding to uncertain fault i
Greek Symbols

λ 
: Constants used for lexicographic optimization
β1 , β2 , β3 , β4 

33
 ACCEPTED MANUSCRIPT

Γ : Flexibility of sensor network

Φ : System (overall) robustness
φ∗ : Maximum meaningful value of system robustness

PT
φi : Maximum meaningful value of slack for uncertain fault i
φ∗f i : Maximum slack required for fault i to account

RI
for uncertainty in occurrence probability of fault i
φ∗si : Maximum slack required for fault i to account

SC
for uncertainty in sensor failure probabilities

Appendix A. CSTR Example: Construction of Bipartite Matrix B

U
In this section, we discuss the approach for obtaining the fault-variable bi-
AN
partite matrix B that is used in sensor network design. A jacketed, continuous
stirred tank reactor (CSTR) shown in Figure A.6 , involving a second order
exothermic irreversible reaction A + B → P , is taken as illustration. Equations
M
A.1 - A.4 represent the dynamic model of the jacketed CSTR [29].
D
TE

Tj
CAi
A Jacket
EP

qf
B
CBi
C

CA, CB , CP , T
AC

Figure A.6: Jacketed CSTR [29]

34
 ACCEPTED MANUSCRIPT

 
dCA qf (CAi − CA ) −E2
= − a2 exp CA CB (A.1)
dt V RT
 
dCB qf (CBi − CB ) −E2
= − a2 exp CA CB (A.2)
dt V RT

PT
 
dCP −qf CP −E2
= + a2 exp CA CB (A.3)
dt V RT
qf (Ti − T ) (−∆H)a2 exp −E


dT 2
CA CB h · s(T − Tj )

RI
RT
= + − (A.4)
dt V ρCp V ρCp

In the above equations, qf is the inlet as well as the outlet volumetric flowrate,

SC
CAi and CBi are the concentrations of the reactants in the inlet stream, and Ti
is the temperature of this stream. CA , CB , CP are the concentrations of species

U
A, B and P in the product (outlet) stream, T is the reactor temperature and
Tj is the jacket temperature. Changes in inlet conditions (CAi , CBi , Ti ) are
AN
assumed to be possible faults. The system (state) variables are CA , CB , CP ,
and T . The parameter values are given in Table A.12. The steady state values
of exogenous variables (or faults) and system variables are given in Table A.13.
M
D

Table A.12: Parameter values for jacketed CSTR

Parameter Value Parameter Value Parameter Value

ρ (kg/m3 ) 2.42 × 104 V (m3 )
TE

790.2 E2 (kJ/kmol) 0.5

R (kJ/kmol.K) 8.314 a2 (m3 /s.kmol) 29.0 s (m2 ) 4.60
∆H (kJ/kmol) −54000 h (kJ/(m2 .s.K)) 0.060 cp,r (kJ/kg.K) 3.0
EP

Table A.13: Values of exogenous and system variables for jacketed CSTR

Parameter Value Parameter Value Parameter Value

C

CAi (kmol/m3 ) 4.661 CBi (kmol/m3 ) 4.9267 Ti (K) 336

Tj (K) 300 qf (m3 /s) 6.2748 × 10−5 CA (kmol/m3 ) 0.1836
CB (kmol/m3 ) CP (kmol/m3 )
AC

0.4493 4.4774 T (K) 348.325

Generation of Signed Digraph

The signed digraph (SDG) corresponding to the jacketed CSTR is obtained

from Equations A.1 - A.4 using the SDG generation algorithm listed in [30].

35
 ACCEPTED MANUSCRIPT

The SDG is presented in Figure A.7. SDG is a directed graph with set of
vertices representing process variables and faults, while edges (with signs + or
-) represent the cause-effect relationship between the corresponding vertices.

PT
For example, the edge between vertices CAi and CA has positive sign thereby
implying that a positive change in CAi from its steady state value leads to

RI
positive change in CA when all other variables are held constant.

CBi

SC
+
CAi
−

U
CB
+ +
−
−
+
AN − −
+
− CA T Ti
−
M
+ +

CP
D

−
TE

Figure A.7: SDG of jacketed CSTR

EP

Generation of Initial Response Matrix

The qualitative signs of the effects of various faults on process variables

C

is extracted from the SDG and used in the sensor network design procedure.
AC

In particular, these effects are extracted as initial responses [30]. The initial
response of a system variable is defined as the sign of the first non-zero response
caused by a fault and is obtained by propagating through the shortest paths
from the corresponding fault node. The initial response matrix for the CSTR
(Figure A.7) is shown in Table A.14 and may be +, -, 0 or ? (indeterminate
which may happen if there are multiple shortest paths from a fault to a variable

36
 ACCEPTED MANUSCRIPT

with opposite effects). In this matrix, the rows and columns correspond to faults
and system variables, respectively. The tags for faults (F1 − F3 ) and for system
variables (S1 − S4 ) are also listed in Table A.14.

PT
Table A.14: Initial response matrix of CSTR using Figure A.7
Faults System Variables

RI
CA (S1 ) CB (S2 ) CP (S3 ) T (S4 )
CAi (F1 ) +1 -1 +1 +1
CBi (F2 ) -1 +1 +1 +1
Ti (F3 ) -1 -1 +1 +1

SC
Generation of Bipartite Matrix

U
For use in sensor network design for various criteria, a fault variable effects
AN
matrix B, labeled bipartite matrix [20], is derived from the initial response
matrix by simply choosing Bij = 1, i = 1, 2, ..., m; j = 1, 2, ..., p if fault i has
a non-zero initial response on variable j and zero otherwise. By measuring
M
at least one variable such that Bij = 1 ensures that fault i can be detected
or observed by the selected sensor network [8]. Additionally, to be able to
D

differentiate or resolve between faults, Raghuraj et al. [8] defined p2 pseudo

faults corresponding to resolution between each pair of original faults. The k th

TE

pseudo fault corresponding to original pair of faults i and i0 affects variable j

if the sign of effects of faults i and i0 on variable j are different. These pseudo
faults are also added to the bipartite matrix B. This bipartite matrix is listed
EP

below:
C
AC

37
 ACCEPTED MANUSCRIPT

S1 S2 S3 S4

PT
 
F1  1 1 1 1 
 
F2
 
 1 1 1 1 
 

RI
 
F3  1 1 1 0 
B= (A.5)
 
 
 
∆F1,2 (F4 )  1 1 0 0 
 

SC
 
∆F1,3 (F5 ) 
 1 0 0 0 

 
∆F2,3 (F6 ) 0 1 0 0

U
where ∆Fi,i0 is the pseudo fault corresponding to resolution between faults Fi
and Fi0 . In Equation A.5, the pseudo faults have been labeled F4 − F6 for the
sake of simplicity.
AN
Fault Occurrence and Sensor Failure Probabilities
M
Bhushan and Rengaswamy [20] considered probabilities of occurrences of
faults and probabilities of failures of sensors in the sensor network design proce-
D

dure. Given the probabilities of occurrences of the original faults, the probability
TE

of occurrence of a pseudo fault is defined as:

fi,i0 = min(fi , fi0 ) (A.6)

EP

where fi , fi0 are the probabilities of occurrences of faults Fi , Fi0 and fi,i0 is the
probability of occurrence of the pseudo fault ∆Fi,i0 .
C

Figure 1 in the main text displays the bipartite matrix B (Equation A.5) for
the CSTR example. The occurrence probabilities of faults and failure probabil-
AC

ities of sensors are also listed in this figure.

Appendix B. Additional Details of TE Process

The list of streams and indicators in TE process flow sheet shown in Figure
3 are given in Table B.15 and Table B.16.

38
 ACCEPTED MANUSCRIPT

Table B.15: List of streams

Stream Number Stream Name
1 Feed of component A

PT
2 Feed of component D
3 Feed of component E
4 Feed of component C
5 Stripper overhead

RI
6 Reactor feed
7 Reactor product
8 Recycle stream at compressor

SC
9 Purge
10 Separation liquid
11 Product
12 Reactor cooling outlet

U
13 Condenser cooling outlet

Symbol
AN
Table B.16: List of indicators
Type of Indicator
FI Flow indicator
JI Power indicator
M
LI Level indicator
PI Pressure indicator
SC Agitator speed indicator
D

TI Temperature indicator
CWS Condenser water supply
CWR Condenser water return
TE

Appendix C. Optimization Formulation for Flexible Networks

EP

Bo and Bing-Zhen [18] proposed optimization formulation for designing a

flexible sensor network for a linear mass balance process. The network is flexible
C

if it can ensure observability of variables not only in the existing process but also
ensure observability in case additional edges (streams) are added in the process.
AC

Flexibility was formally defined in equation 2. To obtain flexible networks, Bo

and Bing-Zhen [18] proposed the following optimization formulation:

39
 ACCEPTED MANUSCRIPT

Formulation V:

z Y
X
min (1 − xj ) (C.1)

PT
i=1 j∈Zi

Subject to constraints 33, 34 and 36 (C.2)

RI
In the above formulation, the definition of xj is same as in Formulation III.
Further, z is the number of paths in the process graph with length not more

SC
than m − k and Zi is the set of edges in the ith path. Thus, Formulation V
minimizes the number of unmeasured paths while ensuring observability of all
the variables.

U
[1] Vaclavek V, Loucka M. Selection of measurements necessary to achieve mul-
AN
ticomponent mass balances in chemical plant. Chem Eng Sci 1976;31:1199.

[2] Madron F, Veverka V. Optimal selection of measuring points in complex

plants by linear models. AIChE J 1992;38(2):227.
M
[3] Meyer M, Lann JML, Koehret B, Enjalbert M. Optimal selection of sensor
location on a complex plant, using a graph oriented approach. Comput
D

Chem Engng 1994;18 (Suppl.):S535.

TE

[4] Ali Y, Narasimhan S. Sensor network design for maximizing reliability of

linear processes. AIChE J 1993;39(5):820.
EP

[5] Ali Y, Narasimhan S. Redundant sensor network design for linear processes.
AIChE J 1995;41(10):2237–49.
C

[6] Bagajewicz MJ. Design and retrofit of sensor networks in process plants.
AIChE J 1997;43(9):2300.
AC

[7] Lambert HE. Fault trees for locating sensors in process systems. Chem
Eng Prog 1977;Aug.:81.

[8] Raghuraj R, Bhushan M, Rengaswamy R. Location of sensors in complex

chemical plants based on fault diagnostic observability criteria. AIChE J
1999;45(2):310.

40
 ACCEPTED MANUSCRIPT

[9] Bhushan M, Rengaswamy R. Design of sensor network based on the SDG of

the process for efficient fault diagnosis. Ind Eng Chem Res 2000;39(4):999.

PT
[10] Chen JY, Chang CT. Development of an optimal sensor placement proce-
dure based on fault evolution sequences. Ind Eng Chem Res 2008;47:7335–
46.

RI
[11] Chang CT, Mah KN, Tsai CS. A simple design strategy for fault monitoring
systems. AIChE J 1993;39(7):1146.

SC
[12] Commault C, Dion JM, Agha SY. Structural analysis for the sensor location
problem in fault detection and isolation. Automatica 2008;44:2074–80.

U
[13] Krysander M, Frisk E. Sensor placement for fault diagnosis. IEEE
AN
Trans on Systems, Man, and Cybernetics-Part A: Systems and Humans
2008;38(6):1398–410.
M
[14] Bagajewicz M. A review of techniques for instrumentation design and
upgrade in process plants. Canadian Journal of Chemical Engineering
2002;80:3–16.
D

[15] Bagajewicz M, Sánchez M. Reallocation and upgrade of instrumentation

TE

in process plants. Comput Chem Engng 2000;24:1945.

[16] Bhushan M, Narasimhan S, Rengaswamy R. Sensor network realloca-

EP

tion and upgrade for efficient fault diagnosis. In: Fourth International
Conference on Foundations of Computer-Aided Process Operations, Coral
Springs, Florida, USA. 2003, p. 443–6.
C

[17] Kolluri S, Bhushan M. Audit of sensor networks for efficient fault diagnosis.
AC

Journal of Process Control 2013;23:881–93.

[18] Bo L, Bing-Zhen C. Study on flexibility of sensor network for linear pro-

cesses. Comput Chem Engng 2002;26:1363–8.

41
 ACCEPTED MANUSCRIPT

[19] Kolluri S, Bhushan M. Reallocation index based sensor network design

for efficient fault diagnosis. In: 11th International Symposium on Process
Systems Engineering - PSE2012; vol. 31. 2012, p. 1527–31.

PT
[20] Bhushan M, Rengaswamy R. Comprehensive design of a sensor network
for chemical plants based on various diagnosability and reliability criteria:

RI
I. framework. Ind Eng Chem Res 2002;41:1826–39.

[21] Narasimhan S, Rengaswamy R. Quantification of performance of sensor

SC
networks for fault diagnosis. AIChE J 2007;53(4):902–17.

[22] Bhushan M, Narasimhan S, Rengaswamy R. Robust sensor network design

U
for fault diagnosis. Comput Chem Engng 2008;32:1067–84.

AN
[23] Bagajewicz M, Cabrera E. New MILP formulation for instrumentation
network design and upgrade. AIChE J 2002;48:2271.
M
[24] Bagajewicz M, Markowski M, Budek A. Economic value of precision in the
monitoring of linear systems. AIChE J 2005;51(4):1304–9.
D

[25] Narasimhan S, Jordache C. Data reconciliation and gross error detection;

Gulf publishing company; Texas; 2000.
TE

[26] Downs JJ, Vogel EF. A plant-wide industrial process control problem.
Comput Chem Engng 1993;17(3):245.
EP

[27] Maurya MR, Rengaswamy R, Venkatasubramanian V. Application of

signed digraphs-based analysis for fault diagnosis of chemical process flow-
C

sheets. Engng Applic Artif Intell 2004;17:501.

AC

[28] Tsai CS, Chang CT. Optimal alarm logic design for mass-flow networks.
AIChE J 1997;43(11):3021–30.

[29] Maurya MR. Integrating causal models and trend analysis for process fault
diagnosis. Ph.D. thesis; Purdue University; 2003.

42
 ACCEPTED MANUSCRIPT

[30] Maurya MR, Rengaswamy R, Venkatasubramanian V. A systematic frame-

work for the development and analysis of signed digraphs for chemical pro-
cesses. 1. algorithms and analysis. Ind Eng Chem Res 2003;42(20):4789.

PT
RI
U SC
AN
M
D
TE
C EP
AC

43
 ACCEPTED MANUSCRIPT

Research Highlights

PT
RI
• Optimal selection of variables for placing sensors essential for process
safety and loss prevention.

SC
• A generic base case sensor network design approach presented.

• Approach maximizes reallocation options enabling low cost retrofit in

future.

U
• Approach demonstrated for fault diagnosis, accurate estimation of key
AN
variables, and observability of variables after a change in process layout.

• Approach used to design sensor networks for various case studies.

M

• Simulation experiments demonstrate low cost retrofit of designed net-

works.
D
TE
C EP
AC