Process Plant Instrumentation Design and Upgrade

Process Plant
Instrumentation
Design and Upgrade
Process Plant
Instrumentation
Design and Upgrade
Miguel J. Bagajewicz, Ph.D.

School of Chemical Engineering and Materials Science
University of Oklahoma
Boca Raton London New York
CRC Press is an imprint of the

Taylor & Francis Group, an informa business
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2000 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business
No claim to original U.S. Government works
Printed on acid-free paper

Version Date: 20161019
International Standard Book Number-13: 978-1-56676-998-3 (Hardback)
This book contains information obtained from authentic and highly regarded sources. Reason-
able efforts have been made to publish reliable data and information, but the author and publisher
cannot assume responsibility for the validity of all materials or the consequences of their use. The
authors and publishers have attempted to trace the copyright holders of all material reproduced in
this publication and apologize to copyright holders if permission to publish in this form has not
been obtained. If any copyright material has not been acknowledged please write and let us know so
we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information
storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, please access www.
copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc.
(CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organiza-
tion that provides licenses and registration for a variety of users. For organizations that have been
granted a photocopy license by the CCC, a separate system of payment has been arranged.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and
are used only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
To all those people from my generation who pursued a
country in which it would be worth living. Some of
them disappeared in the process.
CONTENTS
Preface xiii
Acknowledgments xvii
Chapter 1. Plant Data Management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Introduction 1
Plant Information and Operations Management 1
Model-Based Monitoring 4
Quality of Data 5
References 11
Chapter 2. Instrumentation Design Goals . . . . . . . . . . . . . . . . . . . . . . . 15

Introduction 15
Measured and Key Variables 15
Selection of Monitoring Variables 16
Selection of Key Variables in Control 17
Selection of Measured Variables for Fault Diagnosis 24
Instrumentation Design Goals 25
Upgrading of Instrumentation 27
Nomenclature 28
References 28
Chapter 3. Instrumentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Introduction 31
Flow Rate Instrumentation 31
Level Measurement 46
Temperature Measurement 51
vii
viii Contents
Pressure Measurement 56
Density Measurement 58
On-Line Process Analyzers 60
Transmission and Transformation of Signals 60
Nomenclature 62
References 63
Chapter 4. Errors in Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

Introduction 65
Instrument Properties 65
Measurement Quality 66
Sensitivity and Speed of Response 71
Hysteresis and Dead Band 72
Calibration Curves 74
Accuracy of Different Instruments 76
Nomenclature 78
References 78
Chapter 5. Variable Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

Introduction 81
Model 81
Measurement Equation 82
Graphs and Flowsheets 82
Connectivity of Systems 83
Observability 85
Redundancy 86
Linear Systems 87
Canonical Representation of Linear Systems 87
Other Methods 93
System Degree of Redundancy 95
Quantification of Observability and Redundancy 96
Graphs and Canonical Matrices 100
Nonlinear Systems 110
Full Nonlinear Systems 120
Nomenclature 123
References 125
Chapter 6. Design and Upgrade of Nonredundant and

Redundant Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Introduction 127
Upgrade and/or Design Goals 127
Design for Estimability 128
Contents ix
Design for Estimability Efficiency 139

Compulsory Measurements and the Upgrade Case 140
Sensor Networks for Bilinear Systems 142
Nomenclature 156
References 157
Chapter 7. Data Reconciliation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

Data Reconciliation 159
Background 159
Linear Data Reconciliation 162
Steady-State Linear Data Reconciliation 163
Nonlinear Steady-State Data Reconciliation 169
Dynamic Data Reconciliation 170
Nomenclature 171
References 172
Chapter 8. Design of Precise Sensor Networks . . . . . . . . . . . . . . . . 175

Introduction 175
Cost-Optimal Design 175
Multiple Instruments and Hardware Redundancy 176
Maximum Precision Models 178
Generalized Maximum Precision Model 180
Relation Between Sensor Network Models 181
Solution Procedures for Linear Systems 183
Parameter Estimation in Nonlinear Systems 184
Nomenclature 189
References 190
Chapter 9. Precision Upgrade of Sensor Networks . . . . . . . . . . . . 191

Introduction 191
Upgrade Options 191
Cost Benefit Analysis 192
Upgrade Models Based on Addition of Sensors 192
Model for Resource Reallocation 198
Generalized Model for Resource Reallocation and Upgrade 201
Nomenclature 209
References 209
Chapter 10. Reliability of Nonrepairable

Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Introduction 211
Sensor Service Availability 211
x Contents
Sensor Service Reliability 212

Failure Density and Failure Rate 212
Markovian Model 214
Mean Time to Failure 216
Estimation Availability and Reliability of Variables 217
Determination of Estimation Reliability 218
Estimation Reliability in Nonredundant Systems 221
Availability, Reliability and Degree of Estimability 222
System Availability and Reliability 222
Nomenclature 223
References 224
Chapter 11. Design of Reliable Linear Nonrepairable

Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Introduction 225
Nonredundant Networks Featuring Maximum
Reliability 225
Redundant Networks Featuring Maximum Reliability 237
Redundant Networks Featuring Maximum Reliability and
Hardware Redundancy 245
Redundant and Restricted Networks 246
Nomenclature 250
References 251
Chapter 12. Design of Reliable Bilinear Nonrepairable

Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Introduction 253
Bilinear Multicomponent Systems 253
Energy Networks 266
Nomenclature 266
References 267
Chapter 13. Design of Reliable and Cost-Efficient

Nonrepairable Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Introduction 269
Minimum Cost Model 269
Minimum Number of Sensors Model 270
Solution Procedure 270
Relation to Other Models 270
Limitations of Previous Models 271
Generalized Maximum Reliability Model 271
Nomenclature 278
References 278
Contents xi
Chapter 14. Design of Repairable Sensor Networks . . . . . . . . . . . 279

Introduction 279
Failure Intensity 279
Repair Intensity 280
Expected Number of Repairs 280
Maintenance and Total Cost 284
Residual Precision 284
Nomenclature 289
References 290
Chapter 15. Design of Robust Sensor Networks . . . . . . . . . . . . . . . 291

Introduction 291
Origin of Gross Errors 291
Gross Error Handling 293
Test for Gross Error Presence 293
Gross Error Detection in Dynamic Data
Reconciliation 295
Inaccuracy in Gross Error Detection 296
Multiple Gross Error Identification 296
Gross Error Size Estimation 297
Sensor Network Error Detectability 297
Sensor Network Gross Error Resilience 298
Robust Sensor Networks 299
Minimum Cost Model for Robust Networks 300
Nomenclature 302
References 303
Chapter 16. Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307

Introduction 307
Genetic Algorithms 307
Nomenclature 313
References 313
Chapter 17. Design of Sensors for Process

Fault Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
Introduction 315
Fault Detection, Diagnosis and Alarms 315
Fault Observability 326
Fault Resolution 326
Sensor Network Design 327
xii Contents
Sensor Location for Fault Observability 327

Sensor Location for Fault Resolution 332
Nomenclature 337
References 338
Index 343
PREFACE
The process of obtaining good estimates of process variables that conform to

basic laws has received the name estimation. An example of this is the Kalman
filter. Estimators based on the assumption of steady state have existed since
the seminal article by Kuehn and Davidson [Chem. Eng. Prog., 57, 44(1961)],
and are known as data reconciliation methods. Recently, estimation methods
for dynamic systems, which received the name of dynamic data reconcilia-
tion, have been developed. In reality, these latter methods can be classified as
optimal estimation methods. Names aside, the real issue at stake is not the pres-
ence of noise, but the detection of gross errors, that is, biased instruments and
leaks.
There is a vast amount of literature devoted to the selection and good mainte-
nance of instruments. This literature covers the selection of the right instrument
for a particular range and system, but only after the desired accuracy and relia-
bility of measurement have been established. Little has been written on how to
systematically determine the right accuracy and reliability needed when select-
ing an instrument, much less how much redundancy is needed for a particular
system. The key variables that need estimation come from control require-
ments, as well as monitoring needs for safety, quality control and production
accounting. These are the starting points of the design methodology. This book
concentrates on determining the optimal accuracy and reliability of instruments
and their location. To determine this, certain desired properties of the system of
instruments are used as constraints while the cost is minimized. These proper-
ties, among others, are variable observability, system reliability and precision
of certain variables.
This book is not a textbook. Rather, it is intended to be an organized col-
lection of the most relevant work in this area, published until the middle of
1999. It may, therefore, seem a book for specialists. However, it has been
xiii
xiv Preface
written with the intention of making it readable by regular engineers with some
background in linear algebra, mathematical optimization and graph theory. It
is organized so that the complexity of the sensor network design is addressed
step by step. Thus, after the issue of errors and estimability has been covered
in Chapters 4 and 5, Chapter 6 covers the design of sensor networks that fulfill
redundancy goals. Chapter 7 presents a review of data reconciliation, so that
the design of sensor networks with accuracy/precision goals can be presented
in Chapter 8. Chapter 9 presents models for the upgrade of sensor networks.
Chapter 10 covers new concepts of reliability associated with sensor networks.
Chapters 11 and 12 are devoted to the design of reliable systems. Chapter 13
incorporates maintenance, that is, deals with repairable systems. Chapter 14
reviews the theory of gross error detection preparing the ground for Chapter 15,
which presents techniques that help design sensor networks that are capable of
filtering gross errors appropriately. Finally, Chapter 16 covers the use of genetic
algorithms to solve some of the problems posed in the book, and Chapter 17
is devoted to the incipient work on instrumentation location for fault detection.
To avoid appendices, the book blends new concepts with the presentation
of known background material. For example, well-known concepts from graph
theory, such as spanning trees, cutsets and cycles, are blended with the new
concept of estimability in Chapter 5.
One of my former advisors (I had three) taught me some elements of the art
of dreaming. He had a sign posted outside his office. It said something like this:
The great projects are
r dreamt by the crazy
r executed by the audacious
r enjoyed by the happy
He was crazy and audacious, as required. However, he did not have the
tenacity to hold on to his project. He has, nonetheless, instilled in me what he
was preaching and doing. On a much smaller scale, I also had a dream, worked
with tenacity, and am enjoying the result, despite the sour taste of having to let
this manuscript go to the publisher without being absolutely sure it is perfect,
as I would have preferred. Oscar Wilde said it very well: “In this world there are
only two tragedies. One is not getting what one wants, and the other is getting
it.”
I spent endless hours away from the family and from myself. Neither my
family nor probably myself would have benefited from not writing this book,
simply because I would be shamelessly pursuing some other lovely chemical
engineering endeavors. My wife has asked me not to apologize for the time
stolen from the family, so I will not. At least, tradition states, I should dedi-
cate it to her and my children. I am sure they will agree that the dedication I
have chosen brings back many buried ghosts from our common past, and pays
Preface xv
tribute to those victims of state terrorism executed during the summary and
barbaric killings of the 1970s in my country Argentina. There are other more
civilized ways to handle dissent of any kind, even dissent that is perceived as
evil. My native country is finally finding the strength to remember. I hope it is
also acquiring the wisdom to prevent these tragedies from being repeated.
ACKNOWLEDGMENTS
When I accepted this challenge, I naively thought that writing a book is just
a matter of knowing the material and putting together a nice story about it. It
is after all, my first book, so I should be forgiven. Very soon I realized what
nobody had told me, or if I had heard it, I had ignored. Knowledge cannot be
put together so easily, especially when you are the first one writing about an
issue, because there is nowhere to look. After I started to organize the material,
I realized that there are many holes in the story, many unanswered questions
that make the story lose continuity. I postponed, as is probably customary, the
end date of the book, so I would be able to answer some of these questions.
Mabel Sánchez has been a remarkable brain to have available to accomplish
this task. Although Qiyou Jiang did not work directly in sensor location matters,
his outstanding work on data reconciliation has influenced my thinking to such
an immeasurable extent. I regret not having him around anymore. Both of them
have contributed to what have been, so far, the most academically productive
years of my life. My deep thanks are also extended to my students Mariano
Savelski, Hernán Rodera and José Soto, who patiently proofread some chapters
of the manuscript and provided constructive criticisms. My student, Margiori
Rivas, patiently read the entire book and warned me of many mistakes. For this I
am very grateful. Finally, Dr. Shankar Narasimhan provided valuable feedback
on sections of the book. The great staff of the School of Chemical Engineering
and Materials Science at the University of Oklahoma, who make my office
life so livable, have proven, once again, that intelligence is the ability to laugh
at oneself without regret. They have facilitated so many things that I cannot
enumerate.
xvii
CHAPTER 1
Plant Data Management
INTRODUCTION
Instrumentation is needed in process plants to obtain data that are essential

to perform several activities. Among the most important are control, the assess-
ment of the quality of products, production accounting (sometimes called yield
accounting) and the detection of failures related to safety. In addition, certain
parameters that cannot be measured directly, such as heat exchanger fouling
or column efficiencies, are of interest. Finally, new techniques, such as on-line
optimization, require the construction of reliable computer models for which
the estimation of process parameters is essential.
This book concentrates on the tasks of determining the optimal set of mea-
sured variables and selecting the accuracy and reliability of the corresponding
instruments. The goal is to obtain sufficiently accurate and reliable estimates
of variables of interest while filtering bad data due to possible instrument mal-
function. An additional goal is to observe and diagnose single and multiple
process faults.
First, this chapter highlights the different activities in which data play an
important role. Next, model-based monitoring and quality of data, including an
overview of the different ways of guaranteeing the desired quality, are discussed.
This background material provides the foundation for the goals of instrumen-
tation design and upgrade.
PLANT INFORMATION AND OPERATIONS MANAGEMENT
In the last decade, the chemical process industry has been incorporating new
technologies in the form of several computer programs to help gather, filter,
1
2 PLANT DATA MANAGEMENT
organize and use the information of plants for several technical and management
activities. Through these different software packages, quality of products and
cost containment have improved considerably. Efficiency of operations, which
in the past relied on the expertise of local unit operators, can now rely on inter-
active and computer-based supervisory control activities at the plant-wide level.
In addition, yield accounting, operations planning and maintenance scheduling
benefit from more reliable and accurate data. In view of these new advances in
data processing, even management procedures are undergoing revision.
Most process plants are designed to run at steady-state conditions. In practice,
these conditions are not strictly met because plants are subject to minor unpre-
dictable changes. However, the assumption of steady state is still successfully
used to perform most of the gathered data analysis, with the obvious exception
of control. This analysis, followed by decision making, covers several activities
of plant operation. Briefly, some of these activities are:
r Operation monitoring: This is a short-term activity in which the data are
used, either as part of control loops or as means to change control settings.
r Fault detection: This includes the detection of instrument malfunction and
equipment failure and the assessment and quantification of leaks.
r Performance analysis: This activity is typically performed on a daily basis
and covers what in the oil-refining industry is called yield accounting, oil
accounting or oil movement.
r Process modeling: Simulation has proven to be an efficient tool for the pro-
cess engineer. It provides means to analyze alternative processing conditions,
to evaluate retrofit changes and detect operational problems.
r Operations planning: Typical activities, such as scheduling, fouling moni-
toring, heat exchanger cleaning and catalyst activation, rely on data from the
process and its adequate modeling.
r Production planning: This is performed on a larger time scale. Several soft-
ware packages have been developed to address this problem, especially for
the oil industry.
r Maintenance planning: Production losses increase with poor maintenance.
As frequency of a given maintenance plan is increased, cost increases. Natu-
rally, a trade-off is obtained where total cost of production loss and mainte-
nance is minimum. If the maintenance plan can be modified by using tech-
niques that allow early identification of potential problems, the overall cost
is reduced.
r Parameter estimation: The estimation of parameters is of paramount impor-
tance in practice. Typically, these parameters are column efficiencies, heat
transfer coefficients, flash vaporization efficiencies, etc. All of them are im-
possible to be measured directly. In many cases, the precision and location
Plant Information and Operations Management 3
of the existing instrumentation are insufficient to obtain good quality esti-

mates. A typical example is the prediction of fouling in crude fractionation
preheating trains, which is information of vital importance for the optimiza-
tion of the cleaning cycles.
r On-line optimization: Data from the plant are used to “tune” the parameters
of a simulation model, which in turn are used as a basis for an optimization of
the plant operation. As a result, a new steady-state condition that maximizes
profit is obtained. This new technology was introduced in industry in the 1980s
and has resulted in millions of dollars of savings in operating costs (from 2% to
20%) for different scenario applications: crude fractionation plant: (Mullick,
1993); distillation (Smith, 1996); ethylene plants (Lauks et al., 1992); olefin
plants, catalytic cracking, hydrocrackers, bisphenol-A plants, FCC units, and
other cases referenced by Zhang et al. (1995) and Brydges et al. (1998). The
cycle of Figure 1.1 depicts how real-time optimization cycles operate in plants
(Forbes and Marlin, 1996). Not only hierarchical optimization and control
loops are already in practice producing millions of dollars of savings, but
there are also plans to perform closed loop optimization, that is, automatic
update of control setpoints.
r Enterprise resource planning: This term was coined recently to refer to a
planning activity, based mainly on optimization procedures, that will com-
prise several of the above described planning activities (operations, produc-
tion, maintenance) in addition to finance and management planning. (Badell
et al., 1998; Bunch, 1998; Grosdidier, 1998). Thus, a vertical integration be-
tween business and plant operations using optimization models is starting to
take place (Badell and Puigjaner, 1998). An integration of all these activi-
ties, from operation and monitoring to long-term planning, is also known as
Plant-Wide Management and Control (Swanson and Stewart, 1994; Pelham
and Pharris, 1996). Some companies have started to integrate management
Estimate parameters Update the model
Filtering of gross errors & Optimize & Update

Data reconciliation controller set points
Plant & Controllers
FIGURE 1.1. Real-time optimization.

and enterprise resource planning with the above activities (Benson, 1995;
Natori and Tjoa, 1998; Harkins, 1999). The idea of operating plants in a
“hands-off” mode was for decades a dream of plant managers and plant
decision-making teams. With reliable data already available and with the
help of the undergoing research, this dream is starting to become a reality
(Koolen, 1994).
It is clear that quality of data is a key aspect for all the above activities. Some
of them, such as on-line optimization, cannot be successfully performed without
accurate and consistent data. This challenges designers to determine the best
sensor network in terms of number of instruments, location and quality that is
needed for this whole cycle to work properly. The choice of instrumentation
has therefore ceased to be a localized and independent problem in the plant to
become one that is global and multiobjective.
MODEL-BASED MONITORING
Model-based monitoring consists of the use of a combination of models

(formal descriptions of process behavior) and on-line measurements to achieve
the following goals:
a. Produce estimates of measured and unmeasured variables
b. Identify malfunctioning instrumentation
c. Identify unsafe and/or faulty operating conditions and their origin
d. Identify events that can impact efficiency and quality of products
Kramer and Mah (1993), in an excellent overview of this issue, discussed
several scenarios in which good estimates of data can be obtained. Data rec-
onciliation and gross error detection are techniques that help to accomplish
tasks (a) and (b) and constitute a particular case of the concept of data rectifi-
cation. Although data reconciliation relies on analytical constraints and mostly
on least square estimation, data rectification can obtain these estimates by
using techniques such as Kalman filtering, pattern recognition, neural net-
works, principal component analysis and partial least squares. Likewise, the
detection of faults relies on techniques based on statistics, but several other
techniques can be used. Tasks (c) and (d) are direct consequences of using a
model.
The field of sensor network design and upgrading has traditionally relied on
model-based concepts. Almost all work addressing monitoring goals include:
—the ability of the sensor network to provide estimates of variables of interest
—the ability to guarantee certain accuracy through data reconciliation
—reliability
Quality of Data 5
—capability of identifying gross errors through statistical model-based tech-

niques
—capability of identifying process faults
QUALITY OF DATA
Quality of data is a loose term that is often used to refer to several properties
that a set of data should have. These properties can be condensed in the following
three broad attributes:
r accuracy
r precision or reproducibility
r reliability
Accuracy is the ability of an instrument to measure the correct or “true” value.
In turn, precision can be defined as the ability of a sensor to reproduce a value
within a certain interval. Thus, an instrument can be precise and not accurate.
This happens when repeated measurements of the same variable fall within
a small interval that does not contain the true value. Conversely, an accurate
instrument may have poor reproducibility, but the value of the mean of many
of its measurements may be close to the true value. Figure 1.2 illustrates these
concepts.
Although these concepts are fairly simple, they are sometimes confused.
Indeed, it is common to see engineers to refer to precise systems as accurate
and vice versa. Moreover, some books and manuals use the concept of precision
as being part of a more general concept of accuracy that includes both the
variance of the measurements and the deviation of the mean from the true
value. Unfortunately, instead of proposing a new name, accuracy is still used,
increasing the confusion.
Reliability of data is defined as the probability that the data will be actually
not absent during a certain period of time. In turn, availability of data is the
probability of data not being absent at a certain instance in time. In this regard,
reliability is a more stringent requirement.
Accurate and imprecise
Inaccurate and precise True value
FIGURE 1.2. Accuracy and precision.

Summarizing, precision is a desired property related to the quality of the

instrument and the existing noise in the system. Accuracy and reliability are
more related to the probability of instrumentation failure. If the failure is not
total but it only constitutes a bias, accuracy is compromised. When a sensor
goes off-line, the failure is total and the signal is no longer present.
Software or Analytical Redundancy
In addition to direct measurements, there are indirect ways of determining

process variables. Therefore, we further classify data acquisition as follows:
r data acquisition supported solely by instrument readings
r data acquisition enhanced by software
When data acquisition is solely supported by instrument measurements, each
particular value of a variable of the system is directly associated to its source, the
instrument that measures it. It is well known that in a world not constrained by
cost, precise and failure-proof instrumentation would suffice to obtain precise
and reliable estimates. However, because in almost all cases, the probability of
a sensor failure is not negligible, redundancy is used as means of guaranteeing
data availability. With redundancy comes discrepancy from one instrument
to another measuring the same variable. This might not matter if the noise
associated with the signal is small enough. However, because the level of noise
is usually not small, readings have to be reconciled. Thus, redundancy is classi-
fied as:
r hardware redundancy
r software or analytical redundancy
Hardware redundancy is a term used when two or more sensors are used to
measure the same variable. This is the case of two thermocouples measuring the
temperature inside a vessel or a pipe in an environment where this temperature
is the same (for example, a steam boiler). Another example is the use of two
flowmeters installed to measure the same flow.
Software or analytical redundancy is a term used when a set of measurements
of different variables is supposed to satisfy a mathematical model. A simple
example of this is a unit with several input and output streams. If one measures
the flow rate of all streams with a single instrument per flow measured, there
is no hardware redundancy. However, because the sum of the input flow rates
has to be equal to the sum of the outputs, redundancy is now analytical. In
other words, two estimates are available for each flow rate, one from its direct
measurement and the other obtained by using the material balance equations,
which constitutes a conflict that needs to be resolved. Data reconciliation deals
with the task of determining the best statistical estimates of all these variables.
Quality of Data 7
Data Reconciliation
Because instrument readings are inaccurate and do not even obey basic con-
servation laws, there is a need to determine the best estimates out of a conflicting
set of readings. However, in many facilities, plant operators still assume that
readings are accurate enough for monitoring and control purposes. Because im-
balances create conflicts with production accounting clerks, balances have been
forced by the use of several heuristic techniques. For example, the elimination
or manual correction of measurements that in the experience of the operator are
less trustworthy has been the most popular. Data reconciliation techniques were
introduced in the 1980s particularly in the oil industry. Not only are these tech-
niques able to enhance the accuracy of plant data, but they are also capable of
detecting and filtering instrument bias and identifying leaks. Several books out-
line several aspects of data reconciliation (Mah, 1990; Madron, 1992; Veverka
and Madron, 1997; Romagnoli and Sánchez, 1999; Narasimhan and Jordach,
2000). In addition, hundreds of articles have been devoted to the problem.
Data have to be accurate and consistent. Accuracy is obtained by proper
instrument selection and calibration. Consistency is tested at a first level by
the plant operator and later by the process engineer. Underneath each of these
evaluations is a model based on empirical knowledge of the plant or on basic
laws of nature (mass and energy conservation). In the case of total mass conser-
vation, the mathematical form of this model is a simple linear set of equations.
When component balances are made, the set of equations is bilinear. The full
energy balances involving temperatures, compositions, flow rates and pressures
as variables are also nonlinear, especially because they include the evaluation of
thermodynamic properties. Thus, data reconciliation is a systematic way of per-
forming this assessment, taking into account the precision of each contributing
measurement and making use of statistical tools.
In addition to the value of the estimate of each measured variable, data recon-
ciliation is also capable of providing estimates for unmeasured variables. When
these unmeasured variables are process parameters, the technique received the
name of parameter estimation.
Various ways have been proposed to validate data without the need of using
data reconciliation. In principle, the term “soft sensor” was coined for the use
of existing process and laboratory data to infer the value of a certain measure-
ment (Martin, 1997; Hong et al., 1999). These soft sensors are based on neural
networks and wavelet decomposition.
Precision and Software/Analytical Redundancy
Software or analytical redundancy increases the precision of estimates. It

is well known that the precision of estimates produced by data reconcilia-
tion is higher than those associated with the data provided by the redundant
measurements. In other words, the standard deviation associated with the es-
timates is lower than those of the individual measurements. For production
accounting and planning, this reduction in the level of uncertainty is important.
In the oil industry, a small percentage gained in the uncertainty can lead to differ-
ent financial decisions, especially nowadays when more and more sophisticated
computer programs are used to make predictions and decisions.
Reliability and Software/Analytical Redundancy
If data are to be reliable, the instrumentation used has to be reliable too.

However, because balance equations and other more complex models can be
used for the estimation of variables, redundancy provides a way of enhancing
reliability. Therefore, through the use of data reconciliation, one can guarantee
that when a particular instrument measuring directly a variable fails, an estimate
of this variable is still available through the model. Thus, one can afford to use
less reliable instrumentation and reduce cost, as long as a whole scheme is in
place to perform the estimation through software.
Detection of Gross Errors in Instrumentation
In the absence of gross errors, data reconciliation reduces typically to a

simple optimization problem where the adjustments to the measurements are
minimized, requiring that the adjusted values satisfy a plant model, usually
material and energy balances. It is precisely the presence of gross errors or
outliers that makes the data reconciliation task difficult because they need to be
identified and eliminated.
Instrument malfunction is a general term that relates to situations that range
from miscalibration to total failure. In the absence of redundancy, miscalibration
or bias cannot be detected, unless the deviation is so large that it becomes
obvious. Redundancy, and especially software or analytical redundancy, is the
only way to contrast data and determine possible malfunctions of this sort by
filtering the data to detect and/or estimate or eliminate these biases. In addition,
process systems often have leaks that can also be detected through software
or analytical redundancy. This approach was called gross error detection by
one group of researchers (Romagnoli and Stephanopoulos, 1980; Mah and
Tamhane, 1982; Crowe et al., 1983; Crowe, 1994; Bagajewicz and Jiang, 1997,
1998) and lately sensor/data validation or signal reconstruction by another group
(Dunia et al., 1996; Qin et al., 1997; Tay, 1996; Tham and Parr, 1996; and
others).
Identification of gross errors is performed by using different techniques.
The most popular techniques are those rooted on statistical hypothesis testing.
In these tests, the null hypothesis is that there are no gross errors, whereas
the alternative hypothesis is that at least one gross error exists in the set of
Quality of Data 9
measurements considered. The set under analysis could be the whole system
(global tests), a particular measurement (measurement tests) and/or a particular
unit (nodal or constraint tests). The different contributions to this problem are
briefly reviewed in the next paragraph.
Reilly and Carpani (1963) were the first to propose the global test and the
nodal test. The measurement test was proposed by Mah and Tamhane (1987)
and Crowe et al. (1983). Almasy and Sztano (1975) proposed the maximum
power measurement test, which was the object of several follow-up articles.
Madron (1985) proposed an alternative method of about the same power as
the maximum power test. The nodal test was introduced by Reilly and Carpani
(1963) and discussed later by Mah et al. (1976). New techniques relying on
the generalized likelihood ratio were proposed by Narasimhan and Mah (1987,
1988), which were later proven equivalent to the measurement test. Johnston
and Kramer (1995) proposed a bayesian approach to the whole problem of
reconciliation and gross error detection. Later, principal component analysis
(PCA) was applied to gross error testing by Tong and Crowe (1995). Since
then, a large number of PCA-based sensor validation techniques have been
presented (Dunia et al., 1996; Qin et al., 1997; Tay, 1996; and others). Finally,
Rollins and Davis (1992) introduced an unbiased estimation technique that uses
Bonferroni tests.
In an effort to depart from the statistically based techniques, neural network
techniques have also been proposed as tools for gross error detection (Gupta and
Narasimhan, 1993; Himmelblau, 1994; Karjala and Himmelblau, 1994, 1996;
Reddy and Mavrovouniotis, 1998). Each of the aforementioned methods has
its own problems, especially when several gross errors are present at the same
time.
Estimation of Gross Errors
Once a gross error or a set of gross errors is detected, the following two
questions need to be answered:
r Should the corresponding measurements be eliminated or should they be
independently estimated?
r How is a gross error related to a bias, especially when averages of several
measurements are reconciled? How to filter process variations from gross
errors? Should historical data be used?
In trying to answer these questions, several methods have been developed
throughout the past years. Three kinds of strategies help identify multiple gross
errors:
r Serial elimination (Ripps, 1965; Serth and Heenan, 1986; Rosenberg et al.,
1987), which identifies one gross error at a time by using some test statistic
and eliminates the corresponding measurement until no gross error is de-

tected. Different existing commercial software use this technique based on
the measurement test and recently by means of principal component tests.
r Serial compensation (Narasimhan and Mah, 1987), which identifies the gross
error and its size, compensates the measurement and continues until no error
is found.
r Simultaneous or collective compensation (Keller et al., 1994; Kim et al.,
1997; Sánchez and Romagnoli, 1994; Sánchez, 1996; Sánchez et al., 1999),
which proposes the estimation of all gross errors simultaneously. In addition,
Jiang and Bagajewicz (1999) proposed a serial identification with collective
compensation strategy (SICC) for dynamic systems and steady-state cases.
Finally, the unbiased estimation technique (UBET), proposed by Rollins and
Davis (1992), makes the identification first and then a simultaneous estima-
tion.
Many researchers have evaluated the performance of these approaches. Se-

rial elimination is simple but has the drawback of losing redundancy and is not
applicable to gross errors that are not directly associated with measurements,
e.g., leaks (Mah, 1990). Serial compensation is applicable to all types of gross
errors. However, its results completely depend on the accuracy of estimation for
the size of gross errors (Rollins and Davis, 1992). Collective compensation is
considered to be more correct but computationally too intensive and impracti-
cal (Keller et al., 1994). Nevertheless, the results from simultaneous/collective
compensation methods seem to be more accurate. For example, the simulta-
neous estimation method developed by Sánchez (1996) and later modified by
Sánchez et al. (1999) is very accurate. However, it is still not suitable for large
systems, because it becomes combinatorially expensive.
Commercial software vendors claim that the introduction of data reconcili-
ation technology has produced millions of dollars of savings. Oil losses come
from inaccurate computations of transactions and from tank evaporation and
actual spills. Imbalances between input and output measurements in refineries
can be of the order of 0.5–1% (Miles and Jellfs, 1988) or even higher, depending
on the quality of the instrumentation. In a typical refinery processing 150,000
barrels of crude per day this represents close to a million dollars per year in
oil that is unaccounted for. Any project aiming at the identification of these
losses will realize this type of savings. However, losses are not the only issue.
Not only biased instrumentation can lead to underestimation of produced goods
but can also affect monitoring, thus reducing the efficiency of plant operation.
At the same time, proper identification of biased instrumentation allows better
maintenance and prevents accidents. Some of these features, such as leak de-
tection, are not available commercially and will represent a novel technological
application.
References 11
Detection of Process Faults
Faults propagate throughout the process, altering the readings of instruments

(pressures, temperatures, flow rates etc). Thus, sensors should be able to deter-
mine departures from normal operation. The quality of data needed for this task
is, therefore, different. Although in normal operation the focus is on accuracy
of estimation of key variables, it shifts to observability and proper diagnosis of
faults in abnormal situations. Thus, one of the problems is to prevent the false
diagnosis because of faulty instrumentation.
REFERENCES
Almasy G. A. and T. Sztano. Checking and Correction of Measurements on the Basis of Linear
System Model. Problems of Control and Information Theory, 4, 57 (1975).
Badell M. and L. Puigjaner. Short-Term Planning from the Business Level in ERP Systems with
Vertical Integration. Paper 240j. Annual AIChE Meeting, Miami (1998).
Badell M., A. Espuna and L. Puigjaner. Using ERP Systems with Budgeting Optimization Tools for
Investment Decision Making. Paper 239b. Annual AIChE Meeting, Miami (1998).
Bagajewicz M. and Q. Jiang. An Integral Approach to Dynamic Data Reconciliation, AIChE J.,
43, pp. 2546 (1997).
Bagajewicz M. and Q. Jiang. Gross Error Modeling and Detection in Plant Linear Dynamic Rec-
onciliation. Comp. & Chem. Eng., 22, 12, pp. 1789–1810 (1998).
Benson R. S. Computer Integrated Management: An Industrial Perspective on the Future. Comp.
Chem. Eng., 19, Suppl., pp. S543–S551 (1995).
Brydges J., A. Hrymak and T. Marlin. Real Time Optimization of a FCC Recovery Section. Pro-
ceedings of FOCAPO (1998).
Bunch P. R. Integration of Planning and Scheduling Systems with Manufacturing Processes. Paper
235g. Annual AIChE Meeting, Miami (1998).
Crowe C. M. Data Reconciliation. Progress and Challenges. Proceedings of the Process Systems
Engineering (PSE) International Symposium, Kyongju, Korea (1994).
Crowe C. M., Y. A. Garcia Campos and A. Hrymak. Reconciliation of Process Flow Rates by
Matrix Projection. I. The Linear Case. AIChE J., 29, pp. 818 (1983).
Dunia R., J. Qin, T. F. Edgar, T. J. McAvoy. Identification of Faulty Sensors Using Principal
Component Analysis. AIChE J., 42, pp. 2797–2812 (1996).
Forbes J. F. and T. E. Marlin. Design Cost: A Systematic Approach to Technology Selection for
a Model-Based Real-Time Optimization Systems. Comp. & Chem. Eng., 20, No 6/7, 717–734
(1996).
Grosdidier P. Understand Operation Information Systems. Hydrocarbon Processing. Sept. (1998).
Gupta G. and S. Narasimhan. Application of Neural Networks for Gross Error Detection. Ind. Eng.
Chem. Res., 32, 8, pp. 1651 (1993).
Harkins B. Turning Knowledge into Profit. Chemical Engineering, March (1999).
Himmelblau D. M. Rectification of Data in a Dynamic Process Using Artificial Neural Networks.
Proceedings of the Process Systems Engineering (PSE) International Symposium, Kyongju,

Korea (1994).
Hong S. J., J. H. Jung, C. Han. A Design Methodology of a Soft Sensor Based on Local Models.
Comp. & Chem. Eng. Suppl., pp. S351–S354 (1999).
Jiang Q. and M. Bagajewicz. On a Strategy of Serial Identification with Collective Compensation
for Multiple Gross Error Estimation in Linear Data Reconciliation. Ind. & Eng. Chem. Res., 38,
5, pp. 2119–2128 (1999).
Johnston L. P. M. and M. A. Kramer. Maximum Likelihood Data Rectification. Steady State Systems.
AIChE J., 41, 11 (1995).
Karjala T. W. and D. M. Himmelblau. Dynamic Data Rectification by Recurrent Neural Networks
vs. Traditional Methods. AIChE J., 40, 11, pp. 1865 (1994).
Karjala T. W. and D. M. Himmelblau. Dynamic Rectification of Data via Recurrent Neural Nets
and the Extended Kalman Filter. AIChE J., 42, pp. 22–25 (1996).
Keller J. Y. M. Darouach and G. Krzakala. Fault Detection of Multiple Biases or Process Leaks in
Linear Steady State Systems. Comp. & Chem. Eng., 18, pp. 1001 (1994).
Kim I., M. S. Kang, S. Park and T. F. Edgar. Robust Data Reconciliation and Gross Error Detection:
The Modified MIMT Using NLP. Comp. & Chem. Eng., 21, 7, pp. 775–782 (1997).
Koolen J. L. A. Plant Operation in the Future. Comp. & Chem. Eng., 18, pp. S477–S481 (1994).
Kramer M. A. and R. S. H. Mah. Model-Based Monitoring. FOCAPO Proceedings, Crested Butte,
(1993).
Lauks U. E., R. J. Vanbinder, P. J. Valkenburg and C. van Leeuwen. On-Line Optimization of an
Ethylene Plant. ESCAPE-1, Supplement to Comp. & Chem. Eng., 16, Suppl., pp. S213–S220
(1992).
Madron, F. A New Approach to the Identification of Gross Errors in Chemical Engineering Mea-
surements. Chem. Eng. Sci., 40, pp. 1855 (1985).
Madron F. Process Plant Performance, Measurement Data Processing for Optimization and
Retrofits. Ellis Horwood, West Sussex, England (1992).
Mah R. S. H. Chemical Process Structures and Information Flows. Butterworths, Stoneham
(1990).
Mah R. S. H., G. M. Stanley and D. M. Downing. Reconciliation and Rectification of Process Flow
and Inventory Data. Ind. Eng. Chem., Process Des. Dev., 15, 1 (1976).
Mah R. S. H. and A. C. Tamhane. Detection of Gross Errors in Process Data. AIChE J., 28, 828
(1982).
Mah R. S. H. and A. Tamhane. Detection of Gross Errors in Process Data. AIChE J., 33, pp. 1514
(1987).
Martin G. Consider Soft Sensors. Chem. Eng. Prog., July (1997).
Miles J. and P. A. M. Jelffs. Computer Aided Loss Investigation and Monitoring. The Second Oil
Loss Conference. P. A. M. Jelffs, editor. John Wiley & Sons, Ltd. (1988).
Mullick S. Rigorous On-Line Model (ROM) for Crude Unit Planning, Scheduling, Engineering and
Optimization. AIChE Spring Meeting (1993).
Narasimhan S. and C. Jordache. Data Reconciliation & Gross Error Detection. An Intelligent Use
of Process Data. Gulf Publishing Company, Houston, TX (2000).
Narasimhan S. and R. S. H. Mah. Generalized Likelihood Ratio Method for Gross Error Identifi-
cation. AIChE J. 33, 9, pp. 1514–1521 (1987).
References 13
Narasimhan S. and R. S. H. Mah. Generalized Likelihood Ratios for Gross Error Identification in
Dynamic Processes. AIChE J., 34, pp. 1321 (1988).
Natori Y. and I. B. Tjoa. To Innovate Chemical Plant Operation by Applying Advanced Technology
and Management. Proceedings of FOCAPO (1998).
Pelham R. and C. Pharris. Refinery Operations and Control: A Future Vision. Hydrocarbon Pro-
cessing, July (1996).
Qin S. J., H. Yue and R. Dunia. Self-Validating Inferential Sensors with Application to Air Emission
Monitoring. Ind. Eng. Chem. Res., 36, 5, pp. 1675–1685 (1997).
Reddy V. N. and M. L. Mavrovouniotis. An Input-Training Neural Network Approach for Gross
Error Detection and Sensor Replacement. Inst. of Chem. Eng. Trans IchemE, 76, part A, May
(1998).
Reilly P. M. and R. E. Carpani. Application of Statistical Theory of Adjustment to Material Balances.
13th Can. Chem. Eng. Conf., Montreal, Que (1963).
Ripps D. L. Adjustment of Experimental Data. Chem. Eng. Progr. Symp. Ser. 55, 61, 8–13 (1965).
Rollins D. K. and J. F. Davis. Unbiased Estimation of Gross Errors in Process Measurements.
AIChE J., 38, 4, pp. 563–572 (1992).
Romagnoli J. and M. Sánchez. Data Processing and Reconciliation for Chemical Processes Oper-
ations. Academic Press (1999).
Romagnoli J. and G. Stephanopoulos. On the Rectification of Measurement Errors for Complex
Chemical Plants. Chem. Eng. Sci. 35, 5, pp. 1067–1081 (1980).
Rosenberg J., R. S. H. Mah and C. Iordache. Evaluation of Schemes for Detecting and Identifying
Gross Errors in Process Data. Ind. Eng. Chem. Res., 26, pp. 555 (1987).
Sánchez M. Ph.D. Thesis. Universidad Nacional del Sur (Argentina) (1996).
Sánchez M. and J. Romagnoli. Monitoreo de Procesos Continuos: Análisis Comparativo de
Técnicas de Identificación y Cálculo de Bias en los Sensores. AADECA 94-XIV Simposio
Nacional de Control Automático, Argentina (1994).
Sánchez M., J. Romagnoli, Qiyou Jiang and M. Bagajewicz. Simultaneous Estimation of Biases
and Leaks in Process Plants. Comp. & Chem. Eng., 23, 7, 841–858 (1999).
Serth R. and W. Heenan. Gross Error Detection and Data Reconciliation in Steam Metering Systems.
AIChE J., 32, pp. 733 (1986).
Smith O. Development and Benefits of On-Line Process Modeling Applications. AIChE Spring
Meeting (1996).
Swanson I. and R. Stewart. Towards Full Plant-Wide Management and Control. PetrochemAsia 94
International Conference, Singapore, Nov. (1994).
Tay M. E. Keeping Tabs on Plant Energy and Mass Flows. Chem. Eng., September (1996).
Tham M. T. and A. Parr. Succeed at On-Line Validation and Reconstruction of Data. G. F. Nalven,
editor. AIChE, New York (1996).
Tong H. and C. M. Crowe. Detection of Gross Errors in Data Reconciliation by Principal Compo-
nent Analysis. AIChE J., 41, 7, pp. 1712 (1995).
Veverka V. V. and F. Madron. Material and Energy Balances in the Process Industries. Elsevier
(1997).
Zhang Z., P. R. Pike and T. Hertwig. An Approach to On-Line Optimization of Chemical Plants.
Comp. Chem. Eng., 19, Suppl., pp. S305–S310 (1995).
CHAPTER 2
Instrumentation Design Goals
INTRODUCTION
In this chapter, the goals of instrumentation design are presented in further

detail. First, the concept of key variables is introduced. Then, different types
of criteria for the selection of key variables are presented. Finally, the sensor
network design goals are discussed in detail.
MEASURED AND KEY VARIABLES
Traditionally, the fields of monitoring and process control relied heavily on

an a priori selection of measured variables. The same can be said about the
emerging field of parameter estimation in on-line optimization. The traditional
concept used is to measure directly the variable whose value one desires to
estimate. When one departs from this concept, a new view of the field of sensor
network design emerges.
Software or analytical redundancy can provide better and more reliable esti-
mates through the use of steady-state and dynamic models. When the variable
of interest is measured, its estimate is improved. Thus, in general, one can state
that measured variables are not necessarily the same as the key variables. This
prompts the following definitions:
Definition: Key variables are the variables for which their estimated values
are of interest for control, monitoring and parameter estimation.
Definition: Measured variables are the variables whose measurements al-

low reliable and accurate estimates of key variables.
15
16 INSTRUMENTATION DESIGN GOALS
Thus, the optimal choice of measurements can now be performed by mak-

ing use of software or analytical redundancy while guaranteeing reliabil-
ity/availability and good gross error filtering properties. This is the objective
of this book.
In the next sections, the selection of key variables in the field of process
monitoring control and alarm is reviewed.
SELECTION OF MONITORING VARIABLES
Monitoring of a process is performed for several purposes. Table 2.1 summa-

rizes some of the monitoring purposes and the variables of interest. No compre-
hensive systematic methodology exists to optimally select these variables. Nev-
ertheless, singular value decomposition was used for this purpose in the context
of parameter estimation in on-line optimization (Krishnan et al., 1992a, b).
Parameter estimation is based on a model. For example, column efficiencies,
heat exchanger heat transfer coefficients and flash vaporization efficiencies can
only exist if a model of the associated equipment exists. Therefore, mathemati-
cal expressions derived from these models are used to estimate the parameters.
In this regard, as long as these equations can be solved, parameter estimation
poses no other challenges than their sensitivity to the measurement of some
variables. These problems can be mitigated by using software or analytical
redundancy.
In on-line optimization, there is an additional problem: the sensitivity of the
optimum to the parameters. This has, in turn, an effect on the desired accuracy
TABLE 2.1. Key Variables in Monitoring.
Monitoring Goal Key Variables
Safety Fault tree analysis, process hazard analysis (PHA) and

hazard and operability studies (HAZOP) determine
chains of events and variables to monitor to prevent
accidents. This information is, in turn, used to assess
fault occurrence on-line.
Production accounting Flow rate and composition of input and output process
streams.
Quality control Composition of streams, especially products, or
variables that indirectlyinfluence them.
Parameter estimation These parameters are normally not possible to measure
directly. Examples are heat exchanger fouling and tray
efficiencies that are used in on-line optimization.
Environmental impact Environmental consciousness and regulations dictate
that certain variables should be monitored on a regular
basis to assess a possible damage on the environment.
of the estimates of these parameters. Krishnan et al. (1992a, b) presented an

analysis of which parameters should be used. They also discuss in detail how
plant-wide models can be used to estimate the needed parameters, thus present-
ing an alternative to the traditional estimation methods that focus on measure-
ments around the units where parameters are used. In addition, they introduce a
methodology to select measurements. This methodology is based on a combi-
nation of singular value analysis (SVA) and sensitivity analysis. The procedure,
however, does not rely on cost analysis.
SELECTION OF KEY VARIABLES IN CONTROL
Although most of the design methods of instrumentation networks for mon-

itoring purposes rely on steady-state models, the design of control schemes is
based on dynamic models. The accuracy needed in the variables of interest is
related to control loop performance, which many times supersedes cost. There
are some guidelines and models to determine these variables. Although they
are typically suggested by the designer, some heuristic methods exist (Seborg
et al., 1989; Ogunnaike and Ray, 1994). Mathematical programming techniques
exist for particular cases such as the decentralized discrete regulator problem
with output feedback (Oloomi and Sawan, 1989).
To design a control system, an appropriate number of controlled and ma-
nipulated variables need to be selected. The number of manipulated variables
cannot exceed the degree of freedom, which is defined mathematically as the
number of process variables minus the number of independent equations. In
addition, if offsets cannot be tolerated, the number of manipulated variables
should be larger than the controlled variables. In the following sections, guide-
lines to select manipulated and key variables are first discussed. The technique
of singular value analysis to determine key variables is reviewed next followed
by an overview of the concept of inferential control.
Classification of Process Variables in Control Theory
In control system design theory, process variables are classified as either

inputs or outputs, as is shown in Figure 2.1.
The output variables are process variables that ordinarily are associated with
exit streams of a process or conditions inside units, such as compositions,
temperatures, levels and flow rates. To satisfy plant and control objectives, a
subset of the output variables is selected as variables to be controlled (i.e.,
controlled variables). Typically, the inputs are associated with inlet streams
(e.g., feed composition or feed flow rate) or environmental conditions (e.g.,
ambient temperature). In general, controlled variables are measured on-line,
and their measurements are used for feedback control. However, this is not
x1 y1
x2 y2
Input x3 PROCESS y3 Output
• •
variables • • variables
xM yN
FIGURE 2.1. Process variables.
always the practice. The case in which measured and controlled variables do not
coincide has received the name of inferential control (Brosilow and Tong, 1978),
which is primarily used in distillation to infer composition from flow rate and
temperature measurements. Measured variables that are not controlled variables
received the name of secondary measurements (Weber and Brosilow, 1978) in
the control literature. As it was discussed at the beginning of this chapter, these
concepts are generalized in this book, so that all measurements contribute to
obtain good estimates of key variables.
Selection of Controlled Variables
The consideration of control objectives has led to a number of guidelines

for the selection of controlled variables from the available output variables. We
now reproduce a list presented by Seborg et al. (1989).
r Guideline 1: Select variables that are not self-regulating. A common example
is liquid level in a storage vessel with a pump in the exit line.
r Guideline 2: Choose output variables that may exceed equipment and oper-
ating constraints (e.g., temperatures, pressures and compositions).
r Guideline 3: Select output variables that are a direct measure of product
quality (e.g., composition and refractive index) or that strongly affect it (e.g.,
temperature and pressure).
r Guideline 4: Choose output variables that seriously interact with other con-
trolled variables. The steam header pressure for a plant boiler that supplies
several downstream units is an example of this type of output variable.
r Guideline 5: Choose output variables that have favorable dynamic and static
characteristics. Ideally, there should be at least one manipulated variable that
has a significant direct and rapid effect on each controlled variable.
Key variables for control may not be state variables but, rather, some combi-
nation of them that indicates the system performance, such as product recovery
ratios, pressure ratios in compressors, process loads etc.
Selection of Manipulated Variables
Based on the plant and control objectives, a number of guidelines have been
proposed for the selection of manipulated variables from among the input vari-
ables:
r Guideline 6: Select inputs that have large effects on the controlled variables.
Typically, the steady-state gain between the manipulated and controlled vari-
able should be as large as possible.
r Guideline 7: Choose inputs that rapidly affect the controlled variables. In
other words, avoid time delays.
r Guideline 8: The manipulated variables should affect the controlled variables
directly rather than indirectly. For example, in a system heated by steam injec-
tion in a jacket, the exit temperature of the process stream is better controlled
by manipulating the inlet steam flow, rather than the exit condensate, be-
cause the steam rate has more direct effect on the pressure of the shell and
consequently on the saturation temperature.
r Guideline 9: Avoid recycling disturbances. That is, it is preferable not to
manipulate an inlet or a recycle stream, which can be avoided by manipulating
a utility stream.
These guidelines may be in conflict, especially when a variable has a larger

steady-state gain (Guideline 6) but slower dynamics (Guideline 7), which calls
for a trade-off analysis.
Selection of Measured Variables
The recommendations of control practitioners are that output variables used

as controlled variables are the obvious choice. They also recommended to
additionally measure selected input variables and output variables, because
recorded measurements of manipulated inputs provide useful information for
tuning controllers and troubleshooting control loops. Finally, measurements of
disturbances in inputs can be used in feed-forward control schemes. In choosing
outputs to measure and in locating measurement points, both static and dynamic
considerations are important. Seborg et al. (1989) suggest:
r Guideline 10: Reliable, accurate measurements should be considered.
r Guideline 11: Select measurement points that have an adequate degree of
sensitivity. For example, in distillation columns a product composition is often
controlled indirectly by regulating a temperature near the end of the column
if an analyzer is not available. For high-purity separations, the location of the
temperature measurement point is very important.
r Guideline 12: Select measurement points that minimize time delays and time
constraints. Reducing dynamic lags and time delays associated with process
measurements improves closed-loop stability and response characteristics.
r Guideline 13: Select measurements that can filter disturbances appropriately.
In distillation columns, when temperature is used to assess composition,
disturbances due to the presence of components that are not anticipated in
the feed can make control inappropriate (Downs and Ogunnaike, 1994).
In addition, Friedmann and Moore (1972) pointed out that if key variables
cannot be measured directly or the available instruments do not provide the
desired accuracy, then other measurements are needed to provide an estimate
of the key variables through software or analytical redundancy.
Singular Value Analysis
Singular value decomposition has been suggested as a more systematic way

of selecting controlled and manipulated variables. This procedure allows the
determination of the best multiloop technique by capturing the “decoupled”
open loop gains of the system (Arkun and Ramkrishnan, 1984; Lau et al., 1985;
Moore, 1986, 1987; Skogestad and Postlethwaite, 1996, Chapter 10).
Consider a process that has n manipulated variables. We assume that a steady-
state process model is available and that it has been linearized as follows:
C = KM (2.1)
where C is the vector of n controlled variables, M is the vector of m manipulated
variables (both expressed as deviation variables), and K is the steady-state gain
matrix.
Normally, n = m and one requirement is that K should be nonsingular if the
controlled variables are to be independently regulated. A small eigenvalue of K
indicates that larger changes in one or more manipulated variables are needed
to control the process.
The singular values of K are nonnegative numbers defined as the positive
square roots of the eigenvalues of K T K . The first r singular values are positive
numbers where r is the rank of matrix K T K . The remaining n − r singular
values are zero. Finally, the condition number (CN) is defined as the ratio of
the largest and smallest nonzero singular values (Moore, 1986, 1987). If K is
singular, then it is ill-conditioned and by convention CN = ∞. The concept of
a condition number can also be extended to nonsquare matrices (Klema and
Laub, 1980). Thus, the condition number is a positive number that provides a
measure of how ill-conditioned the gain matrix is. Processes with poorly con-
ditioned K matrices tend to require large changes in the manipulated variables
to influence the controlled variables. A justification of this statement is given
by Seborg et al. (1989, p. 691).
L
79 Distillate
73 D=1,277 lb-mole/hr
1,763 lb-mole/hr zD=93 % C3
64 Sidedraw
F64=828 lb-mole/hr
z64=90 % i-C4
38
882 lb-mole/hr
Sidedraw
15 F15=883 lb-mole/hr
z15=90 % n-C4
5 V
1,764 lb-mole/hr
Bottoms
B=1,421 lb-mole/hr
zB=95 % i-C5
FIGURE 2.2. Four components column. Adapted from Seborg et al., 1989.
Roat et al. (1986) analyzed the choice of manipulated variables for a complex,
four-component distillation column (Figure 2.2). The choices are taken from
Seborg et al. (1989) and reproduced in Table 2.2. The four components are
propane, isobutane, n-butane, and isopentane.
There are six possible manipulated variables. In addition, ratios of these
variables are also permissible. Table 2.2 shows the condition numbers for six
schemes that have been evaluated for the column. Because three of the strategies
have roughly the same low CN , these were subjected to further evaluation using
dynamic simulation. Based on simulation results, the best control strategy was
number 4.
When the process gain matrix is nonsquare, the singular value decomposition
(SVD) is obtained as follows:
K = U 6V T (2.2)
where U is a n × m orthonormal matrix. Its columns are called “left singular

vectors.” Matrix V is also an m × m orthonormal matrix, with columns called
“right singular vectors.” Finally, 6 is an m × m diagonal matrix of scalars called
the “singular values,” which are organized in descending order.
TABLE 2.2. Condition Numbers for the Gain Matrices.
Controlled Variables
z D = Mole fraction of propane in distillate D
z 64 = Mole fraction of isobutane in tray 64 side-draw
z 15 = Mole fraction of n-butane in tray 15 sidedraw
z B = Mole fraction of isopentane in bottoms B
Possible Manipulated Variables
L = Reflux flow rate B = Bottoms flow rate
D = Distillate flow rate F64 = Sidedraw flow rate at tray 64
V = Steam flow rate F15 = Sidedraw flow rate at tray 15
Strategy Number Manipulated Variables Condition Number
1 L/D, F64 , F15 , V 9030

2 V/L , F64 , F15 , V 60100
3 D/V, F64 , F15 , V 116000
4 D, F64 , F15 , V 51.5
5 L , F64 , F15 , B 57.4
6 L , F64 , F15 , V 53.8
Note: In each control strategy, the first controlled variable is paired with the first manipulated
variable and so on. Thus, for strategy 1, x D is paired with L/D and x B is paired with V .
Moore (1987) presented an ethanol column example, where two manipulated

variables are chosen (distillate flow rate D and condenser heat load Q). The
process gain matrix is given by:
 
∂ T1 ∂ T1
∂D ∂Q 
 ∂T
 2 ∂ T2 


K =  ∂ .D ∂Q 
 . .. 

(2.3)
 . . 
∂T
n ∂ Tn 
∂D ∂Q
and describes the temperature sensitivity on each tray with respect to each
of the two manipulated variables. In his example, U is a 50 × 2 matrix, V
is a 2 × 2 matrix and the condition number is CN = 66.4. The article also
suggests two methods to locate the sensors needed. In this case, two temperature
measurements in trays need to be chosen. One of these methods is based on the
location of the principal components in the left singular vectors. The other is a
more detailed analysis based on a global search.
The most sensitive element (the largest) in each column of the left singular
matrix is chosen as the measured variable. After the sensor locations have been
selected, it is important to always check the singular value analysis (SVA) of the
smaller system consisting of the m × m system of manipulated and measured
variables selected in the previous step. The objective is to determine if the

condition number and the singular values are still roughly the same magnitude.
The condition number for this system can be worse or better than predicted by
the overall analysis including all the variables. If the condition number from the
partial analysis is considerably worse, a more detailed global method of sensor
selection may need to be considered. This is called the intersensitivity method.
The intersensitivity method is based on a global search of all possible sen-
sor combinations. A partial singular value analysis can be performed for each
combination of sensors that make physical sense. According to Moore (1987),
this procedure is more time consuming and in most cases does not yield results
that are significantly different from the principal component method.
Singular value analysis has also been suggested as part of the selection of op-
timization variables in on-line optimization schemes (Krishnan et al., 1992a,b;
Glemmestad et al., 1997). Glemmestad et al. (1997) also suggested the use of
an alternative procedure based on the study of the influence of the disturbances
in the objective function of on-line optimization for each measured set.
Inferential Control
In inferential control, process measurements that can be obtained rapidly are

used to infer the value of the controlled variable. For example, as analyzed in the
previous section, if the overhead product stream in a distillation column cannot
be analyzed on-line, sometimes measurement of the top tray temperature can be
used to infer the actual composition. For a binary mixture, the Gibbs phase rule
indicates that there is a unique relation between composition and temperature
if pressure is constant. Therefore, a thermodynamic equation can be used to
relate the temperature of the top tray to the overhead composition.
For the separation of multicomponent mixtures, approximate methods to es-
timate compositions must be used. Based on process models and plant data,
simple algebraic correlations can be developed to relate the mole fraction of the
heavy key component to several different tray temperatures (usually near the top
of the column). Then, the overhead composition can be inferred from available
temperature measurements and used in the control algorithm (Moore, 1986). If
necessary, the parameters in the correlation may be updated when new compo-
sition measurements become available. For example, if samples are sent to the
analytical laboratory once per hour, the correlation parameters can be adjusted
so that the predicted values agree with the measured values.
The concept of inferential control can be used for other process operations,
such as chemical reactors where composition is normally the controlled vari-
able. Selected temperature measurements can be used to estimate the outlet
composition when it cannot be measured on-line. However, when inferential
control does not perform satisfactorily, incentive exists to introduce other on-
line measurements for feedback control. Consequently, there is considerable
interest in the development of new instrumentation, such as process analyzers,

which can be used on-line and which exhibit very short response times.
Havre and Skogestad (1996) and Cao et al. (1997) introduced the concept of
output effectiveness (OE) for measurement selection. This OE is defined as the
square root of the diagonal elements of scaled open loop transfer function mul-
tiplied by its generalized (Moore-Penrose) inverse. The details of this procedure
and its justification can be found in the original article and in Cao et al. (1998),
who extended the use of the OE to secondary measurements.
Finally, practitioners and researchers are starting to realize that techniques
such as multiple regression, principal component analysis, partial least squares
and neural networks, which are popular in monitoring, can be used effectively
in control (Deshpande et al., 1996; Soroush, 1998).
Interaction of Design and Control
It is the current state of the art that process design precedes the design of the
control system and they constitute separate activities. In systems control theory,
it is a well-known fact that this approach has serious limitations because the
plant design determines the process dynamic characteristics. To ameliorate
these deficiencies, Luyben and Floudas (1994) proposed singular value analy-
sis, as well as other measures to construct a multiobjective framework for the
design of process systems and their control schemes simultaneously. Downs
and Ogunnaike (1994) also discuss the design for controllability and opera-
tions. Covering all these design techniques is outside the intended scope of this
book. Nevertheless, as this field continues to develop, it will produce improved
lists of key variables for which a sensor network will have to be built.
SELECTION OF MEASURED VARIABLES FOR FAULT DIAGNOSIS
Alarms are important elements in monitoring. They can be broadly classified

as quality loss alarms and safety alarms. In the first case, alarms are set to
monitor if a particular variable is within an expected range. Hardware and
software redundancy of measurements is thus of importance. In addition, data
reconciliation and, in particular, gross error detection become a must. In the
case of safety alarms, the problem is similar, but the goals are not quite the same.
If quality is lost, the system can continue functioning, and this quality loss can
be measured by a monetary loss. However, when safety is compromised, then
the equipment or even the plant should be shut down.
Faults originate somewhere in the process, and the disturbances propagate to
places where a sensor can detect them. Several techniques have been developed
to establish the existence of a fault (observability), determine its nature (diag-
nosis) and take appropriate action (alarm and or shutdown), assuming that the
instrumentation is free of biases and is well calibrated. A few books cover this
Instrumentation Design Goals 25
field in detail (Himmelblau, 1978; Pau, 1981; Gertler, 1998). Several techniques
of diverse origins are used nowadays: fault trees, cause-effect digraphs, neural
networks, knowledge-based systems and model-based statistical approaches
such as principal component analysis and partial least squares. The field is very
much intertwined with the data reconciliation field to the extent that the same
techniques are used sometimes in both fields under different names.
Fault diagnosis poses an even more ambitious goal than simple control and
instrument malfunction diagnosis. Indeed, for each pattern of signals that are
suspect of process malfunction, one needs to be able to sort if the pattern is
being created by instrument malfunction or if it is a genuine process fault, or
if it is a combination of both. With valuable production at stake, a shutdown
based on false alarms is an issue of increasing concern. For this reason, there
is increasing interest in the use of simultaneous instrument and process fault
detection.
INSTRUMENTATION DESIGN GOALS
Several goals exist for every designer of a set of instruments in a plant.

r Cost: The simplest objective function used in design has been the sum of the
annualized capital investment cost and the operating cost. The simplest de-
sign objective in the case of sensor networks has been the overall annualized
investment cost. Operating costs that are related to electricity consumption
have always been neglected, leaving maintenance costs as the primary com-
ponents.
r Estimability: This term was coined to designate the ability to estimate a
variable using hardware or software, and it will be formally introduced in
Chapter 5. In the absence of more sophisticated goals, estimability can be used
as a substitute goal. The connection between redundancy and estimability will
be explored in detail in Chapter 5.
r Precision: In many cases, specific levels of precision can be developed for
particular variables in a system. For example, accounting might be interested
in closing the material balance within a certain percentage or even request
that the precision of each stream crossing battery limits has a particular value.
Parameter estimation is also a software-based activity for which the parameter
precision depends on the set of sensors used.
r Reliability: If information on the frequency of instrumentation failure is
available, it is possible to set as a design goal that the software availabil-
ity or reliability of variables be larger than prespecified thresholds.
r Gross error robustness: A robust sensor network is defined as a network that
has good performance in the presence of gross errors. This includes:
— Gross error detectability: This term is coined after the ability (or more
precisely the statistical power) to detect gross errors.
— Residual precision: After gross errors are detected, the level of precision
associated with the estimates deteriorates, usually because the variables
with gross errors are eliminated from the set. Precision availability is a
level of precision achieved when data reconciliation is performed after a
certain number of gross errors are eliminated.
— Gross error resilience: Undetected gross errors have a smearing effect on
the estimates obtained by data reconciliation. This effect can be quantified,
and a robust network is capable of containing this smearing to a desired
small level.
r Maintenance: Until recently, the costs associated with maintenance have not
been considered part of the design of sensor networks. Two aspects are of
importance: the cost of preventive and corrective maintenance.
r Fault diagnosis and alarms: Only a few articles addressing alarm needs in
this context have been published. Tsai and Chang (1997) proposed to design
alarm logic minimizing the expected loss, which is assessed in terms of
probabilities. This line of work builds up on the early analysis performed by
Inoue et al. (1982) and Kohda et al. (1983). Lately, Raghuraj et al. (1999)
used signed directed graphs to propose optimal sensor networks that would
guarantee fault observability.
Although all these goals seem equally important, cost has been the traditional
objective function used in design. However, in some cases, other goals, such
as precision, have also been used as objective functions in design. This book
concentrates on the minimum cost model. The simplest of these models is:
Minimize {Investment Cost}

s.t.
Desired Level of Estimability of Key Variables
The unknowns of this problem are the number and location of measurements
as well as the type, precision, and reliability of each instrument. Throughout
the book, other constraints will be added until a complete minimum cost model
is presented. This model is the following:
Minimize {Total Cost}

s.t.

 Desired Level of Estimability of Key Variables


 Desired Level of Precision of Key Variables

 Desired Level of Reliability of Key Variables


Desired Level of Gross-Error Robustness
Upgrading of Instrumentation 27
where the total cost includes the maintenance cost, which regulates the avail-
ability of variables, a concept that substitutes reliability when the system is
repairable.
The above complete model covers one important aspect of monitoring, which
is to filter relatively good data for monitoring and eventually control purposes.
However, it does not cover the optimal design of a sensor networks and a logic
for alarm systems. Even though important attempts are being made to address
this issue (Tsai and Chang, 1997; Raghuraj et al., 1999), a model based on
cost-efficient alarm design is yet to be produced. Likewise, the aforementioned
direct incorporation of control performance measures as additional constraints
of a cost-optimal model is a matter that has not been investigated yet. Finally,
methods for cost-optimal instrumentation design corresponding to the imple-
mentation of several other monitoring procedures, like principal component
analysis (PCA), projection to latent structures (PLS), wavelet analysis, and
neural networks, among others, have not been yet proposed.
While the above model seems complete for process monitoring in the absence
of severe faults, once this latter is considered a more ambitious problem arises.
From the exclusive point of view of fault detection, the problem of the design
of instrumentation is as follows:

s.t.

 Desired Observability of Faults


 Desired Level of Resolution of Faults
 Desired level of Reliability of Fault Observation



Desired level of Gross-Error Robustness in the Sensor Network
The combination of both goals, that is, the design of a sensor network, capable
of performing estimation of key variables for production accounting, parameter
estimation for on-line optimization as well as fault detection, diagnosis and
alarm (all of this in the context of detecting and assessing instrument biases and
process leaks) is perhaps the long term goal of this field.
UPGRADING OF INSTRUMENTATION
When reconciliation methods and new control strategies gain acceptance

in practice by process and instrumentation engineers, the issue of instrument
revamping becomes critical. Replacing an existing measurement device with a
more accurate one is the first solution offered to the problem of increasing data
accuracy. However, this is not always the most economical option.
Industry wants to determine how a limited amount of resources (typically

fixed capital) can be allocated to maximize certain goals, typically accuracy. In
addition, certain resources can be reallocated, as is the case of laboratory time
and chemicals for analysis, thermocouples, pressure gauges and even flowme-
ters. There is increasing interest for this type of problems. A commercial soft-
ware (DATACON from Simulation Sciences, 1993) includes tools to perform
(albeit rudimentary) some of these studies. Recently, researchers from British
Petroleum and Imperial College studied how to perform optimal reallocation
(Alhéritière et al., 1997, 1998a,b). This issue will be covered throughout the
book and with especial emphasis in Chapter 9.
NOMENCLATURE
c: vector of controlled variables

CN : condition number
D: distillate flow rate
Fi : Flow rate
m: vector of manipulated variables
Q: condenser heat load
U: matrix of left-singular vectors
V: matrix of right-singular vectors
xi : input variable
yi : output variable
zi : composition
Greek Letters
6: matrix of singular values
REFERENCES
Alhéritière C., N. Thornhill, S. Fraser and M. Knight. Evaluation of the Contribution of Refinery
Process Data to Performance Measures. AIChE Annual Meeting, Los Angeles (1997).
Alhéritière C., N. Thornhill, S. Fraser and M. Knight. Cost Benefit Analysis of Refinery Process
Data: Case Study. Comp. & Chem. Eng., 22, Suppl., pp. S1031–S1034 (1998a).
Alhéritière C., N. Thornhill, S. Fraser and M. Knight. Cost Benefit Analysis of Process Data in
Plant Performance Analysis. AIChE Annual Meeting, Miami (1998b).
Arkun Y. and S. Ramkrishnan. Structural Sensitivity Analysis in the Synthesis of Process Control
Systems. Chem. Eng. Sci., 39, 7 & 8, pp. 1167 (1984).
Brosilow C. and M. Tong. Inferential Control of Processes. II. The Structure and Dynamics of
Inferential Control Systems. AIChE J., 24, pp. 492–500 (1978).
References 29
Cao Y., D. Rossiter and D. H. Owens. Screening Criteria for Input and Output Selection. ECC97
CDROM, number 773, Brussels, Belgium (1997).
Cao Y., D. H. Owens and D. Rossiter. Output Effectiveness and Scaling Sensitivity for Secondary
Measurement Selection. Trans IChemE, 76, Part A, October (1998).
DATACON Manual. Application: Optimal Location of Instrument. Simulation Sciences (1993).
Desphande P. B., S. S. Yerrapragada, E. F Jacob, R. N. Mohanty and S. L. Joshi. Improve Control
with Software Monitoring Technologies. Hydrocarbon Processing, Sept. (1996).
Downs J. J. and B. Ogunnaike. Design for Control and Operability: An Industrial Perspective.
Proceedings of FOCAPO 94 (1994).
Friedmann P. G. and J. A. Moore. For Process Control. . . Select the Key Variable. Chemical
Engineering, June (1972).
Gertler J. J. Fault Detection and Diagnosis in Engineering Systems. Marcel Dekker, New York
(1998).
Glemmsatd B., S. Skogestad and T. Gundersen. On-Line Optimization and Choice of Optimization
Variables for Control of Heat Exchanger Networks. Comp. & Chem. Eng., 21, Suppl., pp. S379–
S384 (1997).
Havre K. and S. Skogestad. Input-Output Selection and Partial Control. Proc. IFAC’96. 13th World
Congress, San Francisco, CA, Vol. M, pp. 181–186 (1996).
Himmelblau D. Fault Detection and Diagnosis in Chemical and Petrochemical Processes. Elsevier,
Amsterdam (1978).
Inoue K, T. Kohda, H. Kumamoto, I. Takami. Optimal Structure of Sensor Systems with Two Failure
Modes. IEEE Transac. on Reliability, Vol. R-31, 1 (1982).
Klema V. C. and A. J. Laub. The Singular Value Decomposition: Its Computation and Some
Applications. IEEE Trans. Auto. Control, AC-35, 164 (1980).
Kohda T., H. Kumamoto, K Inoue. Optimal Shutdown Logic for Protective Systems. IEEE Transac.
on Reliability, Vol. R-32, 1 (1983).
Krishnan S., G. Barton and J. Perkins. Robust Parameter Estimation in On-line Optimization. Part
I. Methodology and Simulated Case Study. Comp. & Chem. Eng., 16, pp. 545–562 (1992a).
Krishnan S., G. Barton and J. Perkins. Robust Parameter Estimation in On-line Optimization. Part
II. Application to an Industrial Process. Comp. & Chem. Eng., 17, pp. 663–669 (1992b).
Lau H., J. Alvarez and K. F. Jensen. Synthesis of Control Structures by Singular Value Analysis.
AIChE J., 31, 3, pp. 427 (1985).
Luyben M. L. and C. A. Floudas. Analyzing the Interaction of Design and Control. 1. A Multiob-
jective Framework and application to Binary Distillation Synthesis. Comp. & Chem. Eng., 18,
10, pp. 933–969 (1994).
Moore C. Application of Singular Value Decomposition to the Design, Analysis and Control of
Industrial Processes. Proc. Am. Control Conf., pp. 643 (1986).
Moore C. Determining Analyzer Location and Type for Distillation Column Control. 14th An-
nual Meeting of the Federation of Analytical Chemistry and Spectroscopy Societies. Detroit
(1987).
Ogunnaike B. A, W. H. Ray. Process Dynamics, Modeling and Control. Oxford University Press
(1994).
Oloomi H. and M. E. Sawan. Optimal Sensor Placement in Decentralized Control Systems. Int. J.
Systems Sci., 20, 6, pp. 939–943 (1989).
Pau L. F. Failure Diagnosis and Performance Monitoring. Dekker, New York (1981).
Raghuraj R., M. Bhushan and R. Rengaswamy. Locating Sensors in Complex Chemical Plants
Based on Fault Diagnostic Observability Criteria. AIChE J., 45, 2, pp. 310–322 (1999).
Roat S. D., J. J. Downs, E. F. Vogel and J. E. Doss. The Integration of Rigorous Dynamic Model-
ing and Control System Synthesis for Distillation Columns: An Industrial Approach. Chemical
Process Control CPC-III, Edited by M. Morari and J. J. McAvoy. CACHE-Elsevier, New York
(1986).
Seborg D. E., T. F. Edgar, D. A. Mellichamp. Process Dynamics and Control. John Wiley (1989).
Skogestad S. and I. Postlethwaite. Multivariable Feedback Control, Analysis and Design. John
Wiley & Sons (1996).
Soroush M. State and Parameter Estimation and their Applications in Process Control. Comp. &
Chem. Eng., 23, pp. 229–245 (1998).
Tsai C. S. and C. T. Chang. Optimal Alarm Logic Design for Mass-Flow Networks. AIChE J., 43,
11, pp. 3021 (1997).
Weber R. and Brosilow C. The Use of Secondary Measurement to Improve Control. AIChE J., 24,
pp. 485–492 (1978).
CHAPTER 3
Instrumentation
INTRODUCTION
The purpose of this chapter is to provide an overview of the different types of

instrumentation used in process plants. The chapter focuses primarily on flow
rate, level, temperature and pressure instruments. Although density, concentra-
tion and pH measurements, among others, are sometimes performed on-line,
they are more often made in the laboratory. However, some on-line density
measurements are discussed. It is also customary to include valves as part of
instrumentation, but because the book does not discuss valve selection and lo-
cation, their analysis is omitted. Finally, a brief discussion on new emerging
tendencies in signal processing is added. A complete survey of instrumentation
can be found in the handbooks edited by Liptak (1997) and Webster (1999).
FLOW RATE INSTRUMENTATION
Flowmeters are used to monitor and control fluid flow. They are divided into
five basic categories based on the method of measurement:
r differential pressure meters
r velocity meters
r positive displacement meters
r mass meters
r variable area meters
Several factors play an important role in determining what type of me-
ter is best suited for a particular operation. Most often, these factors include
31
32 INSTRUMENTATION
Flow
Venturi Tube
Concentric Segmental
Flow
Eccentric
Orifice plate
Flow
Flow nozzle
FIGURE 3.1. Differential pressure flowmeters.
application, fluid properties, desired performance, method and/or ease of instal-

lation, safety, environmental impact, being many times the cost an overriding
factor.
Differential Pressure Meters
Differential pressure meters measure the difference in pressure between the

two sides of a restriction in a confined stream. There are three general types
(Figure 3.1):
r orifice meters
r venturi meters
r flow nozzles
They are based on a restriction to the flow of fluid produced by reducing the
diameter of the pipe.
The equations governing these devices can be derived using the mechanical
energy balance (Bernoulli equation).
v2
P +ρ + gh = constant (3.1)
2
where P is the pressure, v the velocity and ρ the density. This form of the
Bernoulli equation assumes that the density is constant and friction losses are
neglected. Applying this equation to an orifice or a venturi one obtains:
v22 − v12
P1 − P2 = ρ (3.2)
2
where subscripts 1 and 2 refer to upstream and downstream conditions, respec-

tively. Now, the total flow rate is related to the velocity through the continuity
equation,
W = ρ A 1 v1 = ρ A 2 v2 (3.3)
one obtains:
· ¸
1 − (A2 /A1 )
P1 − P2 = ρv22 (3.4)
2
After introducing the diameters of the pipe and the restriction, this equation
can be further rearranged using Equation (3.3) as follows:
√
2π 2 1 p
Wtheo = D2 p (P1 − P2 )ρ (3.5)
4 1 − (D2 /D1 )4
This equation provides a theoretical relationship when density is assumed

constant. It applies well for liquids. When gases are considered, then the me-
chanical energy equation is rewritten as follows:
Z P2
dP
v22 − v12 = −2 (3.6)
P1 ρ
where the change of height is assumed zero and friction losses are again ignored.
This integral can be evaluated by assuming that the gas undergoes an adiabatic
expansion. This is expressed by:
µ ¶k
1
P = constant (3.7)
ρ
Upon integration, and after a few manipulations involving the continuity

equation, one obtains:
√
2π 2 Y1 p
Wtheo = D2 p (P1 − P2 )ρ1 (3.8)
4 1 − (D2 /D1 )4
where Y1 is the adiabatic gas expansion factor. The expression for Y1 is

Ã £ ¤ !1/2
[1 − (D2 /D1 )4 ][k/(k − 1)](P2 /P1 )2/k 1 − (P2 /P1 )(k−1)/k
Y1 = £ ¤
1 − (D2 /D1 )4 (P2 /P1 )2/k [1 − (P2 /P1 )]
(3.9)
34 INSTRUMENTATION
Thus, the equation for gases has been put into the form of Equation (3.5) for
incompressible fluids with the inclusion of a correcting term (Y1 ). This addi-
tional factor is only needed for the compressible case.
Deviations from the theoretical behavior predicted by Equations (3.5) and
(3.8) are rooted in the measurement of the pressure and the corresponding
diameter of the vena contracta. For this reason, a discharge coefficient C is
introduced.
Wtrue
C= (3.10)
Wtheo
The discharge coefficient is calculated at testing conditions, that is, at standard

temperature and pressure. Further corrections need to be performed when actual
conditions are different from the testing conditions. First, Y1 departs from the
value at testing conditions and affects the value of the discharge coefficient.
This variation in C is typically presented as a function of the Reynolds number
(Figure 3.2).
In addition, variations in the temperature of the fluid induce changes in the
discharge coefficient. For compressible fluids, laboratory tests are conducted
by using steam or air as the working fluid. As the fluid changes, the gas ex-
pansion factor also changes. Therefore, several equations exist to correct Y1
(Miller, 1996).
FIGURE 3.2. Variation of the discharge coefficient with the Reynolds number. Adapted
from Upp, 1993.
Testing and Base Conditions
The conditions under which the manufacturer calculates the discharge coef-
ficient are called testing conditions, whereas the typical average conditions at
which the meter measures the flow are called base conditions. The deviations
produced by conditions that are different from the testing or base conditions
(temperature, pressure and fluid composition), which affect the discharge co-
efficient and the fluid density, are a source of errors. The correction of these
deviations between actual and base or testing conditions are called compen-
sation and is performed by measuring temperature and pressure so that the
changes in the discharge coefficient can be inferred. Certain resistance exists
to perform flow compensation to the signal that the operator sees. Rather, it
is preferred to perform this compensation on plant historians and production
accounting databases.
Velocity Meters
Velocity meters measure fluid velocity and use the relationship Q = v A to

calculate volumetric flow rates, where Q is the volumetric flow rate, v the
fluid velocity and A the meter cross-sectional area. There are five basic types
of velocity meters: electromagnetic, vortex, turbine, ultrasonic and pitot. The
signal produced by the meter is linear with fluid velocity; therefore, the errors
associated with square-root extraction that are typical of differential pressure
meters are eliminated.
Electromagnetic meters, or magmeters (Figure 3.3), are governed by
Faraday’s Law of Electromagnetic Induction, V = k0 BLv, where V is the in-
duced voltage, B the magnetic field, L the conductor length through the fluid
(usually the diameter) and v the fluid velocity.
Thus, the volumetric flow rate is obtained by:
π V
Q = k1 L (3.11)
4 B
where k1 is a constant that is determined during testing or calibration.
These meters operate by producing a magnetic field perpendicular to the
fluid flow. The resulting voltage, which is proportional to the fluid velocity,
is measured by two electrodes. Fluids used in magmeters must be electrically
conductive and nonmagnetic. Consequently, they can be used in most water-
based operations.
Turbine meters consist of a bladed turbine rotor that spins as a result of the
action the fluid exerts on it (Figure 3.4). Because the fluid induces an angular
velocity to the rotor, the rotation is proportional to the flow rate and a linear
relationship is obtained.
36 INSTRUMENTATION
FIGURE 3.3. Electromagnetic meter. (a) Reprinted from Miller, 1996, reproduced with
permission of The McGraw-Hill Companies.
In a Vortex meter (Figure 3.5), the flow is split into two streams forcing the
fluid to roll up into a well-defined vortex. After the vortex is formed, it sheds,
and a second vortex begins to form on the opposite side (downstream) of the
element. The formation time of these vortices is proportional to the velocity.
The vortex meter consists of a vortex-generating element spanning across the
diameter of the pipe and a detector to convert the shedding energy into an
FIGURE 3.4. Turbine meter. Reprinted from Miller, 1996, reproduced with permission of
The McGraw-Hill Companies.
FIGURE 3.5. Vortex meter. (a) Reprinted from Dolenc, 1996, with permission from
Rosemount Measurment, Eden Prairie, MN.
electric signal. The pressure fluctuates as a result of the vortex shedding, and
the detector picks up its frequency. Thus, the frequency of shedding of the
vortices from the bluff body is given by (Connel, 1996):
St v
f = (3.12)
ω
where St is the Strouhal number (known to remain constant for Reynolds num-
bers from 10,000 to 1,000,000), f is the frequency of shedding (Hz), v the
stream velocity (m/sec) and ω the width of the bluff body (m).
Both vortex and turbine are called linear flowmeters of the pulse-frequency
type. Because the equations are linear, one can obtain a mean meter coefficient
(K factor), which is the quotient of the pulse frequency and the volumetric flow
rate. Therefore, the governing equation for these flowmeters is:
f
Q= (3.13)
K
where f is the frequency in pulses per second (Hz). The K factor is plotted
versus flow rate over a range. Typical signature curves for a turbine and a vortex
flowmeter are shown in Figure 3.6. The K factor varies; therefore, a mean value
between the maximum and minimum is chosen.
In ultrasonic meters, sound waves are used to determine the velocity of the
fluid. In a Doppler meter, a constant frequency acoustic wave is transmitted
into a fluid. Entrained solids or bubbles reflect back the sound, therefore, the
frequency of the reflected wave changes by an amount proportional to the
38 INSTRUMENTATION
Linearity envelope
K K
Mean K Linearity
(– %)
Linear range
Lower limit. Usually
given by Re=10,000
Flow rate Flow rate
(a) Turbine (b) Vortex
FIGURE 3.6. Turbine meter and vortex meter signature curves. Adapted from Miller, 1996.
velocity. In a transit time meter (Figure 3.7), an acoustic pulse is introduced

diagonally in both upstream and downstream directions.
The difference in time for pulses to travel through the fluid is used to calculate
the velocity. As inferred by Figure 3.8, the time the signal takes to travel forward
and backward is given by the following equations:
d
tA = (3.14)
s − v cos α
d
tB = (3.15)
s + v cos α
where s is the velocity of the sound in the fluid. Therefore, inverting and taking
the difference of both expressions, one gets:
d
v= ( f B − f A) (3.16)
2 cos α
where f A and f B are the frequencies at each receiver. In other words, the flow
rate is proportional to the difference in frequencies.
c Panametrics,
FIGURE 3.7. Transit time ultrasonic meter. Reprinted from Dolenc, 1996. °
Inc., 1996.
A B
a
d
FIGURE 3.8. Schematics of ultrasonic meter.
The Pitot tube (Figure 3.9) consists of a small tube inserted horizontally in
the pipe. This tube creates a stagnation point in the fluid flow. At this point, the
pressure exerted by the fluid is larger than the pressure at static holes located
in the wall.
Applying the Bernoulli equation between the wall and the static holes one
obtains
vs2
Ps − P0 = ρ (3.17)
2
where Ps and vs is the pressure and velocity at the stagnation point, whereas
P0 is the static pressure. From this equation, one can derive an expression for
the velocity:
s
2(Ps − P0 )
vs = (3.18)
ρ
In the case of gases (typically above 200 ft/sec), compressibility becomes

important and the fluid is assumed to be isentropically compressed at the point
Flow Flow
Low
DP Openings for static pressure
pressure side
High pressure
side
(a) Side-Wall Static Tap (b) Static Tube

FIGURE 3.9. Pitot tube schematics.
40 INSTRUMENTATION
of impact. Thus, the following equation is derived:

v " #
u
u 2k µ P0 ¶ µ P0 ¶(k−1)/k
vs = C t −1 (3.19)
k − 1 ρ0 ρ0
The coefficient C is close to 1. We now express the mass flow rate as a function
of the pressure drop and density for the incompressible case as follows:
s
√ (Ps − P0 )
Q = vs A = 2K pitot 4Pitot (3.20)
ρ0
where K pitot is the pitot coefficient, which takes care of the departures from
testing conditions, and 4Pitot is the compressibility factor. Because of the form
of this equation and the fact that pressure difference is measured, pitot meters
are sometimes also classified as differential pressure meters.
Positive Displacement or Volumetric Meters
Positive displacement (PD) meters, also called linear meters, directly measure
volumetric flow rates by letting the flow pass through compartments of known
volume. The fluid passage is then counted and multiplied by the individual
volume of each segment. Measurements are taken for a given amount of time,
and the output is transmitted on a volume/time basis using a counter.
There are several types of PD meters, including the rotary sliding vane, the
lobed impeller and the gear meter (Figure 3.10).
In the rotary sliding vane meter, spring-loaded vanes move in and out of the
moving rotor. A known amount of fluid is trapped between the vanes and the
outer wall, which allows calculation of the volume/revolution ratio. In lobed
impeller and gear meters, fluid is trapped between the impellers or gears and
the wall. Thus, the same linear relationship between frequency and flow is es-
tablished, as in Equation (3.13). These meters are used for product transactions
at the battery limits (usually called cash transactions) and are subject of several
ANSI standards, which makes them expensive.
A typical signature curve for these types of flowmeters is shown in Figure 3.11.
Some detailed features about positive displacement meters are discussed by
May (1971).
Mass Meters
Mass flowmeters are classified into four types: Coriolis, Thermal, Heated
Element and Temperature Rise. The Coriolis meter (Figure 3.12) is based on
Newton’s Second Law of Motion (F = ma). In this meter, fluid flows through
FIGURE 3.10. Positive displacement meters. Reprinted from Dolenc, 1996, with permis-
sion from Brooks Instrument, Hatfield, PA U.S.A.
two flow tubes causing them to twist. The amount of twist is proportional to the
mass flow rate, which is measured by magnetic sensors attached to the tubes.
As the fluid rotates due to direction changes with angular velocity Ä, a force
F is exerted on the fluid. Thus, the governing equation of these meters is
F
Q= (3.21)
2LÄ
where L is the length of the pipe. In practice, the pipe is oscillated by electro-
magnetic forces at the natural frequency of the structure. Magnetic sensors on
True volume
K
1.01
1.00
0.99
Slip
Theoretical volume
Operating range
Flow rate
FIGURE 3.11. Positive displacement meter: signature curve. Adapted from Miller, 1996.
42 INSTRUMENTATION
c Rosemount Inc.,
FIGURE 3.12. Coriolis flowmeter. Reprinted from Pomroy, 1996, °
2000.
the tubes measure the displacement. The twist angle relates to the mass flow
rate, whereas the frequency of vibration indicates the density of the fluid.
The thermal mass-flow meter operates by measuring thermal properties of the
fluid (Figure 3.13). One type of thermal mass-flow meter, the thermal anemome-
ter, measures the amount of heat removed from an electrically heated probe. As
the fluid flows through the probe, its temperature changes and, consequently, a
variation of resistance is observed. To maintain the temperature constant, the
power is varied, and it is this power variation that is used to infer the flow rate.
Temperature Sensor Velocity Sensor
Wires
FIGURE 3.13. Immersible thermal flowmeter.

• •• •
Sensor Tube
Q2
Q
Q1
Laminar By-pass
FIGURE 3.14. Bypass thermal flowmeter.
A second sensor is often introduced to measure the temperature and adjust the
power on the first sensor.
The mass flow rate is given by the following relationship:
µ· ¶ ¸
1 E v2 /Rv k2 1/m
Q= − (3.22)
k1 Tv − T k1
where E v is the electrical power, Rv is the resistance, (Tv − T ) is the temper-

ature difference, which is maintained constant, and k1 , k2 and m are empirical
constants.
The temperature-rise flowmeter measures the temperature of the fluid up-
stream and downstream of a heater device and calculates the flow rate from the
change in temperature (Figure 3.14). Typically, a bypass of the flow is arranged
so that these temperature measurement elements can be properly installed. A
laminar flow element induces a linear pressure drop, ensuring that the flow
rates through the two fluid paths are proportional. Thus, the ratio of the flows
is known.
These flowmeters are governed by the following expression:
µ ¶
Q2 q
Q = 1+ (3.23)
Q 1 C p 1T
where Q 2 and Q 1 are the mass flow rates through the two paths, q is the heat
added to the fluid, C p is the fluid heat capacity and 1T is the temperature
difference.
A recent review of mass flowmeters is offered by Gibson (1998). He points
out the emergent popularity of Coriolis and Thermal mass meters prompted by
the increased precision requirements in the process industry.
44 INSTRUMENTATION
Pressure
Force
AF Weight
v2 v2
Aa Buoyancy
Force
Float
v1
Pressure
Force
(a) Schematics (b) Force Balance Schematics

FIGURE 3.15. Rotameter.
Variable Area Meters
Area flowmeters use a float inserted in a vertical tube, exposing a variable

area to the flow (Figure 3.15). The float stays at a certain height by virtue of a
balance between buoyancy, weight and pressure forces.
Assuming that the downward pressure is just the static pressure, whereas the
upward pressure force is the total pressure (static + dynamic), one can write a
force balance to obtain the following relation for the volumetric flow rate:
s µ ¶
VF ρE
Q = K R Aa 2g −1 (3.24)
AF ρ
where K R is the correction factor, which is one under ideal conditions, Aa is the
free area left by the float, A F the area of the float, VF the volume of the float and
ρ F the density of the float. For compressible fluids, there is a compressibility
factor that must be included.
Applicability and Selection Criteria
Differential pressure meters are the most widely used flowmeters. To main-
tain accuracy, they require 5-40 straight pipe diameters upstream of the meter.
Orifices are the most widely used differential pressure meters, and their ad-
vantages include durability, low cost for large pipe diameters and the ability
to be used at a variety of temperatures. However, these flowmeters have high
sensitivity to changes in density and viscosity. In addition, they have a tendency
to erode. Venturi meters are used in high flow rate applications and are more
expensive. However, in such applications, they may result in significant perma-

nent pressure losses and diminished accuracy. Flow nozzles are used when ca-
pacity is the important issue. They work well in high-velocity, high-temperature,
and high-turbulence situations. They are also capable of handling larger solids
than orifice meters. Although flow nozzles are less expensive than venturis, they
result in higher permanent pressure losses and are less accurate.
Pitot tubes are used in streams that carry clean liquids or gases. These tubes
are sensitive to changes in the velocity profile of the fluid; therefore, several
ports are placed across the diameter of the pipe and an average value for the
differential pressure is obtained. The small openings of these meters tend to
become plugged if particles are present in the fluid. In addition, they produce a
low differential pressure, which makes repeating measurements a problem.
Positive displacement (PD) meters are most often used in hydrocarbon batch-
ing, blending and custody transfer operations. They perform well with viscous
fluids but are not recommended for gases because of a lack of sealing or for
dirty fluids due to the presence of gears and vanes. Advantages of PD meters
include their accuracy when using viscous fluids and their high flow range.
Disadvantages include the lack of sealing around the gears, the loss of pressure
due to the removal of energy by the meter, and the coating of gears by extremely
viscous fluids leading to a volume reading higher than the actual flow.
Coriolis meters are the most accurate instruments (Pomroy, 1996) and until
1999 are mainly used in critical control loops and in the management of high
value fluids. Benefits of these meters include their high degree of accuracy, large
flow range and their tight control ability. In addition, they do not need recalibra-
tion, do not obstruct the fluid flow and can be used with different types of fluids.
Thermal mass-flow meters are effective with clean, low-density gases. They
are used to measure gas flow in various applications such as ducts, pilot plants,
purge streams and leak testing and are most commonly used in dopant gas
flows for semiconductor production. Benefits include their ability to accurately
measure gas flows. Disadvantages include the coating of probes by dirty fluids
and the need for fluid uniformity.
Variable area meters, such as rotameters and purge meters, can be used in both
gas and liquid applications. Because their operation is governed by equilibrium
of forces, their benefits include a small pressure loss, low cost and ease of
installation. One disadvantage is the wide fluctuation in accuracy.
Because there are no mechanical parts that impede fluid flow, magmeters are
excellent for streams containing particles or corrosive chemicals. Other advan-
tages include their ability to provide accurate voltage signals in both laminar
and turbulent flow and their low sensitivity to changes in density, pressure and
viscosity. A disadvantage of these meters is the potential of the electrodes to
become coated, thereby producing errors in the voltage reading. However, this
problem can be virtually eliminated by maintaining a fluid velocity between 8
and 15 ft/sec.
46 INSTRUMENTATION
Table 3.1 summarizes some of the features of flowmeters (Dolenc, 1996;

Furness, 1998; McMillan et al., 1998; Pomroy, 1996; Smith, 1978; Upp, 1993;
Zientara, 1972).
Accurate measurement of mass flows requires the measurement of density or
the pressure and temperature from which the density or density changes can be
inferred. The most direct solution is to install a density transmitter. However,
this is expensive. For liquids, a density/temperature correlation can be made,
so that if the concentration is fairly constant, corrections can be made on the
basis of temperature variations. In the case of gases, knowledge of the absolute
pressure is also needed for such calculations. Another criteria related to the
range of measurement, linearity and accuracy will be discussed in Chapter 4.
LEVEL MEASUREMENT
Level measurement is essential, even when one assumes steady state. It pro-
vides one more element of flow rate data redundancy, because the difference in
levels can be representative of a steady flow leaving or feeding a tank. They are
divided into five basic categories based on the method used:
r differential pressure meters
r displacer-based meters
r signal-reflection-type meters
r radio-frequency admittance meters
r radiation-based meters
Level instruments based on floats and switches are omitted. The former are
being abandoned, and the latter do not offer continuous monitoring. Neverthe-
less, meters based on switches are still used for safety purposes.
Differential Pressure Meters
This is the most frequently used device. It consists of measuring the difference
in pressure between two points in a vessel (Figure 3.16).
The difference in pressure is given by:
1P = ρl gh + ρv g(H − h) (3.25)
where h is the difference of level between the liquid-vapor interface and the
lower arm of the meter and H is the total height difference between the arms.
Inverting the equation, one gets:
1P − ρv g H 1P
h= ≈ (3.26)
(ρl − ρv )g ρl g
TABLE 3.1. Applicability of the Main Types of Flowmeters.
Clean/Dirty Clean/Dirty Corrosive Slurries Max Press. Temp. Range

Technology Liquid Gas Liquid Part./Fibrous Steam (psig) (◦ F)
Differential pressure
Nozzle (+ + +)/(+) (+ + +)/(++) (++) (−)/(−) (+ + +) 4000 ≤1000
Orifice (+ + +)/(+) (+ + +)/(++) (+) (−)/(−) (+ + +) 4000 ≤1000
Venturi (+ + +)/(++) (+ + +)/(++) (++) (−)/(−) (++) 4000 ≤1000
Velocity
Electromagnetic (+ + +)/(+ + +) (−)/(−) (+ + +) (+ + +)/(+ + +) (−) 5000 −40–350
Turbine (+ + +)/(−) (++)/(−) (−) (−)/(−) (−) 6000 −450–6000
Vortex (+ + +)/(++) (+ + +)/(++) (+) (+)/(−) (+ + +) 1500 −330–800
Ultrasonic Doppler (−)/(+ + +) (−)/(++) (+) (++)/(+) (−) 1500 −50–900
Ultrasonic transit time (+ + +)/(−) (+ + +)/(−) (+) (−)/(−) (+) 1500 −50–900
Volumetric
Positive displacement (+ + +)/(+) (−)/(−) (++) (+)/(−) (−) 300–1500 −30–600
Mass
Coriolis (+ + +)/(+ + +) (+)/(+) (+ + +) (+)/(+) (+) 1400–5700 −400–800
Thermal mass (−)/(−) (+ + +)/(+ + +) (−) (−)/(−) (−) 4500 32–150
Temp. rise (+ + +)/(+ + +) (+)/(+) (++) (++)/(++) (+) 4500 32–150
Variable area (+ + +)/(−) (+ + +)/(−) (++) (−)/(−) (+) 200–6000 ≤1000
Note: (+++) = acceptable; (++) = normally acceptable; (+) sometimes acceptable; (- ) not acceptable.
47
48 INSTRUMENTATION
Differential
Pressure
H Transmitter
FIGURE 3.16. Differential pressure level measurement.
In this formula, one can neglect terms containing the vapor density and do
not incur in any gross error.
Displacement-Based Meters
This type of level meter operates on the basis of buoyancy. The displacer
(Figure 3.17) is immersed in a side arm chamber, sometimes inside the vessel,
and is restrained by an elastic device whose motion or distortion is directly
proportional to the buoyant force. The reading is taken by either measuring
directly the vertical position or indirectly through torque devices.
Signal Reflection-Type Meters
These meters are based on the concept of sending a signal and measure its
reflection from the surface (Figure 3.18). Two types of signals can be used:
ultrasonic and electromagnetic. Based on the speed of the signal, the time of
Displacer
Transmitter
Displacer
Chamber
FIGURE 3.17. Displacer.

Level Measurement 49
Ultrasonic
Transmitter
Ultrasonic
Transducer
FIGURE 3.18. Ultrasonic level meters.
response is directly proportional to the distance from the source/receiver to the

liquid-gas interface.
Radio Frequency Admittance Meters
In radio frequency (RF) meters the change of impedance between two ca-
pacitor electrodes is used to measure the level (Figure 3.19). The first electrode
is typically the wall of the vessel, whereas the second is a rod immersed in the
fluid.
Once a voltage is imposed between the two electrodes, one can measure
the current and the capacitance. The change of level in the vessel results in a
change of impedance. In RF devices, capacitance increases as the level in the
vessel increases, because the air, which is a low dielectric, is replaced by a
RF
Transmitter
RF sensing
element
FIGURE 3.19. Radio frequency level meters.

50 INSTRUMENTATION
Radioactive
Source
Radioactive
Detector
FIGURE 3.20. Nuclear radiation-based level meters.
higher dielectric liquid. A constant voltage is applied to a rod or cable (sensing

element) in the process. The radio frequency current that results is monitored
to infer the level of the process material. Admittance systems are capacitance
devices with the ability to ignore coatings on the sensor.
Radiation-Based Meters
These meters consist of a source of gamma rays that is located on one side of
the vessel and a detector installed on the opposite side (Figure 3.20). Because the
liquid has lower transmitance than the vapor, the detector is able to determine
its level.
Strapping Tables
Many vessels are not cylindrical, and they have a varying cross-sectional area.
Therefore, although the level can be measured by several of the above-described
devices, the actual volume of material inside the vessel is not a linear function
of it. Instead of making use of such a nonlinear analytical relation, industry
actually makes use of a table. One of course can, and in many cases it has been
done, create a function by simply fitting the table values. Strapping tables are
a vivid example of old practices to which industry holds-on to unnecessarily.
A primary benefit of differential pressure level meters is that they can be

readily installed on a vessel and can be easily removed using block valves.
However, these meters are subject to errors because of density variations of
the liquid, which are typically caused by product composition or product tem-
perature changes. If accurate measurements are to be made, these variations
must be compensated. Therefore, their main application is for clean liquids
(Connell, 1996).
Displacement type level meters work well with clean liquids, and they are
accurate and applicable over a wide range of fluid densities. However, they
are also affected by changes in product density. Because the displacer is in the
process fluid, solids can deposit on it and cause a calibration shift. Displacers
should only be used for relatively nonviscous, clean fluids and perform better
when used for short spans. One disadvantage shared by these devices and the
differential pressure level devices is that two penetrations (nozzles) into the
vessel are required (Sholette, 1996).
Ultrasonic level meters do not have moving parts, and they do not come
in contact with the process material. In addition, calibration is relatively easy.
However, dust, heavy vapors, surface turbulence, foam and even ambient noise
can affect their performance. Moreover, because they are based on the velocity
of sound in the vapor phase, temperature and pressure can also affect their
performance.
Radio frequency (RF) technology produces highly accurate measurements
in storage tanks and some process vessels. Its primary disadvantage is cost,
which can be justified for tank gauging and inventory control. In addition, RF
admittance is fairly application specific. Indeed, temperature, pressure, chem-
ical compatibility and electrical properties affect measurement accuracy and
need to be carefully considered. Capacitance measurements are also subject
to large errors caused by conductive coatings. Admittance technology avoids
this problem by correcting the output for coatings. A second sensor is added to
monitor the dielectric constant and then compensate the calibration based on
this information.
Because nuclear systems require a radioactive source mounted on one side
of the vessel and a radiation detector or receiver on the other side, no tank
penetration is needed to make a successful level measurement. Typical process
concerns, such as very high temperatures or pressures, or corrosiveness of the
liquid to be measured, are irrelevant. High cost, estimated to be two to four
times the cost of other technologies, prevents this technology to be widespread.
In addition, licenses, approvals and periodic inspections are also required. Ac-
curacy, linearity and speed of response are also of concern. Finally, because
they rely on signal attenuation, these devices require the density to be relatively
constant (McMillan et al., 1998).
Table 3.2 summarizes the application suitability of level instrumentation.
TEMPERATURE MEASUREMENT
Temperature-measuring devices are divided into three groups:

r thermal-expansion thermometers
r electrical devices
r radiation-based devices
52 INSTRUMENTATION
TABLE 3.2. Applicability of the Main Types of Pressure Measuring Devices.
Changing
Technology T > 300 ◦ F P > 600 psi Liquids Vapors Density
Differential pressure Y Y Y Y N
Displacer Y Y Y Y N
RF admittance/capacitance Y Y Y Y Y
Ultrasonic N N Y N Y
Electromagnetic N N Y Y Y
Source: Parker (1998), Carsella (1998).
Thermal Expansion Thermometers
These temperature-measuring devices are based on the thermal expansion

of either liquids or solids. In the case of liquids, the sensitive element (a bulb)
contains a liquid or gas, which changes its volume or pressure with tempera-
ture. Those systems based on volume changes are completely filled with liq-
uid, whereas those based on pressure changes are filled with gas. Liquid in
glass thermometers are not used in automatic control systems, because read-
ings are typically manual. Therefore, the so-called filled-system thermometers
have been used more extensively (Figure 3.21). Changes in gas pressure are
carried through a capillary, sometimes through large distances (up to 250 ft) to
a bourdon tube. The motion of the bourdon is used directly or amplified by a
mechanical linkage to drive a pen in a recorder.
FIGURE 3.21. Filled system. Reprinted from Perry, 1984, reproduced with permission of
The McGraw-Hill Companies.
FIGURE 3.22. Bimetallic thermometers. Reprinted from Perry, 1984, reproduced with
Another type of thermal expansion thermometers are bimetal thermometers

(Figure 3.22), which are made of strips of two metals that have different ex-
pansion rates. Because these strips are tied in one end, the other end deflects.
Typically, they are wound in a spiral or helix arrangement in which the loose
end rotates with temperature changes.
Electrical Devices
Thermocouples are made of two different metallic wires that are connected.
When the two junctions are at different temperatures, an electric current flows
(Figure 3.23).
FIGURE 3.23. Thermocouple.

54 INSTRUMENTATION
However, when the circuit is open, no current flows and an electromotive

force (E) is measured. The thermal electromotive force E is a measure of the
difference in temperature through the following relation.
E ≈ a(T1 − T2 ) + b(T1 − T2 )2 (3.27)
where a and b are constants.

In control systems, the reference junction (cold junction) is usually located
at the emf measuring device. This junction may be held at constant temperature
such as in an ice bath or a thermostat oven, or it may be at ambient temperature
but electrically compensated so that it appears to be held at a constant tem-
perature. A thermocouple is inherently an inaccurate device (Wightman, 1972)
because it is not normally possible to maintain a good quality control on the
wire used.
Resistance thermometers are based on the change in electrical resistance
when a metal undergoes a change in temperature. Industrial resistance ther-
mometers are usually constructed of platinum, copper or nickel, but semicon-
ducting materials such as thermistors are in increased use. Different forms of
bridge circuits are used in industrial resistance thermometry, the most common
being the Wheatstone bridge. The change in electrical resistance of a material
with a change in temperature, the coefficient of resistance, is positive and es-
sentially constant over large portions of their range. Typical resistance versus
temperature curves for platinum, copper and nickel are given in Figure 3.24.
Thermistors operate on the same principle, but there is no metal involved.
They are made of semiconductors such as mixtures of oxides of nickel,
Relative
Resistance
(Rt/Ro) 6
Nickel
Copper
4
Platinum
-200 0 200 400 600 800 o

C
FIGURE 3.24. Typical resistance curves. Adapted from Perry, 1984.

manganese, copper, iron, magnesium, titanium and other metals sintered at

high temperatures. Finally, integrated circuit sensors are electronic devices
that provide a linear output of voltage as a function of temperature in a certain
range.
Radiation-Based Devices
Pyrometers measure the temperature of an object by determining the quantity

and characteristics of the energy that it radiates. There are two groups: (1) op-
tical pyrometers in which the brightness of a hot object is compared with that
of a source of standard brightness and (2) radiation pyrometers, which either
measure the rate of energy emission per unit area over a relatively broad range
of wave lengths or compare the radiation at two wave lengths.
Many thermal expansion thermometers are relatively inexpensive, but they

are not very amenable to the modern computer-based data acquisition systems.
Although possible, conversion from mechanically based indicators to electronic
signals require an investment that makes thermocouples and thermoresistance
thermometers more competitive for this purpose. Among the latter, the use of
thermocouples is less expensive and more popular. Table 3.3 illustrates some
of the applicability criteria and ranges of use.
TABLE 3.3. Applicability of the Main Types of Temperature Measuring Devices.
Temp. Range
Technology (◦ F) Advantages Disadvantages
Thermal expansion
Filled thermal −300–1000 Simple Factory repair only
Bimetallic −80–800 Low cost Changing calibration
Simple with handling
Electrical
Thermocouple −440–5000 Small size Cold junction can affect
Wide range calibration
Thermoresistance −430–1800 Small size Self-heating
Accurate Drift
Expensive
Thermistor −150–600 Narrow span Nonlinear response
No cold junction
Optical
Pyrometer 0–7000 No contact Fragile
Wide range Nonlinear scale
Source: Johnson (1998), Zientara (1972).
56 INSTRUMENTATION
PRESSURE MEASUREMENT
Process pressure measuring devices are divided into three groups:

r liquid column devices
r elastic element devices
r electrical sensing devices
Liquid Column Devices
In liquid column devices, pressure is measured by determining the height of

a liquid column. If the density of the liquid is known, this height is a measure
of the pressure. Most forms of liquid column pressure measuring devices are
commonly called manometers. When the liquid columns are contained in glass
or other transparent tubes, the liquid height can be observed visually. Depending
on the pressure range, water and mercury are the most frequently used. Because
the density of the liquid used varies with temperature, this fact must be taken
into account for accurate pressure measurements.
Elastic Element Devices
Elastic element pressure measuring devices rely on the deformation of an

elastic material (usually metallic). The magnitude of the deformation is pro-
portional to the applied pressure. There are three types: bourdon tube, bellows,
and diaphragm.
r Bourdon-tube elements: Probably the most frequently used process pressure-
indicating device is the C-spring bourdon-tube pressure gauge (Figure 3.25).
As pressure increases the bourdon tube end moves resulting in different po-
sitions of the pointer.
r Bellows elements: These are axially elastic cylinders with folds. The bellows
can be used unopposed, or it they can be restrained by an opposing spring
(Figure 3.26). The pressure to be measured may be applied either to the
inside or to the space outside the bellows, whereas the other side is exposed
to atmospheric pressure.
r Diaphragm elements: There are two types of diaphragm elements: Unop-
posed and opposed by a spring. The first type (Figure 3.27a) consists of one
or more capsules each composed of two flat or corrugated diaphragms bonded
together by soldering, brazing or welding. The second type of diaphragm is
used for containing the pressure and exerting a force on the opposing elastic
element. The diaphragm is a flexible or slack membrane made of rubber,
leather, impregnated fabric or plastic. The movement of the diaphragm is
opposed by a spring, which determines the deflection for a given pressure
Pressure Measurement 57
FIGURE 3.25. C-type bourdon manometer. Reprinted from Perry, 1984, reproduced with
(Figure 3.27b). This type of diaphragm is used for measurement of extremely

low pressure, vacuum or differential pressure.
Electrical Sensing Devices
These devices take advantage of the property that electrical resistance of

conducting solids changes with diameter and length. As pressure is exerted on
wires, these dimensional changes result in an increase in the electrical resistance
of the conductor. These devices are called strain gauges. Bonded strain gauges
are those that are bonded directly to the surface of the elastic element whose
strain is measured (Figure 3.28).
FIGURE 3.26. Bellow elements. (a) Unopposed bellow. (b) Spring-loaded bellow. Re-
printed from Perry, 1984, reproduced with permission of The McGraw-Hill Companies.
58 INSTRUMENTATION
FIGURE 3.27. Diaphragm pressure devices. (a) Unopposed. (b) Opposed. Reprinted from
Perry, 1984, reproduced with permission of The McGraw-Hill Companies.
Although elastic elements are used in practice to measure pressure locally,

direct pressure readings in electric form are only possible if there is a way
to transform the pressure reading to a voltage signal. Strain gauge pressure
transducers provide a number of advantages. They cover the range from 0.15 to
660 bar and temperatures from −160◦ C to 350◦ C and are capable of providing
0–15 MV signals by proper use of electrical bridges. Piezoelectric devices
operate more or less in a similar fashion. Cost is of course an issue irrespective of
the transducer selected. Problems exist because overpressure, vibration, thermal
cycling etc., which affect their performance. A good guideline for selection is
offered by Omega (1998).
DENSITY MEASUREMENT
Liquid density is typically measured in the laboratory by means of floating

type densitometers and weighing techniques. More sophisticated measurements
can be achieved on line by the use of radiation based gravitometers or continuous
weighing U-tubes based on force balance restoration.
FIGURE 3.28. Bonded strain gauge. Reprinted from Perry, 1984, reproduced with per-
mission of The McGraw-Hill Companies.
Density Measurement 59
FIGURE 3.29. Continuous weight densitometers. (a) Direct weight. (b) Vibrating tube.
c CRC Press, Boca Raton,
Reprinted with permission from Wightman, 1972. Copyright °
Florida.
Radiation-based densitometers are based on the same principle as the radia-

tion-based level measuring devices. A gamma ray is passed through the pipe,
which is filled with the substance whose density is measured. The absorption
rate is measured and related to the density.
Continuous weighing U-tubes exist in two types: direct weight measurement
and vibrating frequency measurement; both types are described in Figure 3.29.
The direct weight measurement is performed by allowing the tube to pivot
horizontally on across-leaf suspension as in the weights of a conventional beam
balance. Any deviation in weight is detected optically and amplified.
In the vibrating tube densitometer a drive coil and a pick up coil are suspended
between the two tubes, which in turn are mounted on a pair of antivibrating
mounts. Once the drive coil has exerted a force, the frequency of vibration is
measured. This frequency is given by the following relation:
r
1 S
f ≈ (3.28)
2 I
where S is the stiffness and I is the inertia. The stiffness is a constant and is
related to Young’s modulus of the sampling tube. The inertia is a function of
the volume and the density of the fluid.
Continuous weighing densitometers and radiation-based densitometers are

expensive. Because density measurements are only critical to determine flow
60 INSTRUMENTATION
TABLE 3.4. Applicability of the Main Types of On-Line Analyzers.
Technology Sample Range
Gas chromatography
Thermal conductivity detector Gas, vapor or 0.1–100%
Ionization detector volatile liquid ppm to 100%
Infrared analyzer Usually gas or 0.1–100%
vapor
Thermal conductivity Usually used 0.1–100%
for hydrogen
Ultraviolet absorption Liquids 0.01–100%
(especially
aromatics)
Colorimeters Usually liquids 0.1–100%
Source: Foster (1975).
rate, unless density needs to be monitored for quality control purposes, industry
relies on temperature and pressure measurements to compensate flow rate-
measuring devices.
ON-LINE PROCESS ANALYZERS
Typical on-line analyzers consist of pH meters, spectrophotometers, oxy-

gen analyzers, mass spectrometers, moisture analyzers, liquid and gas chro-
matographs. The application of these analyzers to processes is discussed by
Utterback (1976). Usually, these instruments can be calibrated to measure the
concentration of one component directly. The advantages of using process an-
alyzers are that information about off-spec products is available sooner, envi-
ronmental monitoring is done in situ, and, of course, there is a reduction in
laboratory costs (Foster, 1975). However, they present maintenance challenges
that prevent them from becoming more popular. Table 3.4 summarizes their
applications.
TRANSMISSION AND TRANSFORMATION OF SIGNALS
Pneumatic transmission used to be the main method of conveying measure-

ment signals to control rooms. In the last decades, industry has undergone a
transformation to electronic transmittal equipment. Devices that perform this
operation are called transducers. For example, differential pressure transducers
transform the pressure difference into an output voltage. Ideal transducers are
linear, but in practice they are affected by zero offsets, hysteresis etc.
Transmission and Transformation of Signals 61
Differential
Temperature Pressure Pressure
signal signal signal
P
T DP
‚
D P* (P/T)
· K
P/T
Square True mass
flow
Root
FIGURE 3.30. Computation of mass flow for gases.
In addition, many signals have to be electronically manipulated. For example,

a square root of the signal from differential pressure flowmeters and pitot tubes
must be taken to obtain the volumetric flow rate. Thus, one should expect the
accumulation of errors in all these transformations. This process is illustrated
in Figure 3.30. The assumption is that the density is given by a gas law of
choice (ρ = PM/ZRT) and therefore roughly proportional to the quotient P/T .
In the case of liquids, there is a direct relationship between density and inverse
temperature, almost linear in the range desired. Thus, the same type of scheme
as in Figure 3.30 can be used to obtain the final signal. A detailed formula
to obtain mass flow rate containing several correction factors was developed
several years ago by L. Spink from Foxboro (Connel, 1996).
Smart Sensors and Digital Protocols
In the last decade, digital communications protocols have been introduced.

These protocols allow transmitters and other equipment to communicate with
control systems, valves etc. Communication in these systems is two ways, from
the sensor to the control room and vice versa. The new name of smart sen-
sors has been coined for those instruments that are associated with these new
technologies. Examples of these protocols are HART (Highway Addressable
Remote Transducer) and Fieldbus. HART is an open (nonproprietary) stan-
dard protocol. Fieldbus is an all-digital, serial, two-way communications sys-
tem, serving as a Local Area Network (LAN) for instruments used in process
control and manufacturing automation applications. The demand for fieldbus
technology is being driven by the desire of manufacturing end users to move
away from older, centralized plant control strategies to distributed control in the
field. Fieldbus is neither owned by any individual company nor regulated by a
62 INSTRUMENTATION
single nation or standards body. The technology is controlled by the Fieldbus

Foundation, a not-for-profit organization consisting of more than 100 of the
world’s leading controls and instrumentation suppliers and end users. In addi-
tion, with smart sensors, the number of trips to the field, which only 36% result
in corrective action (Masterson, 1999), can be drastically reduced, calibration
can be performed from the shop. Finally, these sensors prompted changes in
maintenance practice, as “just in time” maintenance (predictive maintenance)
can be performed. Masterson (1999) provided several examples of compa-
nies that have saved anywhere from 50,000 to 250,000 U.S. dollars in annual
savings.
The use of these technologies is eliminating the need for digital-to-analog
conversion of signals, and vice versa, performed nowadays. Also an increase
in the reliability of data obtained from flowmeters based on digital outputs of
pulse counts and frequency is observed (McMillan et al., 1998).
NOMENCLATURE
A: cross-sectional area
B: magnetic field
C: discharge coefficient
C p: heat capacity at constant pressure
Cv : heat capacity at constant volume
d: distance
D: diameter
Ev : electrical power
E: electromotive force
f: frequency of shedding in vortex meters (Hz)
F: force
H: height
h: height difference of level
I: inertia
k: ratio of gas heat capacities (C p/Cv )
ki : constants
K: mean meter coefficient
K R: correction factor
L: length
P: pressure
q: heat
Q: volumetric flow rate
Rv : resistance
S: stiffness
s: velocity of sound in fluid
References 63
St : strouhal number
t: time
T: temperature
v: velocity
V: induced voltage
W: mass flow rate
Y1 : adiabatic gas expansion factor
Greek Letters
ρ: density
4Pitot : pitot compressibility factor
Ä: angular velocity
ω: width of bluff body
REFERENCES
Carsella B. Popular Level-Gauging Methods. Chemical Processing, Dec. (1998).

Connel B. Process Instrumentation Applications Manual. McGraw Hill (1996).
Dolenc J. W. Choose the Right Flow Meter. Chemical Engineering Progress, pp. 22, Jan. (1996).
Foster R. A. Guidelines for Selecting Online Process Analyzers. Chemical Engineering, March
(1975).
Furness R. A. Flowmeter Classifications and Applications. Chemical Processing. Fluid Flow. An-
nual, pp. 72 (1998).
Gibson W. D. Taking the Measure of Mass Flow. Chemical Engineering, Sept. (1998).
Johnson F. L. Temperature Measurement and Control Fundamentals. Chemical Processing, June
(1998).
Liptak B., editor. Instrumentation Handbook. Third Edition. Chilton (1997).
May D. L. Accurate Flow Measurements with Turbine Meters. Chemical Engineering, March
(1971).
Masterson J. S. Reduce Maintenance Costs with Smart Field Devices. Hydrocarbon Processing,
January (1999).
McMillan G. K., G. E. Mertz and V. L. Trevathan. Trouble Free Instrumentation. Chemical Engi-
neering, Nov. (1998).
Miller R. W. Flow Measurement Engineering Handbook. McGraw Hill (1996)
Omega. Force Related Measurements. Transactions in Measurement and Control. Vol. 3, Putnam
Publishing and Omega Press LLC (1998).
Parker S. Selecting a Level Device Based on Application Needs. Chemical Processing. Fluid Flow
Annual. pp. 75 (1998).
Perry R. H. and D. Green. Perry’s Chemical Engineers’ Handbook, 6th Edition. McGraw-Hill, Inc.
New York (1984).
Pomroy J. Selecting the Right Flowmeter. Chemical Engineering, May (1996).
64 INSTRUMENTATION
Sholette W. Pick the Proper Level Measurement Technology. Chemical Engineering Progress,
October (1996).
Smith C. L. Liquid Measurement Technology. Chemical Engineering, April (1978).
Upp E. L. Fluid Flow Measurement. Gulf Publishing Company (1993).
Utterback V. C. Online Process Analyzers. Chemical Engineering, June (1976).
Webster J. G., editor. Measurement, Instrumentation and Sensors Handbook. CRC Press, LLC
(1999).
Wightman E. J. Instrumentation in Process Control. CRC Press, Butterworth & Co, London (1972).
Zientara D. E. Measuring Process Variables. Chemical Engineering, Sept. (1972).
CHAPTER 4
Errors in Measurement
INTRODUCTION
This chapter is devoted to analyzing the basic aspects of errors in instru-

mentation measurements. A few general definitions related to instrumentation
properties are first offered. In a second part, elements contributing to mea-
surement quality, such as precision, systematic errors, hysteresis, dead band,
sensitivity and speed of response, are covered. Following are the terminology
and concepts used in calibration curves. Finally, in the last part, a brief review
of the accuracy of different instruments is offered. A more complete coverage
of the material discussed in this chapter can be found in the handbooks on
instrumentation edited by Liptak (1997) and Webster (1999).
INSTRUMENT PROPERTIES
Range, Span and Rangeability
The interval within which a certain variable is measured or transmitted is

called range. It is expressed by stating the lower and upper range values. The
span is simply the difference between the upper and lower range values. Finally,
the rangeability is defined as the ratio of the highest to the lowest value that an
instrument can measure with the same accuracy.
Influence Variables
Many variables influence the performance of an instrument. Ambient temp-

erature and humidity for example affect the readings and introduce bias and/or
65
66 ERRORS IN MEASUREMENT
alter the precision. These are called influence quantities (Miller, 1996), and
their effects are not linearly additive.
Readability
This is defined as the smallest scale increment to which the reading may be
determined expressed as a percent of the full scale (Perry, 1984). This property
of sensors is mostly related to visual reading, but it also appears in digital or
analog reading devices.
MEASUREMENT QUALITY
Measurements are subject to errors, no matter how much the conditions and
the apparatus used to make them are improved. As it was briefly discussed in
Chapter 1, precision and accuracy are terms that are often confused. In this
section, elements contributing to their definition are discussed.
Precision
Precision of an instrument is defined as the closeness of agreement among

a number of consecutive measurements of a variable that maintains its value
static. It bears no relation to the true value of the variable being measured.
Figure 4.1 illustrates the concept with a set of measurements of a temperature
of boiling water. Figure 4.1a corresponds to one thermometer, and Figure 4.1b
corresponds to another that has a similar average but larger deviations from the
average.
Statistical theory is used to define precision. If the distribution of errors is
assumed to be gaussian, then the variance of this distribution is estimated by
105 105
103 103
Temperature
Temperature
101 101
99 99
97 97
95 95
1 250 1 250
Time Time
(a) (b)
FIGURE 4.1. Precision.

using the standard deviation of a sample.

sP
n
i=1 (x i − x)2
s= (4.1)
n
where xi are the sample values, x the average and s the standard deviation. A
statistically unbiased estimate of the variance of a normal distribution σ 2 is the
modified variance ŝ 2 where
r
n
ŝ = s (4.2)
n−1
As n measurements are taken, statistical estimation theory states√that the

population mean µ has p% probability to be in the interval x ± z p ŝ/ n. The
parameter z p is called confidence coefficient and is obtained by determining
what value of x is such that the area under the curve in the interval [x, −x] is
equal to p. In turn, p is called the confidence level, or confidence limit. For
example for p = 0.95 (95%), the confidence coefficient is z p = 1.96, as it is
illustrated in Figure 4.2.
Because in practice the number of measurements is finite, the use of the
t-student distribution is more appropriate. The confidence coefficient of this
distribution (t p ) is defined as is the p% confidence coefficient for (n − 1)
degrees of freedom. When n ≥ 30 the t-student and the normal distribution
are practically the same. Precision (σ P ) is thus defined as half of the confidence
FIGURE 4.2. Normal distribution.

interval when one measurement is taken, that is:
σ P = t p ŝ (4.3)
Precision is more formally called repeatability or reproducibility by the

International Society for Measurement and Control (ISA) and other literature
sources. Repeatability is defined by ISA standards as the closeness of agree-
ment among a number of consecutive measurements of the output for the same
value of the input under the same operating conditions (ambient temperature,
ambient pressure, voltage etc), approaching from the same direction. Repro-
ducibility is defined the same way, but approaching from both directions. Thus,
repeatability does not include hysteresis, dead band and drift effects (discussed
below), whereas reproducibility does. Precision is thus a loose term coined by
practitioners that does not have a standard definition. Therefore, its usage is also
ambiguous because it can either refer to repeatability or reproducibility. From
now on, unless indication to the contrary, it will substitute for reproducibility.
Origin of Fluctuations
Fluctuation of measurements has varied sources. For example, pressure mea-

surements are affected by small fluctuations originated in pumps and/or vibra-
tions of compressors and other factors. Process temperatures are affected by
ambient temperature fluctuations etc. Turbulent flow is by definition full of
fluctuations, and even laminar flows are subject to such variations, because
flow is driven by pressure differences and depends on density. Thus, there are
two kinds of fluctuations: (1) inherent to the process and hence to the variable
measured and (2) external disturbances on the measurement process.
Assumption of Normal Distribution
Random errors are assumed to be normally distributed. This assumption is

based on the theory that errors are the product of innumerable sources, and,
consequently, the central limit theorem of statistical theory applies. This the-
orem states that a sum of a large number of disturbances each having its own
distribution tends to give a disturbance with a normal distribution. Measured
values are related to the state variables that they measure through a series of
signal transformations. These transformations involve, among others, the use
of measurement devices, transducers, electronic amplifiers and final reading
instruments. Thus, signals are distorted when these nonlinear transformations
are performed. One example is the square rooting performed when differential
pressure signals are used to measure flow rate. Even if the original distribu-
tion is gaussian, the resulting distribution is skewed. Bagajewicz (1996) proved
that this effect can have an impact of sizable proportions in data reconciliation
105
104
103
Temperature 102
101
100
99
98
97
96
95
1 250
Time
FIGURE 4.3. Systematic error.
practice. In addition, noise is usually assumed to have a normal distribution.

However, this is sometimes not true. For example, when noise has a narrow
frequency band width, a Rayleigh distribution is more appropriate (Brown and
Glazier, 1964). Finally, it is a well-known fact that oscillating signals have prob-
ability distributions that are not normal (Himmelblau, 1970). The consequences
of this nonnormal noise probability distribution are not explored here.
Despite these counterexamples, the underlying assumption that all errors are
gaussian is common to all the theories of data reconciliation and gross error
detection in process plants.
Systematic Error (Bias)
Consider a variable that maintains its value static, for example, the boiling
temperature of a pure fluid. Assume that a number of consecutive measurements
of this variable is made and their mean is calculated. The systematic error of
the instrument is defined as the closeness of agreement of the mean value of the
measurements with the true value of the variable. It is also called bias. Figure 4.3
illustrates the concept with a set of measurements of a temperature of boiling
water. The measurements in this figure have a systematic error of about +2◦ C.
When the true value is known the amount of bias (δ) can be estimated by
subtracting the mean value of all measurments from the true value(x̂):
δ = x − x̂ (4.4)
Thus, when the measurements are higher than the true value, the bias is pos-
itive and the reading is said to be high. Conversely, a negative bias corresponds
to a low reading. When true values are not known, other instruments are needed
to determine a good estimate of these “true values.” This process is called cal-
ibration. For example, temperature meters can be calibrated by using systems
with well-known values of temperature, such as the boiling or melting point of
pure substances. Flowmeters, in turn, need to be calibrated with the aid of some
other more precise instruments.
Classification of Systematic Errors (Biases)
Biases can be classified in two major categories: constant and variable biases.
Sources for constant biases are:
—use of incorrect assumption in the calibration procedure. The assumption of
ideal gas behavior in a gas flowmeter calibration is one example.
—corrections not performed in the calibration procedure
—unknown errors in reference standards
—incorrect installation of the instrument. For example, installing a flowmeter
close to an upstream elbow.
—zero shift
Sources of variable biases are:
—drift in the voltage supply to the instrument
—span shift
—wear of the instrument. For example, the orifice edge of a flowmeter can be
affected by particles of a dirty stream.
Drift is defined as a change in output over a specified period of time for a
constant input. Shifts are, in turn, independent of time and correspond to errors
in the measurement range. Span and zero shifts are illustrated in Figure 4.4.
Span shift
Zero shift
100
% Output
Specified
characteristic
100
% Input
FIGURE 4.4. Span and zero shift. Adapted from Liptak, 1997.
104
103
Temperature 102
101
100
99
98
97
96
0 Time 250
FIGURE 4.5. Gaussian signal.
Outliers
An outlier is defined as a measurement that can never be explained, calculated,

estimated or anticipated. Human errors, electrical voltage surges and wiring
problems are typical sources of outliers. In Figure 4.5, a signal without outliers
is presented, whereas in Figure 4.6, a signal with small variance and outliers is
presented.
Figure 4.7 shows the effect of a high signal-wire resistance that causes erratic
behavior and introduces a large number of outliers. In turn, Figure 4.8 shows a
well-wired instruments with an occasional outlier.
Sensitivity and Speed of Response
Sensitivity of an instrument is the smallest change in the measured variable

to which an instrument reacts with a change in reading. Speed of response, or
lag, is a dynamic characteristic that describes the reaction of an instrument to a
measured variable that changes with time. Few sensitivity and lag problems are
104
103
102
Temperature
101
100
99
98
97
96
0 Time 250
FIGURE 4.6. Gaussian signal with outliers.
104
103
102
101
100
99
98
97
96
0 250
FIGURE 4.7. Effect of signal-wire resistance; badly wired instrument. Adapted from
McMillan et al., 1998.
encountered because most measuring devices have acceptably good response.

Fine sensitivity causes cycling and results in poor control. For example, ther-
mocouples are often put in thermowells to protect them from the process fluid.
The presence of these thermowells create a time lag in the response of the
instrument.
Hysteresis and Dead Band
Hysteresis, illustrated in Figure 4.9, is a phenomenon where an output cor-

responding to an increasing input varies through a certain path, which differs
from the path the output follows. This occurs when the input is decreased from
the maximum reached until the original value. The discrepancy between the
paths is typically larger if the interval of change is larger. Dead band, illustrated
104
103
102
101
100
99
98
97
96
0 250
FIGURE 4.8. Effect of signal-wire resistance; properly wired instrument. Adapted from
McMillan et al., 1998.
Output Output Output
Input Input Input
(a) (b) (c)

FIGURE 4.9. (a) Hysteresis. (b) Dead band. (c) Hysteresis and dead band.
in the same figure (and in combination with hysteresis) is a range within which
the input can be changed without observing a change in the output. This is
typically observed when the input direction of change is suddenly reversed.
Linearity
When the expected measured output value is the same variable as the input
value, then the expected output is a straight 45◦ line. This is, for example, the
case when the flow measured by a flowmeter is plotted against the real flow,
after all compensation and conversions have been performed. This diagram is
shown in Figure 4.10.
The conformity between the actual measured value and the characteristic
line is called independent linearity and is obtained by making opposite sign
Independent straight
line
100
Actual calibration
curve
% Output
Maximum positive
deviation
Maximum negative
deviation
100
% Input
FIGURE 4.10. Independent linearity. Adapted from Liptak, 1997.
maximum deviations equal. These deviations from linearity are called static
errors.
Accuracy
The accuracy of a measurement is defined as the degree of conformity with

a standard or true value. Therefore, precision and systematic errors contribute
to the accuracy of an instrument. Thus, a device is said to be accurate if it is
unbiased and precise.
Accuracy (σa ) is many times reported numerically as follows:
σa = δ + σ P (4.4)
In many other textbooks, accuracy is defined in the same way as precision,
ignoring (or assuming) that bias is not present. Precision, bias and accuracy
are also often reported in relative terms, that is, as a percentage of the value
measured. Some typical expressions are given in terms of:
—the measured variable: ±1◦ C
—a percentage of span: ±0.5%
—a percentage of the upper range value: ±0.5%
—a percentage of actual output reading: ±0.3%
The International Society for Measurement and Control (ISA) and the
American National Standards Institute (ANSI) define accuracy as the sum of
linearity, hysteresis and repeatability, compared with two fixed points. As it was
pointed out by Lewis (1998), “nothing is said whether the two fixed endpoints
are in the right place,” and, therefore, the definition he proposes may be thought
of as “relative accuracy.”
Modern nomenclature, not reflected in ISA standards, uses the term “inac-
curacy” instead of “accuracy” to indicate the degree of departure from standard
or ideal values. Although the former term is preferred here, “accuracy” is so
deeply rooted in the sensor location and data reconciliation literature that it will
be nonetheless used throughout this book.
CALIBRATION CURVES
Instruments are used over a certain range. Therefore, one might expect the
precision, bias and accuracy to vary in the range of the measurement. A graph
that shows the ideal values of an output variable as a function of the input
variable is called a specific characteristic curve. For example, an orifice flow-
meter measures the pressure drop as a function of the volumetric flow rate. Thus,
we expect the characteristic curve of such device to be a square root function.
The calibration procedure consists of obtaining the actual value of the output
Calibration Curves 75
Specified characteristic
curve
100 Maximum positive
deviation
Output
Actual calibration
curve
Maximum negative
deviation
100
Input
FIGURE 4.11. Independent conformity. Adapted from Liptak, 1997.
variable by averaging the upscale and downscale readings. The comparison

between the actual and characteristic curves is called independent conformity.
This is illustrated in Figure 4.11.
To construct such curves, another instrument is needed to determine the
“true” value of the input. In practice, the device used for this purpose must have
at least five times better accuracy than the other device and must be free from
systematic errors (Miller, 1996).
Although its shape is known, in many cases the characteristic curve cannot be
constructed because many parameters are not known. For example in the case
of the orifice flowmeter, one might not be able to calculate the actual value but
anticipate that the ideal behavior is proportional to a square root function. Then,
to determine conformity one needs to adjust the characteristic curve. When this
adjustment is done by using the upper range value, the conformity is called
terminal-based conformity. On the other hand, when the initial value is used,
the conformity is called zero-based conformity (Figure 4.12).
Thus, only in a few intervals within the instrument range one can say that
a bias is negligible. Thus, one can think of conformity as the deviation of
the actual measurement from an expected theoretical characteristic curve. In
practice, electronic transformations, such as taking a square root of signals
coming from differential flowmeters, are made to obtain a linear output-input
characteristic curve. In such case, linearity establishes the difference between
the actual and expected values.
Flow Compensation
Measurements are typically reported to the Distributed Control System (DCS)

as percentages of the full span. In the case of differential flowmeters, there
Actual calibration Actual calibration

curve curve
100
100
Output
Output
Specified Specified
characteristic characteristic
curve curve
Maximum Maximum
deviation deviation
100
Input Input 100
(a) (b)
FIGURE 4.12. (a) Terminal based conformity. (b) Zero-based conformity. Adapted from
Liptak, 1997.
should be concern about the accumulation of errors, despite the accuracy of

the flowmeter itself. For example, after the square root of the signal is taken,
this resulting value is multiplied by the so-called meter factor, which is a con-
stant for the particular flowmeter and fluid running through it and is supposed
to contain the effect of density, temperature etc. Moreover, the meter factor
is a result of certain base conditions, which are nothing more than the as-
sumptions used to calculate it, including the flow conditions (temperature and
pressure).
In many cases, the DCS performs automatic adjustment of the meter factor to
compensate for deviations from the base conditions. This procedure is known as
flow compensation. For example, in the case of gases, knowledge of differences
in pressure and temperature can be used to do this compensation (assuming
that the concentration has not changed). In the case of liquids, temperature
suffices. However, when the concentration is not known or it varies, the density
needs to be inferred. In the case of petroleum liquids, the Watson K factor is
a useful tool to transform petroleum fraction laboratory distillation curves into
densities. However, the corrections for liquid are not satisfactory, because the
lab information is not available on real time. With the introduction of databases
(historians), this situation is changing because the data can be adjusted later for
monitoring and accounting calculations.
ACCURACY OF DIFFERENT INSTRUMENTS
Table 4.1 summarizes the accuracy and rangeability one can expect from the
different types of flowmeters described in Chapter 3.
Accuracy of Different Instruments 77
TABLE 4.1. Accuracy and Rangeability of Flowmeters.
Technology Accuracy (%) Rangeability
Differential pressure
Orifice and multivariable flow transmitter 0.5–1.0 R 8:1
Orifice and differential pressure transmitter 1.0–5.0 R 4:1
Velocity
Electromagnetic 0.25–1.0 R 30:1
Turbine 0.15–0.5 R 10:1
Vortex 0.65–1.35 R 25:1
Ultrasonic Doppler 2.0–5.0 R 5–25:1
Ultrasonic transit time 1.0–5.0 R 10–2750:1
Volumetric
Positive displacement 0.15–0.5 R 10:1
Mass
Coriolis 0.10 R 80:1
0.15 R 20:1
0.3 R 10:1
Thermal mass (gases) 1.0 R 50:1
Thermal mass (liquids) 0.5 R 50:1
Variable area (metal tube) 2–10 R 10:1
Source: Pomroy (1996), Dolenc (1996).
Pressure measurements can be as accurate as their theoretical limit (0.01%).

However, in practice, they approach the range of 0.1–1%. Strain gauge pressure
transducers can achieve accuracies of 0.5%. Solid-state pressure transducers
have similar range of accuracy (Wightam, 1972). The accuracy of temperature
and density measurements is given in Tables 4.2 and 4.3, respectively.
TABLE 4.2. Accuracy and Range of Temperature

Measurement Devices.
Technology Accuracy Range (◦ F)
Thermal expansion
Filled thermal 0.5–2.0◦ F –300 to 1000
Bimetallic 1–20◦ F –80 to 800
Electrical
Thermocouple 0.4–0.75% –260 to 2320
(0.2◦ F, best)
Thermoresistance 0.01–0.12% –250 to 850
(0.1◦ F, best) –80 to 150
Thermistor 1–15% –150 to 600
Optical
Pyrometer 0.5–1% −50 to 3000
Source: Johnson (1998), Zientara (1972).
TABLE 4.3. Accuracy and Range

of Densitometers.
Technology Accuracy (sp. gr.)
Buoyancy 0.0001
U-tube (vibration type) 0.001
U-tube (force balance) 0.0005
Radiation (nuclear) 0.0001
Source: Wightman (1972).
NOMENCLATURE
s: standard deviation
ŝ: modified standard deviation
tp: confidence coefficient
p: confidence
x: mean value of all measurements of x
x̂: true value of x
Greek Letters
σa : accuracy
σ P : precision
δ: bias
REFERENCES
Bagajewicz M. On the Probability Distribution and Reconciliation of Process Plant Data. Pro-
ceedings of the Process Systems Engineering (PSE) International Symposium. Kyongju, Korea,
(1994). Comp. & Chem. Eng., 20, N 6/7, p. 813 (1996).
Brown J. and E. V. D. Glazier. Signal Analysis. Reinhold, New York (1964).
Dolenc J. W. Choose the Right Flow Meter. Chemical Engineering Progress, Jan., pp. 22 (1996).
Himmelblau D. M. Process Analysis by Statistical Methods. Wiley (1970).
Johnson F. L. Temperature Measurement and Control Fundamentals. Chemical Processing, June
(1998).
Lewis C. W. More Precise Temperature Measurements. Chemical Engineering, May (1998).
Liptak B., editor. Instrumentation Handbook. Third Edition. Chilton (1997).
McMillan G. K., G. E. Mertz and V. L. Trevathan. Troublefree Instrumentation. Chemical Engi-
neering, Nov. (1998).
Miller R. W. Flow Measurement Engineering Handbook. McGraw Hill (1996).
References 79
Perry R. H. and D. Green. Perry’s Chemical Engineers’ Handbook, 6th Edition. McGraw-Hill, Inc.
New York (1984).
Pomroy J. Selecting the Right Flowmeter. Chemical Engineering, May (1996).
Webster J. G., editor. Measurement, Instrumentation and Sensors Handbook. CRC Press, LLC
(1999).
Wightman E. J. Instrumentation in Process Control. CRC Press, Butterworth & Co, London (1972).
Zientara D. E. Measuring Process Variables. Chemical Engineering, Sept. (1972).
CHAPTER 5
Variable Classification
INTRODUCTION
This chapter presents the classification of variables in their different cate-

gories. These categories are related to the ability to observe the system, that is,
being able to obtain an estimate of the state variables of the system, especially
when they are not measured. The concepts of observability and redundancy and
the methods to determine observable and redundant variables are reviewed first.
The unifying concept of estimability is presented next. Connections to graph
theory are later made. These concepts from graph theory will be used in future
chapters in design procedures. The extension of all these concepts to nonlinear
systems is then covered in the last part of the chapter.
MODEL
Usually, a model for a process plant is given by a mathematical relation,

typically a set of autonomous differential algebraic equations (DAE) such as
the following one:
d x1
= g1 (x1 , x2 ) (5.1)
dt
g2 (x1 , x2 ) = 0 (5.2)
For convenience, the set of state variables x has been partitioned into (x1 , x2 ).
In addition, x is usually divided into measured variables x M and unmeasured
variables xU . In using this model, we depart from the traditional input-output,
or input-state-output representations so popular in control theory.
81
82 VARIABLE CLASSIFICATION
MEASUREMENT EQUATION
State variables describe the behavior of a system. However, these variables

may not be the ones that are being directly measured. Therefore, we introduce
the measurement equation:
z = h(x) + ε (5.3)
where h(x) is a measurement function and ε is the measurement random error.

Thus, the objective of monitoring is to be able to obtain the value of a subset
of state variables of the set (x) by using the knowledge of the measurement
values z. In the simplest case, h(x) is linear. Thus, an estimator of x (calibration
graphs, filters, data reconciliation etc.) performs the inversion of Equation (5.3)
as follows:
e
x = g(z) (5.4)
In the rest of the book, the distinction between x and its estimate e
x will be
omitted with the only exception of Chapter 7, where it is again used.
GRAPHS AND FLOWSHEETS
A graph is defined as a set of vertices connected by edges. A flowsheet is a

graphical representation of a process, also known as a digraph. It consists of
units (nodes) and streams (edges). A graph is called directed when the edges
(streams) have an established direction from one node to the other. Undirected
graphs do not have any preferred direction; therefore, the edges can be traveled
in any direction. Thus, a directed graph is called a digraph. Figure 5.1 illustrates
U2
(purge)
Compressor S3 S4
S9 S8 S7
U5 S1 S6
U1 U3
U4
S2
S5 S9 S5
U2 Cooler S4
S1 (hydrogen) S3
Reactor U3 Flash
S2 (benzene) S8
U1 U5 U4
Heater
Benzene S6 S7
ENV
(a) (b)
FIGURE 5.1. Simplified cyclohexane production process network. (a) Flowsheet. (b) Cor-
responding graph.
Connectivity of Systems 83
the relationship between flowsheet and graphs. Groups of units are sometimes
represented by one node in the graph, as it happens with the set heater-reactor-
cooler, which is represented by node U2 . Because, by definition, an edge has to
connect two vertices, sometimes the environmental node is added when building
the graph of a flowsheet. This type of graph is called the augmented graph, and
is shown in Figure 5.1(b).
CONNECTIVITY OF SYSTEMS
A system is connected if in the corresponding undirected graph there is a

path from every unit to any other unit. Consider the system in Figure 5.2. This
system is connected. However, removal of process U5 makes it a disconnected
system composed of two connected systems.
Although connectivity of small systems can be determined by simple visual
inspection, in large flowsheets this might not be so obvious.
There are two ways to determine if a system is connected: using graph theory
or manipulating the incidence matrix. Both are equivalent and are based on the
same principle.
Graph Approach to the Determination of Connectivity
Take a node (unit). Merge it with any unit to which it is connected through
an edge (stream). Repeat the procedure until no merging is possible. Then
count the number of merged units left. This number is the number of connected
systems. This process is shown in Figure 5.3. Figure 5.3(a) shows the graph
corresponding to Figure 5.2. In Figure 5.3(b), nodes U2 , U3 , U7 and U8 are
merged with the splitters in front of them. In Figure 5.3(c), mixer U4 is merged
with splitter U1 and the second splitter, U6 , is merged with node U5 . The two
nodes left are merged in Figure 5.3(d) to leave only one node. Suppose now
that unit U5 is not originally present. Clearly, the process will end in a graph
similar to the one of Figure 5.3(c), with two nodes (U1 and U6 ) not connected
to each other.
S2 U2
S4 S8 S10
U7
S1 U1
S6 U5
S7
U4 U6
U3 U8
S3 S5 S9 S11
( ):Measured flow rates
FIGURE 5.2. A connected system.

FIGURE 5.3. Steps for the determination of connectivity.
Incidence Matrix Approach to the Determination of Connectivity
The incidence matrix is constructed as follows:

r Each row represents a unit and each column represents a stream.
r In each row, streams entering to the corresponding unit are represented by a
number 1. Streams leaving the unit are represented by a number −1.
The incidence matrix (C) for the process in Figure 5.2 is:
S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11
 
1 −1 −1 U1
 1 −1  U2
 
 1 −1  U3
 
 1 1 −1  U4 (5.5)
C =


 U5
 1 −1 
 1 −1 −1  U6
 
 1 −1  U7
1 −1 U8
The method to determine connectivity based on the incidence matrix is:

r Take the first row of the incidence matrix. Find a column with a nonzero
element.
Observability 85
r Search in that column for another row with a nonzero element. By construc-
tion, this element will have an opposite sign.
r Add the latest row found to the first row and eliminate the row and the column.
r Repeat the procedure for the remaining rows (if any).
r Count the number of rows left. This number is the number of connected
systems.
This procedure is illustrated next for the system shown in Figure 5.2. Suc-
cessive application of the procedure for the first row leads to a new matrix
C 0:
S1 S10 S11
0
(5.6)
C = [1 −1 −1]
which proves that the system is connected. The reader can verify that in
the case where unit U5 is not present the process ends with the following
matrix:
· S1 S6 S7 S10 S11 ¸
0 1 −1
C = (5.7)
1 −1 −1
In the analysis that follows and in the rest of the chapters of the book, we
assume that all systems are connected.
OBSERVABILITY
Observability of a variable in a system can be defined in a broad sense as the

ability to perform an estimate of the variable by using certain measurements per-
formed in the system. When discussing the measurement equation, it was made
clear that often variables are not measured directly, but, rather, some other parts
of the system are measured. Thus, in simple terms, determining the observability
of a variable consists of determining if the function g(z) exists. Kalman (1960)
introduced the concept of observability for linear dynamic systems. The semi-
nal articles by Griffith and Kumar (1971), Kou et al. (1973) and Singh (1975)
discussed observability in nonlinear systems. Stanley and Mah (1981) discussed
this issue of observability of systems described by steady-state models in depth.
In particular, they discuss conditions under which observability can be attained,
conditions that are related to the nature and structure of the mappings h(x) and
g(z). We now concentrate on a variable classification for linear steady-state
systems.
The system shown in Figure 5.2 consists of 5 units and 11 streams. Consider
that the hold-up of units U1 through U8 are negligible. Thus, all balances in the
system do not contain derivatives and are represented by equations of the form
shown by Equation (5.2).
Notice first that all the flow rates after the first split (S2 , S3 , S4 , S5 ) cannot be
calculated, because there is no material balance that can be used to obtain them.
These variables are called unobservable variables. The rest of the unmeasured
variables (S6 , S9 , S10 ,) can be obtained from material balances by using the
measured values. We call these variables, observable variables. This leads to
our first classification:

Measured (M)
½Observable (O)
Variables

Unmeasured (U )
Unobservable (UO)
Thus, we formally define observability as follows:
Definition: A non-measured variable is observable if it can be calculated in

at least one way from the measurements.
REDUNDANCY
Consider now the measured streams in Figure 5.2. If the flow rate of stream
S1 is not measured, it can be estimated in only one way, that is, by using the
value of S7 . However, there is also an additional way: by adding S8 and S11 .
In this last case, we thus say, that the system {S1 , S7 , S8 , S11 } is redundant.
Assume now that stream S7 and S8 are not measured. Then S1 could not be
estimated by any balance equation and its removal makes it unobservable: we
call it nonredundant. Thus, the following definitions follow:
Definition: A measurement is redundant if the corresponding variable can

be calculated in at least one way using the remaining measurements.
Definition: A set of measurements of a system is redundant if all measure-

ments are redundant.
Definition: A measurement of a variable is nonredundant if after removing

this measurement, the variable is unobservable, that is, it cannot be calcu-
lated by using a balance equation involving the other measurements of the
system.
We now complete our classification of variables as follows:

½
 Redundant (R)
measured (M)
 Non-redundant (NR)
Variables ½

 Observable (O)
Unmeasured (U )
Unobservable (UO)
Redundancy is therefore a desirable property of a system because in the case

when an instrument fails, its variable can still be estimated through balances.
Moreover, if the number of different balances that can be used increases, there
will be additional ways to calculate the variable. When more redundant mea-
surements are involved in calculating a variable, it is said that the reliability of
such system and the precision of the estimation increase. There is, therefore, a
need to distinguish these different levels of redundancy.
Hardware Redundancy
We consider now the case where more than one instrument is used to measure
the same variable. This has been traditionally the way redundancy has been
understood. Hardware redundancy has no effect on observability, but it has an
effect on the accuracy and the reliability as well as on other properties of the
sensor network. This will be discussed in Chapter 8.
LINEAR SYSTEMS
The dynamic model of a material balance in a process plant can be represented

by the following differential-algebraic system of equations (DAE):
dw
= Af (5.8)
dt
Cf = 0 (5.9)
where f are the flows of the different streams connecting the units, w are the
hold-ups of the respective units and A and C are system matrices.
CANONICAL REPRESENTATION OF LINEAR SYSTEMS
A mathematical representation is now developed to obtain an analytical way

to determine observable, unobservable and redundant variables. Consider first
the partition of the above system of equations into measured and unmeasured
variables. Let
· ¸
A −I
D= (5.10)
C 0
and
 
f
x =  dw  (5.11)
dt
Thus, the system of differential algebraic Equations (5.8) and (5.9) can be
rewritten as follows:
Dx = 0 (5.12)
By using simple rearrangement of columns in matrix D, the system can be

rewritten in the following way:
· ¸
xU
[DU DM ] =0 (5.13)
xM
where xU corresponds to unmeasured variables and x M corresponds to measured

variables.
Madron (1992) proposed a simple way of determining observable variables:
as in Gauss-Jordan elimination for solving linear systems, a linear combination
of rows combined with column rearrangements is performed to obtain a system
that has the following form:
 
xO
xUO 
[D O DUO D R DNR ]  
 xR  = 0 (5.14)
xNR
where D O and DUO , D R and DNR have the following structure:

 
I
 
D O = 0 (5.15)
0
 
0
 
DUO = G UO  (5.16)
0
 
−G RO
 
D R =  G RUO  (5.17)
GR
 
−G NRO
 
DUO =  G NRUO  (5.18)
0
The column rearrangement that leads to the identity matrix in Equation (5.15)
allows the identification of the observable variables x O . Matrix G UO has typi-
cally more columns than rows and rank equal to the number of rows. Thus, it
forms an undetermined system involving xUO , which will be ignored because
it is insoluble.
The final form of the system of equations is:
Unobservable Redundant
Observable Non-redundant
xO
I 0 - GRO - GNRO
xUO (5.19)
0 GUO GRUO GNRUO =0
xR
0 0 GR 0
xNR
Unmeasured Measured
This is called the canonical form of D. The last set of variables is called
nonredundant; their measurement has to be taken at face value. In the context
of dynamic data reconciliation, these variables have been called self-redundant,
because the set of measurements can still be fitted to a curve.
Rewriting the system, one has
x O = G R O x R + G NRO xNR (5.20)

G RxR = 0 (5.21)
Remark 1: G R represents a graph that corresponds to a redundant system

only. We call this graph, the redundant graph of the system. In addition, the
redundant system matrix is of the form
¥ ¦
G R = I G ∗R (5.22)
Finally
¥ ¦
G R O = 0 G ∗R O (5.23)
Remark 2: When there are redundant measured hold-up, matrix G R has the
following structure
· ¸
A R −B R
GR = (5.24)
CR 0
Therefore, Equations (5.8) and (5.9) are rewritten in terms of its specific
parts as follows:
dw R
BR = AR fR (5.25)
dt
CR fR = 0 (5.26)
Remark 3: Knowledge of x R and xNR allows the determination of the ob-

servable values. This is called co-optation. However, because measurements
do not comply with Equation (5.21), adjustments are needed. The determina-
tion of estimates of x R such that Equation (5.21) is satisfied is called data
reconciliation.
Illustration
Consider the system of Figure 5.2. Its incidence matrix is given by Equa-
tion (5.5). The canonical form of this matrix obtained by Gauss-Jordan factor-
ization is:
S6 S9 S10 S2 S3 S4 S5 S1 S7 S8 S11
1 -1
1 -1 1
1 -1
1 1 -1
C=
1 -1 (5.27)
1 1 -1
1 -1
1 -1 -1
Observable Unobservable
Redundant
S6
S8
U1 U2 U3 U4 U5
S1 S2 S3 S4 S5 S9
S7
( ):Measured flow rates and Hold-ups

FIGURE 5.4. A system with tanks.
If in addition, variable S3 is measured, then the result is:
S6 S9 S10 S2 S4 S5 S1 S7 S8 S11 S3
1 -1
1 -1 1
1 -1
1 -1 1
C=
1 -1 1 (5.28)
1 -1
1 -1
1 -1 -1
Observable
Redundant Non-redundant
We now turn to another example that illustrates the changes in unit hold-up.
Consider the system depicted in Figure 5.4. It contains nine streams, four tanks
and one splitter. Assume that all hold-up and flow rates are measured except
the hold-up of tank U3 , and the flow rates of S2 and S6 .
Matrix D is then given by:
S1 S2 S3 S4 S5 S6 S7 S8 S9 U1 U2 U3 U4
 
1 −1 1 −1
 1 −1 1 −1 
 
D= 1 −1 −1 −1  (5.29)
 1 −1 −1 −1 
1 −1 −1
The canonical form of D is:
S2 S6 U3 S4 S5 S7 S8 S9 U4 U1 U2 S1 S3
1 1 -1 -1
1 1 -1 -1 1 -1
D' = 1 1 -1 1 1 -1
1 -1 -1 -1 -1
1 -1 -1
Observable Redundant Non-redundant

(5.30)
As a result of the procedure, only two equations (last two rows) represent
the redundant system of equations. In addition, all variables are in this case
observable. The corresponding matrices are:
" #
1 0 −1 −1 −1 −1
GR = (5.31)
0 1 0 −1 −1 0
 
0 0 −1 0 0 0
 
G RO =  0 0 −1 0 0 0 (5.32)
−1 0 1 0 0 0
 
0 1 0 1
 
G NRO = 1 1 −1 1 (5.33)
−1 −1 1 0
Thus, U4 , S4 , S5 , S7 , S8 and S9 have been identified as redundant variables.

In particular:
)
A R = [1 0 −1 −1 −1] BR = 1
(5.34)
C R = [0 1 0 −1 −1]
which correspond to the following system of equations:
dw4
= f4 − f7 − f8 − f9 (5.35)
dt
f5 − f8 − f9 = 0 (5.36)
The graph of the redundant system can be depicted as Figure 5.5.

Other Methods 93
S8
S4 U4 S5
U5
S9
S7

FIGURE 5.5. Redundant system.
Thus, the nonredundant (self-redundant) variables are U1 , U2 , S1 and S3 and

the observable variables are U3 , S2 and S6 , as it is indicated by the first three
columns of the canonical form of D.
OTHER METHODS
A few other methods exist for the observability analysis of linear systems.
The pioneering work in this field was performed by Vaclavek (1969) and
Romagnoli and Stephanopoulos (1980). More recently, matrix projection and
QR decomposition have been proposed. We briefly illustrate these two methods
next.
Matrix Projection
This method was proposed by Crowe et al. (1983), and it consists of deter-
mining a matrix P such that PDU = 0. Thus, by premultiplying Equation (5.13)
by P one obtains:
PDU xU + PD M x M = PD M x M = 0 (5.37)
which leaves a system of equation in terms of the measured variables only.

Matrix P is obtained as follows. Assume DU is partitioned as follows:
· ¸
DU,1 DU,3
DU = (5.38)
DU,2 DU,4
Without loss of generality we assume that DU,1 is the largest nonsingular sub
matrix of DU . Then P is given by:
¥ −1
¦
P = −DU,2 DU,1 I (5.39)
Indeed
¥ −1
¦
PDU = 0 −DU,2 DU,1 DU,3 + DU,4 (5.40)
However, the right half columns of DU are linear combinations of the left
−1
half. Thus, −DU,2 DU,1 DU,3 + DU,4 = 0.
QR Decomposition
This method was proposed by Swartz (1989). Let r = rank(DU ). Then the
method consists of multiplying matrix DU by a permutation matrix 5, such
that the first r columns are l.i. Then, DU 5 can be uniquely written as follows:
· ¸
R11 R12
DU 5 = [Q 1 Q2] (5.41)
0 0
where Q 1 and Q 2 are orthonormal and R11 is upper triangular (Stewart, 1973;
Sánchez and Romagnoli, 1996). These matrices can be obtained with one code
instruction in MATLAB. We rewrite Equation (5.41) as follows:
· ¸
R11 R12 T
DU = [Q 1 Q2] 5 (5.42)
0 0
Thus, premultiplying xU by 5T one obtains a permutation of unmeasured

variables [xU,1 xU,2 ], that is, 5T xU = [xU,1 xU,2 ]. Therefore, we premultiply
Equation (5.13) by Q T to obtain
" # x 
R11 R12 Q 1T D M U,2
xU,1  = 0 (5.43)
0 0 T
Q2 DM xM
One equation derived from Equation (5.43) is:
−1
£ T ¤
xU,1 = −R11 Q 1 D M x M + R12 xU,2 (5.44)
It follows that xU,2 is always unobservable. However, the set xU,1 contains
observable and unobservable variables. Indeed, a variable in xU,1 is observable
−1
if the corresponding row of R11 R12 is zero.
Finally, Q 2T spans the null space of DUT , therefore Q 2T DU = 0, indicating that
Q 2T works in the same way as the projection matrix P.
The use of this method for observability analysis will be illustrated later in
this chapter for nonlinear cases.
System Degree of Redundancy 95
SYSTEM DEGREE OF REDUNDANCY
We are now in a position to give a mathematical interpretation of the degree

of redundancy. First, note that matrix GR is of full rank. If GR is not of full rank,
there would be one balance that can be written as a linear combination of the
others. However, by construction, D has as many rows as units, and, therefore,
all balances are linearly independent. Thus, since the degree of redundancy is the
number of l.i. balance equations that can be written with measured quantities,
it is equal to the rank of G R . This suggests the following definition:
Definition: A system has degree of redundancy k, when at least k linearly
independent balance equations can be written by using measured variables
only.
In the case of Figure 5.2, the degree of redundancy is two, as it follows from
Equation (5.27). If S3 is measured, the degree of redundancy is still two.
Property: The maximum degree of redundancy of a system is equal to the
number of units.
The property is self-evident from the fact that this is the maximum number
of l.i. material balances that can written. This prompts the following definition:
Definition: A system has full redundancy when all measurements are redun-
dant.
In addition, the system degree of redundancy is an indication of the appro-
priateness of the measurement location. For example, two systems can have the
same number of measurements but different system degree of redundancy, as
in one case too many variables are redundant and in the other, too many are
nonredundant. Thus. the concept of system degree of redundancy does not fully
encompass the richness of the different types of system redundancies that one
can find. For example, systems can have the same number of measurements
and the same number of units, but different degrees of redundancy.
Consider the system of three units depicted in Figure 5.6(a). Assume all
streams are measured. Thus, it contains three units and four measurements,
resulting in a degree of redundancy of three. Consider now the system in Fig-
ure 5.6(b), and assume that all but S5 and S6 are measured. As in the case
of Figure 5.6(a), this system contains three units and four measurements.
However, its degree of redundancy is one. If the set of measured streams is
x M = {S1 , S2 , S3 , S6 }, then the system degree of redundancy is two.
Thus, the degree of redundancy of a system is a condition that reflects how
effectively a certain number of measurements is distributed throughout the
system. However, we have not found much connection with specific goals,
such as reliability or gross error detectability, resilience or the residual precision
S2 S4
S1 S2 S3 S4 S1 S5
S3 S6
(a) (b)

FIGURE 5.6. Illustration of system degree of redundancy.
under sensor failure, other than the fuzzy statement of saying that “the more
redundancy, the better.” For this reason, one needs to be more specific and talk
about attributes of specific variables at specific locations. This will be discussed
in later chapters.
QUANTIFICATION OF OBSERVABILITY AND REDUNDANCY
We now resort to a few definitions that will allow the quantification of ob-
servable and redundant variables. A unification of both concepts under a single
notion of estimability is presented next.
Estimability
A generalized definition of observability was attempted by Ali and

Narasimhan (1993) to denote as observable any variable, measured or unmea-
sured, for which an estimate can be produced. Estimability is a term preferred
to avoid confusion and reserve the name observable to unmeasured variables,
because it has become popular in the literature. The definition of estimability
is formally presented next.
Definition: A variable Si is estimable if it is measured or unmeasured but

observable.
Degree of Observability of Variables
For convenience, we denote 2( p) as the set of all possible combinations of

p measurements. We call 2 j ( p), the jth element (combination) of this set. We
are now ready for the following definition:
S2 S4
U2
S1 S6
U1 U4
S3 S5
U3
FIGURE 5.7. Illustration of the concept of degree of observability.
Definition: An unmeasured variable Si has degree of observability Oi if:
a. It remains observable after the elimination of any combination 2 j (Oi − 1) ∈

2(Oi − 1), and
b. It becomes unobservable when at least one set 2 j (Oi ) ∈ 2(Oi ) is eliminated.
Notice the words “any” and “at least” in this definition. They are crucial
for the understanding of the concepts. In addition, also notice that the two
conditions need to hold simultaneously.
Consider the system of Figure 5.7 and assume that x M = {S1 , S2 }. Variable
S6 has degree of observability O6 = 1, because just the elimination of the
measurement in S1 makes it unobservable. Part a of the definition does not
apply, because Oi − 1 = 0.
If for example S3 is also measured, that is, x M = {S1 , S2 , S3 }, then S6 would
have degree of observability O6 = 2, because elimination of one measurement
at a time (S1 , S2 or S3 ) would not make it unobservable. However, a deletion
of any of the following two sets (S1 , S2 ), (S1 , S3 ), would render it unobserv-
able. Note, however, that the elimination of the set (S2 , S3 ) would not make S6
unobservable.
Degree of Redundancy of Variables
For convenience, we denote ϑ( p, Si ) as the set of all possible combinations of

p measured variables, not including the measured stream Si . We call ϑ j ( p, Si )
the jth element (combination) of this set. The following definition is then
presented.
Definition: A redundant measured variable Si has degree of redundancy Ri

if:
a. It remains redundant after the elimination of any combination ϑ j (Ri −1, Si ) ∈

ϑ(Ri − 1, Si ), and
b. It becomes nonredundant when at least one set ϑ j (Ri , Si ) ∈ ϑ(Ri , Si ) is
eliminated.
Remark 4: Accordingly, the degree of redundancy of a nonredundant mea-

surement is zero.
In the system of Figure 5.7, for x M = {S1 , S2 }, variable S1 has degree of
redundancy R1 = 0 because it is already nonredundant. If S3 is measured, that
is, if x M = {S1 , S2 , S3 }, then S1 has degree of redundancy R1 = 1, because it
is sufficient to eliminate S2 or S3 to make it nonredundant.
A redundant measurement is such that the variable becomes observable when
the measurement is eliminated. A nonredundant variable, in turn, becomes
unobservable if its measurement is eliminated. Thus, if a variable has degree
of redundancy Ri , the elimination of its measurement will make it a variable
with degree of observability Oi = Ri . For example if for Figure 5.7, x M =
{S1 , S2 , S3 }, then the elimination of S1 , makes it a nonmeasured variable with
degree of observability O1 = 1, because it is enough to eliminate S2 or S3 to
make it unobservable.
These definitions of degree of observability and redundancy were first intro-
duced by Maquin et al. (1991, 1995). The version presented above is the result
of slight modifications made by Bagajewicz and Sánchez (1999). In addition,
Maquin et al. (1991, 1995) introduced the concept of degree of redundancy
of unmeasured variables. This concept is not used in this book, because it is
preferred to leave the concept of redundancy of variables confined to measured
ones.
The concept of estimability, which unifies degree of observability and degree
of redundancy into a single concept, is presented next. This will become im-
portant when the design of sensor networks is attempted, as requirements can
be easier established mathematically.
Degree of Estimability of Variables
Definition: A variable Si (measured or not) has degree of estimability E i if:

a. It remains estimable after the elimination of any combination 2 j (E i − 1) ∈
2(E i − 1), and
b. It becomes unobservable when at least one set 2 j (E i ) ∈ 2(E i ) is eliminated.
To illustrate the above definition, consider the process graph in Figure 5.8,
which includes the environment as a node, and assume all flow rates are mea-
sured, that is, x M = {S1 , S2 , S3 , S4 }.
In this case, ϑ(1, S2 ) = {(S1 ), (S3 ), (S4 )}. Therefore, stream S2 has degree of
redundancy R2 = 1 because just the elimination of S1 makes S2 nonredundant.
Because the elimination of S1 makes S2 nonredundant, and the elimination
of either S3 or S4 does not alter its redundancy status, its degree of estima-
bility is larger than 1. Consequently, it is necessary to evaluate the elements
of the set 2(2) = {(S1 , S2 ), (S1 , S3 ), (S1 , S4 ), (S2 , S3 ), (S2 , S4 ), (S3 , S4 )}. From
S1
U1 U2
S2
S3 S4
ENV
FIGURE 5.8. Illustration of degree of estimability.
this analysis, it can be easily seen that the elimination of (S1 , S2 ) makes S2
unobservable, thus E 2 = 2.
In a similar way it can be shown that R3 = 2 because if all the elements
of ϑ(1, S3 ) = {(S1 ), (S2 ), (S4 )} are individually eliminated, variable S3 stays
redundant but becomes nonredundant if the elements (S1 , S4 ) or (S2 , S4 ) from
ϑ(2, S3 ) = {(S1 , S2 ), (S1 , S4 ), (S2 , S4 )} are deleted. The inspection of 2(2) and
2(3) = {S1 , S2 , S3 ), (S1 , S2 , S4 ), (S2 , S3 , S4 )} helps to conclude that E 3 = 3,
because the elimination of (S2 , S3 , S4 ) renders S3 unobservable.
The following properties are natural consequences of the definition.
Property: The degree of estimability of a nonmeasured variable is equal to

the degree of observability.
Property: The degree of estimability of a measured redundant variable is

its degree of redundancy plus one.
Property: A measured nonredundant variable has degree of redundancy

zero and degree of estimability one.
The degree of estimability is an important property, because it does not
distinguish whether a variable is measured or not. Systems with all variables
with degree of estimability one are nonredundant, and as we shall see later,
feature a minimum number of sensors.
Lemma 5.1: A system where all variables have degree of estimability equal
to one, is a system of nonredundant measured variables and observable unmea-
sured ones.
Proof: If a variable is measured and has degree of estimability one, then

it is nonredundant. If in turn, it is unmeasured, it is observable (because its
estimability is one) and its observability is given by nonredundant variables.
Q.E.D.
Property: When the degree of estimability is larger than one for any vari-
able, then the number of sensors that need to fail before the variable in question
becomes unobservable is equal to its degree of estimability.
The property follows from the definition, and it has connections with the
concept of reliability as it will become apparent in Chapters 10 and 11.
GRAPHS AND CANONICAL MATRICES
In this section, we review a few graph theory concepts related to the canonical
matrix. Most of this material has been extracted from textbooks (Harary, 1972;
Even, 1979) and adapted to our case.
Spanning Trees
Definition: A cycle of a graph is a set of connected edges (streams) such

that the last edge is connected to the first.
Definition: A tree of a graph is a set of connected edges (streams) that does
not form a cycle.
Definition: A spanning tree T of a graph is a tree that connects all vertices
(units) of the graph.
Lemma 5.2: The identity part of D 0 , the canonical representation of matrix
D, corresponds to a set of variables that form a spanning tree.
Proof: Note first that we refer to a canonical form that does not distinguish
measured from unmeasured, that is D 0 = [I D ∗ ]. The canonical part of D 0 (the
identity matrix) corresponds to a tree because there are no cycles, and because it
has a nonzero value in every row, it is connected to all vertices (units). Q.E.D.
Definition: A branch of a spanning tree is any edge (stream) that belongs to
the spanning tree.
Definition: A chord of a spanning tree is any edge (stream) that does not
belongs to the spanning tree.
Lemma 5.3: A system where all variables have degree of estimability E i = 1,
corresponds to a system where all unmeasured streams are given by a spanning
tree.
Proof: If all variables have estimability E i = 1, this implies that all unmea-
sured variables are observable and measured variables are nonredundant. If all
unmeasured variables are observable, then the identity matrix contains all of
them, and therefore, because all the measured variables are nonredundant, there
S2 S4
U2
S1 S6
U1 U4
S3 S5
U3
ENV
FIGURE 5.9. Augmented graph for Figure 5.7.
is no row with zeros corresponding to the columns of the unmeasured variables.

Thus, the identity matrix corresponds to all unmeasured variables that form a
spanning tree. Q.E.D.
Adding one measurement to a system can have a different effect on the

estimability of the rest of the variables, depending on which measurement is
added. Consider the system of Figure 5.7 and assume the measured variables
are given by x M = {S1 , S2 }. The augmented graph for this system is shown in
Figure 5.9.
The set of unmeasured variables is a spanning tree, as it can be observed in
Figure 5.10.
The degree of estimability of all variables is one, because they are all ob-
servable, and x M is a set of nonredundant measurements.
The canonical form of the incidence matrix is:
S3 S4 S5 S6 S1 S2
1 -1
1
1 -1
1 -1 1 (5.45)
1 -1
S4
U2
S6
U1 U4
S3 S5
U3
ENV
FIGURE 5.10. One spanning tree of the system of Figure 5.9.

Assume now that S4 becomes measured. Then, the new canonical form is:
S3 S5 S6 S4 S2 S1
1 1 -1
1 1 -1 (5.46)
1 -1
1 -1
The set {S2 , S4 } is now redundant, whereas S1 remains nonredundant. How-

ever, S3 , S5 and S6 have still a degree of estimability E 3 = E 5 = E 6 = 1, be-
cause it is enough to delete S1 to make any one of them unobservable. In turn,
S2 and S4 increased their degree of estimability to E 2 = E 4 = 2.
Consider now that instead of S4 , stream S5 is measured, that is, x M =
{S1 , S2 , S5 }. Then, the new canonical matrix is:
S3 S4 S6 S5 S1 S2
1 -1 1
1 -1 (5.47)
1 -1
1 -1 1
The set {S1 , S2 , S5 } is now redundant, and their degree of estimability is

E 1 = E 2 = E 5 = 2. There are no nonredundant variables and the degrees of
estimability of all the unmeasured variables is E 3 = E 4 = E 6 = 2.
We now introduce the notion of distance between spanning trees.
Definition: The distance between spanning trees is the given by the number
of streams that are needed to be removed from one of the spanning trees and
repositioned in the other.
Consider the spanning trees of Figure 5.11. The spanning tree of Figure
5.11(a) and the one on Figure 5.10 have distance one, because it is only necessary
to eliminate S3 from Figure 5.10 and add S2 to get Figure 5.11(a). The spanning
trees shown in Figure 5.11(a) and 5.11(b) have also distance one from each
other. However, the spanning tree of Figure 5.11(b) and the spanning tree from
Figure 5.10 have distance two.
As we shall see in future chapters, spanning trees can be used to find sensor
networks featuring minimum number of instruments. These systems have all
variables with degree of estimability equal to one. It is tempting to think that
an enumeration of all spanning trees to later pick the cheapest can be a sound
procedure to design such systems. Unfortunately, this could be computationally
S2 U2 S4 S2 U2
S4
U1
S6 S1
U4 U1 U4
S5 S5
U3 U3
ENV ENV
(a) (b)
FIGURE 5.11. Two spanning tree of the system of Figure 5.9.
costly. The number of spanning trees of a given graph is given by the determinant
of L = A A T (Chen, 1971). Using this, Madron (1992) shows several examples
of how this number can climb to very large values, as large as of the order 109
for regular plants (17 nodes and 49 streams).
Cutsets
Definition: Given a graph, a cutset is defined as the set of edges (streams)

that when eliminated, separate the graph in two disjoint subgraphs.
Consider for example the system of Figure 5.9. The set {S1 , S6 } is a cutset.
When these streams are eliminated, two subgraphs are left, the original flow-
sheet (Figure 5.7) and the environmental node (Figure 5.12). The set {S2 , S4 }
is another cutset. It separates all the nodes from U2 . Finally, the set {S2 , S5 } is
not a cutset because it does not separate the graph into two sets, that is, after
elimination of these two streams, all the units of the graph remain connected.
Cutset {S2, S4}

Set {S2, S5}
S2 S4
U2
S1 s6
U1 U4
S3 S5
U3
ENV
Cutset {S1, S6}
FIGURE 5.12. Cutsets of the system in Figure 5.9.

FIGURE 5.13. A cut that is not a cutset.
In turn, the set {S1 , S3 , S5 , S6 } (Figure 5.13) is not a cutset because the elim-
ination of all these streams leaves three disjoint sets of units, {ENV}, {U3 } and
{U1 , U2 , U4 }. In fact this set is a union of two cutsets, {S1 , S6 } and {S3 , S5 }.
Remark 5: A cutset corresponds to a set of variables with which a material
balance involving a certain number of units can be written.
This was pointed out by Kretsovalis and Mah (1987b) in the context of
process systems. A cutset is, by construction, a set of streams that connects two
subsystems of the graph. Thus, because no other stream is leaving or entering
the subsystems, aside from the ones of the cutset, the sum of all the flows of
the cutset should be equal to zero. This is no other than a material balance.
Definition: A cutset is estimable if at least all but one of its streams are
measured.
This definition relates to the estimability of the variables in the cutset. If all the
variables in the cutset are measured, then all the variables are redundant, with
degree of redundancy one. If one variable is unmeasured, then it is observable,
and all the measured variables are nonredundant. Thus, an estimable cutset is a
cutset in which all the variables have degree of estimability of at least one.
Definition: An estimable cutset is redundant if all its streams are measured.
Property: The number of all cutsets containing a variable Si is equal to the
number of material balances that can be written involving variable Si .
This property is self-evident.
Lemma 5.4: The maximum possible degree of estimability of an unmeasured
variable Si is given by the number of estimable cutsets containing Si .
Proof: Consider all the cutsets containing variable Si . Consider any non-
estimable cutset. If only such cutset is considered, the degree of estimability
of Si is zero. Thus, the elimination of any variable of this cutset, except Si ,
has no effect in the estimability of Si . In fact, the estimability of Si is due

only to estimable cutsets. Assume now that Si belongs to n estimable cutsets
K 1 (i), K 2 (i), . . . , K n (i). Assume first that K 1 (i) ∩ K 2 (i) ∩ . . . · · ∩ K n (i) = Si .
Then, the elimination of a measurement belonging to K 1 (i) makes this cutset
nonestimable, but it has no effect on the observability of Si , because it will still
be guaranteed by the rest of the estimable cutsets. In fact, one variable per cutset
will have to be eliminated to render Si unobservable. Thus, Si has degree of
estimability n. However, if the intersection of any subset of estimable cutsets
contains at least one more variable in addition to Si , then these subset of cutsets
can be made nonestimable by just eliminating any measured variable belonging
to such intersection. Thus, if Si is not measured, the maximum possible degree
of estimability cannot be larger than the number of estimable cutsets. Q.E.D.
Lemma 5.5: The maximum possible degree of estimability of a measured
variable Si is given by the number of redundant cutsets containing Si plus one.
Proof: Assume first that out of the n estimable cutsets, the first m are redun-
dant. Consider again the case that K 1 (i) ∩ K 2 (i) ∩ . . . · · ∩ K n (i) = Si . Then,
the elimination of a measurement belonging to K j (i)( j ≤ m) makes this cutset
nonredundant, therefore the observability of Si does not change. The elimina-
tion of Si itself also has no effect on the observability of this variable, because
it will still be guaranteed by the rest of the measurements of the cutsets. In
fact, two variables per redundant cutset will have to be eliminated to render
Si unobservable. In general, if the measurement of Si is eliminated, then the
m redundant cutsets will become nonredundant, and to make Si unobservable
one additional measurement per redundant cutset will have to be eliminated.
Q.E.D.
Remark: The maximum degree of estimability is an upper bound of the
degree of estimability. As we shall see in the examples below, this does not mean
this degree of estimability is actually possible to be achieved for a particular
flowsheet.
Property: The degree of estimability of an unmeasured variable Si whose
estimable cutsets have only Si as intersection is equal to the number of these
estimable cutsets. If the variable is measured, then its degree of estimability is
the number of the estimable and redundant cutsets plus one.
This property is actually a corollary of the two previous lemmas.
To illustrate the above properties, consider variable S1 in the system of
Figure 5.9. All the cutsets containing S1 are: K 1 (S1 ) = {S1 , S6 }, K 2 (S1 ) =
{S1 , S2 , S3 }, K 3 (S1 ) = {S1 , S2 , S5 }, K 4 (S1 ) = {S1 , S4 , S3 }, K 5 (S1 ) = {S1 , S4 , S5 }.
All these cutsets are shown in Figure 5.14.
Assume all variables, except S1 are measured. This set of measurements
should make estimable all the cutsets containing S1 . In particular, a deletion
S2 S4 K4
U2 K5
S1 S6
U1
U4
K2
S3 S5
U3
K3
ENV
K1
FIGURE 5.14. Cutsets containing stream S 1 .
of one variable per cutset will certainly make S1 unobservable. However, the
minimum number of variables needed to render S1 unobservable is even smaller.
The elimination of the measurements in variables S2 , S4 , and S6 will render S1
unobservable. Thus, in this case, the degree of estimability of S1 is E 1 = 3, lower
than the bound of five given by the number of estimable cutsets. However, if
x M = {S2 , S3 , S6 }, then only K 1 (S1 ) and K 2 (S1 ) are estimable cutsets. Thus, in
this last case, the bound on the degree of estimability of S1 is two, and indeed
this is the minimum number of variables required to render it unobservable.
If in the same example all variables are measured, then all cutsets are redun-
dant, and the bound on the degree of estimability is six. However, the degree of
estimability of S1 is four. Elimination of S1 , S2 , S4 , and S6 is enough to render S1
unobservable. However, if x M = {S1 , S2 , S3 , S6 }, then only K 1 (S1 ) and K 2 (S1 )
are redundant cutsets. Now the maximum degree of estimability is three, which
is indeed the degree of estimability of S1 .
Fundamental Cutsets
Consider the canonical form of the incidence matrix, D 0 = [I D ∗ ]. This

matrix contains as many rows as units are in the system. Thus, it represents
as many linear independent balance equations as it is possible to write in the
system. In other words, it contains n U cutsets, where n U is the number of units.
We call these, fundamental cutsets.
Definition: The nonzero entries of each row of the canonical matrix represent
the fundamental cutsets of the system.
The relation between fundamental cutsets and spanning trees is that a funda-
mental cutset contains exactly one branch of a spanning tree. This property is
well known.
Determination of Cutsets
Because the rows of the canonical matrix represent all the linearly indepen-
dent cutsets of the system, then they are the base of a linear space of vectors
representing linear combinations of balance equations. Because a cutset is a
balance equation, then all cutsets are included in this space. However, cutsets
are represented by entries in each position of the matrix that are restricted to +1,
−1, or 0 such that they leave exactly two disjoint subgraphs after the elimination
of the variables of the cutset.
Thus, to find all the cutsets we resort to the following procedure:
(1) Create the cutset list by putting all fundamental cutsets in the list. Set the
counter k = 2.
(2) Create all linear combinations of k fundamental cutsets, restricting the
coefficients of such combinations to the numbers 1 and −1. This linear
combinations have to be such that absolute values of the entries of the
resulting vector are binary (0,1). If the result has any other existing cutset
as a subset, eliminate this result. Otherwise, include it in the list of cutsets.
(3) k = k + 1
(4) If k < n U perform step (2) (n U is the number of units).
To find all the cutsets that contain a specific variable, step (1) is modified to
include only the row that contains the variable of interest. A canonical form of
the incidence matrix can always be constructed such that the variable of interest
has only one nonzero entry in its corresponding column, that is, it is included
in the identity matrix. Other methods to obtain cutsets exist (Tsukiyama and
Verma, 1980; Fong and Buzacott, 1987).
Consider the system of Figure 5.9. Assume that one wants to determine
all the cutsets containing stream S3 . Then we start from Equation (5.45). The
following combinations of rows have to be explored: (1 ± 2), (1 ± 3), (1 ± 4),
(1 ± 2 ± 3), (1 ± 2 ± 4), (1 ± 3 ± 4).
The fundamental cutset containing S3 is the first row in the cutset list, that
is K 1 (S3 ) = {S1 , S2 , S3 }. If the fundamental cutset containing the variable of
interest is not in the first row, then a simple permutation of rows accomplishes
the objective. Of all the linear combinations of two rows the following are
successful.
r Row one plus row two (1 + 2). The new cutset is K 2 (S3 ) = {S1 , S3 , S4 }.
r Row one minus row three (1 − 3). The new cutset is K 3 (S3 ) = {S3 , S5 }.
r Row one minus row four (1 − 4). The new cutset is K 4 (S3 ) = {S3 , S2 , S6 }.
As one proceeds to perform all the linear combinations of three cutsets, one
obtain five combinations that render elements that are 1 or −1. They are: (1 +
2 − 3), (1 − 2 − 3), (1 + 2 − 4), (1 − 3 + 4), and (1 − 3 − 4). All of these,
K2
S2 S4
U2
S1 S6
U1 U4
K1
S5 K4
S3 U3
K5 K3
ENV
FIGURE 5.15. Cutsets containing S 3 .
except (1 + 2 − 4) contain cutset K 3 (S3 ) as a subset and, therefore, should

be disregarded. The combination (1 + 2 − 4) renders the the cutset K 5 (S3 ) =
{S3 , S4 , S6 }. These cutsets are illustrated in Figure 5.15.
Efficiency of Estimability of Variables
We are interested in the ability of the sensor network to provide the value of
a certain variable, even when a certain number of sensors fail, regardless of po-
sition. As it was described above, this is given by the degree of estimability and
is equal to the number of estimable cutsets whose intersection is the variable in
question. Assume now that no other cutset exists that could be made estimable,
such that its intersection with the existing estimable cutsets is only variable Si .
In such case, we say that the maximum efficiency of the installed measurements
has been obtained. Any other cutset made estimable will increase the number
of measurements, but it will not increase the estimability of the variable. We
now formally define this efficiency.
Definition: The cutset efficiency of estimability of a variable is given by the

quotient of the degree of estimability and the maximum degree of estimability.
For example, consider again the case of Figure 5.7. Consider the case where
S1 is unmeasured and x M = {S2 , S3 , S6 }. Then, only two estimable cutsets ex-
ist that contain variable S1 , namely, K 1 (S1 ) = {S1 , S6 }, K 2 (S1 ) = {S1 , S2 , S3 }
(Figure 5.16). The intersection of these two cutsets is only S1 ; therefore, its
degree of estimability is two. The maximum number of estimable cutsets one
can obtain with three measurements is two. Thus, the maximum degree of es-
timability is two and the efficiency is one. If one adds measurements in S4 and
S5 , the number of estimable cutsets is now five (see Figure 5.14). However,
the degree of estimability of S1 is three. In other words, the efficiency of the
estimability is now 3/5, smaller than before.
S2 S4
U2 K5
S1 S6
U1 U4
K2
S3 S5
U3
ENV
K1

FIGURE 5.16. Fundamental cutsets of the system in Figure 5.10.
Thus, when a variable has efficiency of estimability one, no unnecessary

estimable cutsets have been used. In other words, any additional cutset made
estimable by the addition of measurements may contribute to other goals, such
as the increase of precision, but will not affect the degree of estimability.
We are now in a position to discuss another type of efficiency. Consider the
case in which a certain unmeasured variable has a certain degree of estimability.
Furthermore, assume that cutsets exist whose intersection with all the estimable
cutsets that contain Si is {Si }. Thus, one can eliminate the measurements of one
estimable cutset and introduce measurements in another of these nonintersect-
ing cutsets to make it estimable, and, therefore, maintain the same degree of
estimability. This exchange of cutsets can continue until the minimum number
of measurements is used. This motivates the following definition.
Definition: The minimum estimation cardinality of degree k of a nonmea-
sured variable is the smallest number of measurements that are needed to obtain
a degree of estimability k for the variable.
The minimum cardinality can thus be obtained by finding a combination of k
cutsets whose intersection is variable Si only and whose union has the minimum
number of variables. Once the cutsets that include the variable in question are
known, this is a simple task. It is only required to enumerate the cutsets in
increasing order of cardinality, pick the first k sets and count all measurements
involved. Efficiency, defined as the quotient between the actual number of
measurements involved and the minimum needed, can also be introduced.
Consider again the case of Figure 5.9. Consider again the case where S1 is
unmeasured and x M = {S2 , S3 }. Then, only one estimable cutset exists, namely
K 2 = {S1 , S2 , S3 } and the degree of estimability of S1 is E 1 = 1. However, since
K 1 (S1 ) ∩ K 2 (S1 ) = S1 , then the set x M = {S6 } can achieve the same degree of
estimability, and therefore the minimum estimation cardinality of S1 is one
(Figure 5.17).
S2 S4
U2
S1 S6
U1 U4
K2
S3 S5
U3
ENV
K1

FIGURE 5.17. Estimable cutsets.
NONLINEAR SYSTEMS
Component Balances
In addition to material balances, component balances can be performed.

Assume there are p components. Then the system equations are now:
dwcj
= A f jc j = 1, . . . , p (5.48)
dt
C f jc = 0 j = 1, . . . , p (5.49)
X
P
f jc = f (5.50)
j=1
X
P
wcj = w (5.51)
j=1
Typically, concentrations of species and not component flows are measured;

thus, the relation between component flows and molar (or weight) fractions is
needed. This relation is expressed as follows:
f j = f ⊗ cj (5.52)
where ⊗ indicates the Hadamard product of the two vectors, that is a ⊗ b =

[a1 b1 a2 b2 . . . . . . an bn ].
Now that concentrations have been introduced, the set of Equations (5.49)
contains equations corresponding to splitters. These equations can be substi-
tuted by equations containing equalities of concentrations. The component
balance model becomes:
dw
= Af (5.53)
dt
Cf = 0 (5.54)
¡ w
¢
d w ⊗ cj ¡ f¢
= A f ⊗ cj j = 1, . . . , p (5.55)
dt
¡ f¢
C1 f ⊗ c j = 0 j = 1, . . . , p (5.56)
f
C2 c j = 0 j = 1, . . . , p (5.57)
Equation (5.57) represents now the equality of concentrations in splitters. In

addition, note that Equation (5.56) can also be written as follows:
¡ f¢ f ¡ f¢
C1 f ⊗ c j = (C1 ∗ f )c j = C1 ∗ c j f = 0 j = 1, . . . , p (5.58)
where ∗ represents the Hadamard product of matrix and a vector. This product
is defined as follows: A ∗ b = [a1 a2 . . . an ]b = [(a1 ⊗ b)(a2 ⊗ b) . . . (an ⊗ b)].
In the case of steady state the above system reduces to
Mf = 0 (5.59)
M1 [ f ⊗ c j ] = 0 j = 1, . . . , p (5.60)
C2 c j = 0 j = 1, . . . , p (5.61)
f
where the superscript f in c j has been dropped, as it is no longer needed to
distinguish concentrations of streams from concentrations in units. Also:
· ¸
A
M1 = (5.62)
C1
· ¸
A
M= (5.63)
C
To illustrate the form of these matrices we now resort to the systems in

Figure 5.18(a) and 5.18(b). These systems represent a flash unit followed by a
mixer for Figure 5.18(a) and a splitter for Figure 5.18(b).
The matrices M for both systems are:
S1 S2 S3 S4 S5 S1 S2 S3 S4 S5
· ¸ · ¸
1 −1 −1 1 −1 −1 (5.64)
Ma = Mb =
1 1 −1 1 −1 −1
S3 S3
S1 S1
S2 S2
S4 S5 S4 S5
(a) (b)
FIGURE 5.18. A flash and a mixer.
For the system in Figure 5.12(a) we have M1 = Ma , and matrix C2 does not
exist because there is no splitter. However, for system 5.12(b), matrices M1 and
C2 are:
S1 S2 S3 S4 S5
S1 S2 S3 S4 S5 · ¸
1 −1 (5.65)
M1 = [1 −1 −1 ] C2 =
1 −1
Energy Balances
In addition to material and component balances, energy balances can be

performed. The energy balance equations are:
d(w ⊗ h w )
= A[ f ⊗ h f ] (5.66)
dt
C1 [ f ⊗ h f ] = 0 (5.67)
C2 h = 0
f
(5.68)
where h f is the vector of stream specific enthalpies.

Typically, temperatures are measured. Thus, the relation between enthalpy
flows and temperature is needed. This relation is sometimes expressed assuming
constant and known specific heats as follows:
h f = cp ⊗ T (5.69)
Thus, the same expression can be used for the steady-state version of com-
ponent balances and energy balances. In particular for steady-state systems
we have:
Mf = 0 (5.70)
M1 [ f ⊗ (cp ⊗ T )] = 0 (5.71)
C2 T = 0 (5.72)
The above analysis does not include heat exchangers, because they obey
entirely different balances. In particular, one has to write two algebraic balance
equations for each heat exchanger and one algebraic heat balance for it. Thus,
assume
· ¸
MD
M= (5.73)
MI
where M I corresponds to heat exchangers and any other equipment where

heat is exchanged indirectly, and M D corresponds to the rest of the units in
the system where heat is directly exchanged, except splitters. We assume for
simplicity that the balance equations for units where heat is indirectly transferred
do not contain accumulation terms. Then the material and energy equations
corresponding to units where indirect heat transfer takes places are written as
follows:
MI f = 0 (5.74)
M1,I [ f ⊗ h ] = 0 f
(5.75)
where matrix M1,I contains one row per heat exchanger, that is, half the rows
of M I . We illustrate the structure of this matrix through the following small
example. Consider one heat exchanger (Figure 5.19).
In this case the corresponding matrices are:
S1 S2 S3 S4
· ¸
1 −1 (5.76)
MI =
1 −1
S1 S2 S3 S4
(5.77)
M1,I = [1 −1 −1 1]
which will render one energy balance.
S3
S1 S2
S4
FIGURE 5.19. A heat exchanger.

Observability Analysis of Bilinear Systems
We will consider for simplicity the steady state case. The dynamic case can be
treated the same way as in the case of linear systems, that is, by performing the
same manipulations on D instead of M. Several articles have studied this type
of systems in detail. A simple method, based on Gauss-Jordan factorization, is
presented first. A discussion of other methods follows.
Consider a canonical representation for the bilinear case. The analysis is
restricted to a single component. It also applies to the multicomponent case,
provided that the measurement of concentrations, when performed, is done on
all components of interest. The case of heat exchange will be analyzed as an
extension. Before proceeding with the analysis, a preliminary classification is
required. We classify the set of streams into the following sets:
S = S MU ∪ S UU ∪ S UM ∪ S MM (5.78)
where S MU is the set of streams where the concentration is measured and the
flow rate is unmeasured, S UU is the set of streams where no flow rate or con-
centration is measured, S UM is the set of streams where the concentration is
unmeasured and the flow rate is measured and S MM is the set of streams where
both the concentration and the flow rate are measured. We also considered as
measured all unmeasured streams that are connected to a splitter where at least
one concentration is measured.
We now partition matrices M, M1 and C2 into four parts following the same
classification.
£ ¤
M1 = M1MU M1UU M1UM M1MM (5.79)
¥ ¦
C2 = C2MU C2UU C2UM C2MM (5.80)
We now write the system of Equations (5.59) through (5.61) as follows:
M MU f MU + M UU f UU = −(M UM f UM + M MM f MM ) (5.81)
¡ MU MU ¢ MU ¡ ¢
M1 ∗ c f + M1UU ( f UU ⊗ cUU ) + M1UM ∗ f UM cUM
= −M1MM ( f MM ⊗ cMM ) (5.82)
¡ ¢
C2UM cUM + C2UU cUU = − C2MU cMU + C2MM cMM (5.83)
In these equations, the measured quantities are put in the right-hand side.
This will leave the left-hand side with unmeasured quantities. We rewrite these
equations in matrix form as follows:

 
  f MU
MU UU
¡ M ¢ M ¡ UM ¢  f UU 
 M MU ∗ C MU UU
∗ 

⊗ cUU 
UM
 1 M 1 M 1 f  f UU

C2UU  c 
UM
C2UM
cUU
 UM UM 
M f + M MM f MM
 
= −  M1MM ( f MM ⊗ cMM )  (5.84)
C2MU cMU + C2MM cMM
Thus, the system will be observable if the rank of the above matrix is
equal to the dimension of the vector. Otherwise, some variables will not be
observable. Gauss-Jordan factorization, much in the same way as it was de-
scribed above for linear systems, can be performed on the system matrix ob-
tained from Equation (5.84). The result needs further analysis. Indeed, within
these variables, there will be a set of unobservable variables, one of which
will be ( f UU ⊗ cUU )UO . However, some values of f UU and/or cUU in these
products may be observable, which will render the other also observable.
Therefore, an additional analysis has to be performed to obtain the final
results.
To illustrate the procedure consider a flash unit followed by a mixer as shown
in Figure 5.18(a). Measured and unmeasured flow rates and compositions of
one component are shown in Table 5.1.
Matrix M = M1 is now decomposed into the following matrices (matrix
M UM does not exist):
· S4 S5¸ · S3 ¸ · S1 S2¸
−1 1 −1 (5.85)
M MU
= M UU
= M MM
=
1 −1 1
TABLE 5.1. Measured Variables

in Figure 5.18.
Flow Rate Composition
S1 M M
S2 M M
S3
S4 M
S5 M
Thus, the observability analysis can be done by analyzing the following

matrix,
M MU
S4 S5 S4 S ˜3
-1 f2 - f1
1 -1 f2
-1 f1c1 - f2c2
c4 - c5 f2c2
M1MU ˜ cMU M UU = M1UU
(5.86)
In this matrix, S3⊗ indicates the column that corresponds to matrix M1UU , that
is, the column corresponding to ( f 3UU ⊗ c3UU ), whereas the last column is the
right-hand side of Equation (5.84). The canonical form is:
S3 S5 S3⊗ S4
 
1 f2 − f1
 
 (c2 − c4 ) 
 1 f2 
 (c4 − c5 )  (5.87)
 
 
 1 f 2 c2 − f 1 c1 
 
 (c2 − c5 ) 
1 f2
(c4 − c5 )
which gives the following correct results:
f3 = f1 − f2 (5.88)
(c2 − c4 )
f5 = f2 (5.89)
(c4 − c5 )
f 3 c3 = f 1 c1 − f 2 c2 (5.90)
(c2 − c5 )
f4 = f2 (5.91)
(c4 − c5 )
The analysis can be extended further by realizing that from Equations (5.90)
and (5.88) one can calculate c3 .
We now repeat the same analysis for the case of Figure 5.18(b) and the same
measurements as in the previous case. The submatrices of M, M1 and C2 are:
· S4 S5 ¸ · S3 ¸ · S1 S2¸
−1 1 −1 (5.92)
M MU = M UU = M MM =
−1 −1 1
S4 S5 S3 S1 S2
(5.93)
M1MU = [0 0] M1UU = [−1] M1MM = [1 −1]
S2
· ¸ · S4 S5 ¸
1 −1 (5.94)
C2MM = C2MU =
1 −1
Note that C2UU , M1UM and C2UM do not exist. Thus, the observability analysis
can be done by analyzing the following matrix:
M MU M UU
S4 S5 S3 S3˜
-1 f1 - f2
-1 -1 f2
f1c1 - f2c2 (5.95)
-1
c4 - c2
c4 - c5
M1UU
which has the following canonical form
S3 S3⊗ S4 S5
 
−1 f1 − f2
 −1 f 1 c1 − f 2 c2 
  (5.96)
 −1 −1 f2 
 
 c4 − c 2 
c4 − c 5
S4 S6
S1 S2 S3
S5 S7
FIGURE 5.20. Two heat exchangers in series.
From this matrix, we conclude that only f 3 and f 3 c3 are observable, whereas,
f 4 and f 5 are unobservable. A final analysis renders c3 observable. In addition,
the last two rows of the matrix indicate that the concentrations of S4 and S5 are
redundant.
Presence of Heat Exchangers
In the case where energy balances are being performed surrounding heat
exchangers, the procedure can be applied, but it requires a slight modification.
The equations equivalent to Equations (5.81) and (5.82) are:
¡ ¢
M IMU f MU + M IUU f UU = − M IUM f 2UM + M IMM f 2MM (5.97)
¡ MU ¢ ¡ UM ¢
M1,I ∗ h MU f MU + M1,IUU
( f UU ⊗ h UU + M1,I ∗ f UM h UM
= −M1,I
MM
( f MM ⊗ h MM ) (5.98)
We illustrate these equations through an example. Consider the two heat

exchangers in Figure 5.20.
Table 5.2 summarizes which variables are measured.
TABLE 5.2. Measured Variables

in Figure 5.20.
Flow Rate Temperature
S1 M M
S2 M
S3
S4 M M
S5 M
S6 M
S7 M
The material and energy balance equations are:
F1 = F2 = F3 (5.99)
F4 = F5 (5.100)
F6 = F7 (5.101)
F1 (h 2 − h 1 ) = F4 (h 4 − h 5 ) (5.102)
F2 (h 3 − h 2 ) = F6 (h 6 − h 7 ) (5.103)
Equations (5.102) and (5.103) can be rewritten in a form of products of the

flow rates and enthalpies of the same streams. Matrix M I and M1,I for this
system are:
 S1 S2 S3 S4 S5 S6 S7 
1 −1
 1 −1  (5.104)
M1 = 



1 −1
1 −1
· S1 S2 S3 S4 S5 S6 S7 ¸
1 −1 −1 1 (5.105)
M1,I =
1 −1 −1 1
Application of the above procedure renders all variables in the system

observable.
The theory of observability of bilinear systems started with the articles by
Vaclavek (1969), Vaclavek et al. (1976a, 1976b) and Vaclavek and Loucka
(1976) who first offered the definitions of redundancy and observability and
provided a graph-based necessary and sufficient conditions of unobservability.
All this work has some limitations because of the simplifying assumptions
made (Crowe, 1989). Romagnoli and Stephanopoulos (1980), proposed the first
classification method based on the solvability of the nodal balance equations
by using the output set assignment algorithm (Stadherr et al., 1974). Joris and
Kalitventzeff (1987) also proposed a structural rearrangement of the incidence
matrix, but their method may fail when loops occur leading to singularities
Later, Crowe et al. (1986) proposed the application of matrix projection, and
Kretsovalis and Mah (1987b, 1988a, 1988b) made use of graph theory. Finally,
Maquin et al. (1991) and Sánchez and Romagnoli (1996) make use of orthogonal
transformations, specifically QR decomposition.
FULL NONLINEAR SYSTEMS
A general nonlinear system of equations can be classified by using reduction

to canonical forms too. One way that also works for bilinear systems consists
of linearizing the system around the design or operating point. Once this is
performed, a Gauss-Jordan factorization or a Q-R decomposition can be done
easily. We concentrate only on steady state systems.
Consider the model f (x) = 0. Linearization around its expected operation
point yields
Jx ∼
=d (5.106)
where matrix J represents the Jacobian of f (x) around x0 and d is the cor-
responding constant. Matrix J may be partitioned in submatrixes JM and JU ,
which are related to the vector of measured variables (x M ) and unmeasured
parameters and state variables (xU ), respectively
· ¸
xM
[JM JU ] =d (5.107)
xU
We will illustrate the use of Q-R factorization in the following example. The
flash tank model of Figure 5.21 is taken from Van Winkle (1967).
F1 = F2 + F3 (5.108)
F1 yi1 = F2 yi2 + F3 yi3 (5.109)
X X X
yi1 = yi2 = yi3 = 1 (5.110)
yi3 = ηi yi2 Pi (sat)/P (5.111)
Assume the following structure for vector z:
x T = [F1 y11 y21 y31 F2 y12 y22 y32 F3 y13 y23 y33 P η1 η2 η3 ] (5.112)
F3, y3
T, P, h
F1, y1
F2, y2
FIGURE 5.21. A flash unit.

Full Nonlinear Systems 121
TABLE 5.3. Data for Figure 5.21.
Composition (Molar Fraction)

Saturation
Stream Pressures (mmHg) Feed Vapor Liquid
Flow (mol/hr) 100 49.50 50.50

Ethanol 5287 0.2 0.233 0.167
1-Propanol 2932 0.5 0.436 0.563
2-Propanol 4651 0.3 0.331 0.270
Temperature (◦ C) 137
Pressure (mmHg) 3600
Vaporization efficiency (%) 95
We now assume that the following variables are measured
x M = {F1 , F2 , y12 , F3 , y13 , P} (5.113)
Then, the corresponding matrices (JM , JU , and d) of the linearized model

around the nominal operation point are the following:
 
0.199 −0.167 −50.5 −0.233 −49.5
0.533 −0.595 −0.47 
 
0.267 −0.238 −0.297 
 
 
 
JM = 
 1 
 (5.114)
 −1.395 1 
 
 1 0.001
 
 0.001
0.001
 
100
 100 −50.5 −49.5 
 
 100 −50.5 −49.5
 
 1 1 1 
JU = 


 (5.115)
 1 1 
 1 1 
 
 −0.79 1 
−1.247 1
d T = [0.033 −0.0135 −0.0195 1. 1. 1. 0. 0. 0.] (5.116)

Using Q-R orthogonal factorizations, matrices Q 1 , Q 2 , R11 , R12 and 5 are:
 −1 0.0001 0.0001 0.003 −0.003 0.004 0.0004


−1 0.0001 0.003 −0.003 0.004 0.0004
 
 −1 0.003 −0.003 0.004 0.0004 
−0.01 −0.01 −0.01 −0.301 0.302 −0.387 −0.037 
 
Q1 =  −0.597 −0.060 −0.496 0.473  (5.117)
 0.671 −0.276 −0.557 
 
 −1
0.671 0.468 0.575
0.744 0.074 −0.555 0.365
 
−0.0082
−0.0082
 
−0.0082
 
 0.8164 
 
Q2 = 
 −0.412 
 (5.118)
 −0.404 
 
 0 
 
 0 
0
−100.005 −0.01 −0.01


−100.005 −0.01 49.497 50.4975
 
 −100.005 50.497 −0.0049 −0.005 
 −1.6761 −0.1491 −0.7487 
R11 =  1.4909 −0.4372 −0.325  (5.119)
 
 −1.0618 −0.2267 
−0.2788
0.2453
 
49.4975
 0.5947 −0.2288
 
 0.8945 −0.0229
R12 =
 −1.023
 (5.120)
 0.1708 

−0.2108 −0.1122
0
Nomenclature 123
 
1
 1 
 
 1 
 
 1 
 
 1 
5=


 (5.121)
 1 
 1 
 
 1 
 
 1 
1
After using the permutation matrix one discovers that y33 and η3 are unobserv-
−1
able, that is xU,2 = {y33 , η3 }. We now inspect the matrix R11 R12 to determine
which other variable is unobservable. Therefore,
 
0.9 −0.1246
 −0.9 0.1246 
 
−0.802 0.2467 
−1
R11 R12 =
 1

 (5.122)
 
 0.802 −0.2467
0.7561 0.4023
Thus, the only two nonzero rows correspond to variables y11 and η1 . These are
the two only observable variables. The rest are either measured or unobservable.
A word of caution must be included concerning the above procedures. The
Gauss-Jordan factorization method presented first for bilinear systems, re-
quires additional manipulations after the canonical form of the matrix is ob-
tained. And, in the case of component balances, the use of the normalization
equation for multicomponent systems was not included, and the analysis for
such systems was restricted to a special type of concentration measurements
where all components of interest are measured. Extensions to complete systems
have not been produced, although some discussion has occurred (Sanchez and
Romagnoli, 1996). For fully nonlinear systems the above presented technique of
linearization followed by observability analysis has been questioned (Sanchez
et al., 1992), and alternatives have been proposed (Ponzoni et al., 1997).
NOMENCLATURE
A: incidence matrix
AR: incidence matrix for redundant systems
B: incidence matrix
BR : incidence matrix for redundant systems
cic : concentration of component i in all streams

C: incidence matrix
CR: incidence matrix for redundant systems
D: incidence matrix for linear dynamic systems
Di : submatrices of D
Ei : estimability of variable i
f: vector of flow rate of streams
fR: vector of redundant flow rate of streams
f ic : flow of component c in stream i
Fi : flow rate of stream i
Gi : submatrices of the canonical matrix
h(x): measurement function
hf: enthalpy of material in streams
hw : enthalpy of material in units
I: identity matrix
J: Jacobian matrix
K j: jth cutset of a graph
K j (i): jth cutset containing variable i
M: incidence matrix in bilinear systems
Oi : degree of observability of variable i
P: projection matrix
P: pressure
Q: matrix in Q-R decomposition
Ri : degree of redundancy of variable i
Ri j : submatrices of Q-R decomposition
Si : stream i
S MU : set of streams where the concentration is measured and the flow
rate is unmeasured
S UU : set of streams where no flow rate or concentration is measured
S UM : set of streams where the concentration is unmeasured and the
flow rate is measured
S MM : set of streams where both the concentration and the flow rate are
measured
T: spanning tree
Ui : unit i
w: vector of hold-up of units
wic : hold-up of component c in unit i
x: vector of state variables
e
x: estimates of state variables x
xO : set of observable unmeasured variables
xUO : set of unobservable variables
xM : set of measured variables
xU : set of unmeasured variables
References 125
xR: set of measured and redundant variables

xNR : set of measured and nonredundant variables
yi j : molar fraction of component i in stream j
z: measurement values corresponding to x
Greek Letters
ϑ( p, Si ): set of all possible combinations of p measured variables, not

including Si
ϑ j ( p, Si ): jth element (combination) of ϑ( p, Si )
5: permutation matrix in Q-R decomposition
2( p): set of all possible combinations of p measurements
2 j ( p): jth element (combination) of 2( p)
REFERENCES
Ali Y. and S. Narasimhan. Sensor Network Design for Maximizing Reliability of Linear Processes.
AIChE J., 39, 5, 2237–2249 (1993).
Bagajewicz M. and M. Sánchez. Design and Upgrade of Nonredundant and Redundant Linear
Sensor Networks. AIChE J., 45, 9, 1927–1939 (1999).
Chen W. K. Applied Graph Theory. North Holland Publ., Amsterdam (1971).
Crowe C. M. Reconciliation of Process Flow Rates by Matrix Projection. II. The Nonlinear Case.
AIChE J., 32, 616–623 (1986).
Crowe C. M. Observability and Redundancy of Process Data for Steady State Reconciliation.
Chem. Eng. Sci., 44, 12, 2909–2917 (1989).
Matrix Projection. I. The Linear Case. AIChE J., 29, 818 (1983).
Even S. Graph Algorithms. Computer Science Press (1979).
Fong C. and J. Buzacott. An Algorithm for Symbolic Reliability Computation with Pathsets and
Cutsets. IEEE Trans. Rel., R-36, 34 (1987).
Griffith E. W. and K. S. P. Kumar. J. Math. Anal. Appl., 35, 135 (1971).
Harary F. Graph Theory. Addison-Wesley (1972).
Joris P. and B. Kalitventzeff. Process Measurement Analysis and Validation. XVIII Congress on
the Use of Computers in Chemical Engineering. CEF’87, 41–46 (1987).
Kalman R. E. New Approach to Linear Filtering and Prediction Problems. J. Basic Eng., ASME,
82D, 35 (1960).
Kou S. R., D. L. Elliot and T. J. Tarn. Inf. Contr., 22, 89 (1973).
Kretsovalis A. and R. S. H. Mah. Effect of Redundancy on Estimation Accuracy in Process Data
Reconciliation. Chem. Eng. Sci., 42, 2115 (1987a).
Kretsovalis A. and R. S. H. Mah. Observability and Redundancy Classification in Multicomponent
Process Networks. AIChE J., 33, 70–82 (1987b).
Kretsovalis A. and R. S. H. Mah. Observability and Redundancy Classification in Generalized

Process Networks. I. Theorems. Comp. & Chem. Eng., 12, 7, 671–687 (1988a).
Process Networks. II. Algorithms. Comp. & Chem. Eng., 12, 7, 689–703 (1988b).
Maquin D., G. Bloch, J. Ragot. Data Reconciliation of Measurements. Revue Diagnostic et Surete
de Fonctionnement., 1, 2, pp. 145–181 (1991).
Maquin D., M. Luong and J. Paris. Dependability and Analytical Redundancy. IFAC Symposium
on On-Line Fault Detection in the Chemical Process Industries. Newcastle, UK (1995).
Ponzoni I., M. C. Sánchez, N. B. Brignole. A New Partitioning Algorithm for Classification of
Variables in Process Plant Monitoring. AIChE’97 Annual Meeting, LA, USA (1997).
Chemical Plants. Chem. Eng. Sci., 35, 5, 1067–1081 (1980).
Sánchez M. C., A. J. Bandoni and J. A. Romagnoli. PLADAT: A Package for Process Variable
Classification and Plant Data Reconciliation. Comp. & Chem. Eng., S499–S506 (1992).
Sánchez M. and J. Romagnoli. Use of Orthogonal Transformations in Classification/Data Recon-
ciliation. Comp. & Chem. Eng., 20, 483–493 (1996).
Singh S. N. Int. J. Syst. Sci., 6, 723 (1975).
Stadherr M., W. Gifford and L. Scriven. Efficient Solution of Sparse Sets of Design Equations.
Chem. Eng. Sci., 29, 1025–1034 (1974).
Stanley G. M. and R. H. S. Mah. Observability and Redundancy in Process Data Estimation. CES,
36, 259–272 (1981).
Stewart G. W. Introduction to Matrix Computations. Academic Press, NY (1973).
Swartz C. L. E. Data Reconciliation for Generalized Flowsheet Applications. American Chemical
Society of National Meeting. Dallas, TX (1989).
Tsukiyama R. and M. Verma. An Algorithm to Enumerate All Cutsets of a Graph in Linear Time
per Cutset. J. Assoc. Comput. Mach., 27, 619 (1980).
Vaclaveck, V. Studies on System Engineering. III. Optimal Choice of the Balance Measurements
in Complicated Chemical Systems. CES, 24, 947–955 (1969).
Vaclaveck V. and M. Loucka. Selection of Measurements Necessary to Achieve Multicomponent
Mass Balances in Chemical Plants. CES, 31, 1199–1205 (1976).
Vaclaveck V., M. Kubicek and M. Loucka. Calculation of Material Balances for Chemical Engi-
neering Systems with Due Allowance for Measurement Errors. Theor. Found. Chem. Eng., 9,
242–245 (1976a).
Vaclaveck V., M. Kubicek and M. Loucka. Calculation of Material Balances for Chemical Engineer-
ing Systems with Due Allowance for Measurement Errors—Classification of Stream Parameters.
Theor. Found. Chem. Eng., 10, 256–260 (1976b).
Van Winkle M. Distillation. McGraw-Hill, New York (1967).
CHAPTER 6
Design and Upgrade of Nonredundant

and Redundant Sensor Networks
INTRODUCTION
The purpose of this chapter is to present methodologies for the grassroots

design and the upgrade of sensor networks, with the goal of achieving a certain
degree of estimability for specific variables. After reviewing sensor location
goals, the first part of the chapter discusses the design of linear networks.
Methods to design networks for bilinear systems are discussed afterward.
UPGRADE AND/OR DESIGN GOALS
When the value of all variables is of interest, then the objective becomes to
design a system for which all variables have a degree of estimability of at least
one, that is, all unmeasured variables are observable and all measured variables
are nonredundant. But one may not necessarily be interested in the value of
all variables. Some may remain unobservable. Thus, a design goal should be
estimability of order at least one in only the variables of interest. In the absence
of other goals, estimability of order one should be sufficient.
However, sensors may fail too often, and some reassurance that data for a
variable will remain observable when sensors fail may be desired. Requesting
estimability of higher order is one simple way of addressing this concern.
It might also seem that it is not necessary to distinguish whether a variable
needs to be measured or not, as long as it has a certain degree of estimability.
However, certain variables cannot be measured because there is no space to
install the instrument, there is lack of proper access for calibration, etc. Cost is
excluded, because this issue will be handled by the design procedure itself. In
some other cases, the measurement of some variables may be made compulsory.
127
128 NONREDUNDANT AND REDUNDANT SENSOR NETWORKS
Aside from the economic reasons, which will be handled by the design proce-
dure, reasons for such compulsory choice can be related to ease and/or small
frequency of maintenance. They even can be political, because many produc-
tion accounting personnel feel uneasy about not measuring certain variables
they consider of importance. These fears may not be unfounded, because the
degree of estimability does not take into account the frequency of failure of each
sensor, and, therefore, one very reliable instrument may fail less often than, for
example, a set of two instruments. These issues will be addressed in detail in
future chapters when they can be assessed mathematically using the concept of
instrument reliability.
DESIGN FOR ESTIMABILITY
If for each variable xi there is only one potential measuring device with
associated cost ci , then the total cost is given by:
X
C(q) = ci qi (6.1)
∀i
where q is a vector of binary variables defined by:

½
1 if xi is measured
qi = (6.2)
0 otherwise
Then, the design of the sensor network is an optimization problem that can
be written as follows (Bagajewicz and Sánchez, 1999):
X 
Min 

ci qi 

∀i∈M1


s.t. (6.3)

E k (q) ≥ E k∗ ∀k ∈ M E 




qi ∈ {0,1} ∀i ∈ M1
where E i is the degree of estimability of variable i, E k∗ is the minimum degree

of estimability imposed, M E is the set of variables with required degree of
estimability and M1 is the set of variables where measurements can be placed.
The inequality in the constraint of estimability is essential, because some-
times to achieve a certain degree of estimability in one variable, a larger degree
of estimability than the threshold may be required in others. If instead of in-
equalities, equalities are used, then the problem may be overly constrained and
become infeasible.
One needs to use the above model with caution. Even though a variable can
have a high degree of estimability, its value may be obtained through differences
of large numbers, and, therefore, the precision of such variables may not be
satisfactory. Such shortcomings come from the fact that estimability cannot
directly replace precision goals, nor can it fully replace reliability targets. Thus,
in the context of the above model, if these problems arise, one can increase the
degree of estimability required for variables that exhibit inadequate precision.
This will result in the activation of more cutsets and the consequent improvement
of precision and/or reliability.
Different types of problems arise depending on the degrees of estimability
required for the variables. All these aspects will be explored next. In addition,
if for some reason a measurement in a certain variable should be forbidden
(because of safety, space and other constraints), the corresponding binary vari-
able qi can be a priori set to zero. Similarly, if the variable is to be compulsory
measured, then the corresponding binary variable qi can be a priori set to one.
This is actually one case of instrumentation upgrade in which the existing in-
strumentation is not changed and only additions of new instrumentation are
considered.
Several approaches have been proposed to address this problem. In princi-
ple, the problem is a mixed integer nonlinear programming problem (MINLP),
but it has the added inconvenience that the functions E i (q) cannot be explic-
itly represented by using formal mathematical expressions in terms of q, so
one has to determine estimability by special ad hoc enumeration procedures.
The use of graph theory for special cases is first presented. However, in gen-
eral, this problem can be solved in general by using tree search procedures
(Meyer et al., 1994; Bagajewicz, 1997). Chmielewski et al. (1999) proved that
the problem can in principle be reduced to a traditional MINLP formulation
when precision constraints are used. The tree searching procedure proposed by
Bagajewicz (1997) is presented in this book. Finally, genetic algorithms are
presented in Chapter 16.
Minimal Networks
In this section we present networks that feature the minimum possible number
of sensors, hence the title. These networks arise from requesting estimability
of order one for all the variables. Indeed, consider a network where all the
variables have estimability of order one. The deletion of one measurement will
cause the loss of observability of at least one unmeasured variable, because it
will make that variable unobservable.
It was shown in Chapter 5 that the identity part of the canonical matrix
corresponds to a set of variables that form a spanning tree. It was also shown that
a system where all variables have degree of estimability E i = 1, corresponds
to a system where all unmeasured streams are given by a spanning tree. Thus,
the task of designing such a system consists of determining what spanning

tree has the largest cost. Once that is picked, the rest of the variables will be
nonredundant and will carry the lowest cost. In the special case of equal cost,
any spanning tree suffices.
Direct inspection, however, is costly. As pointed out in Chapter 5, a given
graph can have a large number of spanning trees, and, therefore, direct inspection
can be prohibitive, even when using powerful computers.
The number of sensors of minimal networks is therefore equal to the num-
ber of chords of a spanning tree. Thus, the number of sensors is obtained by
subtracting the number of branches of a spanning tree (n U − 1) from the total
number of streams.
n min = n S − n U + 1 (6.4)
The following algorithm, which avoids the enumeration of all spanning trees,
was presented by Madron (1992) and is based on the notion of Minimum
Spanning Tree (Even, 1979)
(1) Pick the edge with the largest cost. This is the first edge of the tree.
(2) Make a list of all the remaining streams (edges) that form a tree when added
to the list of chosen variables. Pick the one with the largest cost.
(3) Repeat step (2) until n U streams have been picked.
The proof that the above procedure guarantees optimality relies on picking the
stream with largest cost first (Even, 1979). This is one necessary and sufficient
condition for optimality. To explain this, let us consider a spanning tree that
does not contain the stream with largest cost. Because there is a spanning tree
of distance 1 (distance between spanning trees was defined in Chapter 5) that
contains this stream, it is clear that the spanning tree without the largest cost
stream is not optimal. The same argument can be made for any of the rest of
the elements of the spanning tree.
Example 6.1
To illustrate the procedure, we pick again the system of Figure 5.7. Consider
the following costs: c = [20 10 40 10 30 20]
(1) Stream S3 is picked first, because it has the largest cost.
(2) Streams S1 , S2 and S5 form a tree with S3 . S5 is selected because it has the
largest cost.
(3) Streams S1 , S2 , S4 and S6 form a tree with {S3 , S5 }. Streams S1 and S6 have
the higher costs, so both can be included in the spanning tree. Arbitrarily
stream S1 is selected.
S2 S4
U2
S1 S6
U1 U4
S3 S5
U3
FIGURE 6.1. Minimal network.
(4) Streams S2 and S4 form a tree with {S3 , S5 , S1 } (S6 does not form a tree
with {S3 , S5 , S1 } because it would result in a cycle). Both have equal cost,
thus stream S4 is selected.
The final spanning tree (unmeasured variables) is: xU = {S1 , S3 , S4 , S5 }. The
measured variables are x M = {S2 , S6 } and are shown in Figure 6.1.
Example 6.2
Consider the simplified process of hydrodealkylation (HDA) of toluene

(Douglas, 1988) shown in Figure 6.2.
Instrumentation costs are indicated between parentheses after the label of
the stream. By applying the above procedure the set of unmeasured vari-
ables obtained are xU = [S1 S5 S14 S12 S13 S8 S7 ], in which the streams are
listed in the order they have been selected. The measured variables are x M =
[S2 S3 S4 S6 S9 S10 S11 ]. The solution is shown in Figure 6.3.
The unmeasured variables form a spanning tree of the undirected graph that
corresponds to the more expensive sensors. The canonical form of the incidence
U4
S8 (250) S7 (230)
S9 (130)
U5 U3
S11(150) S6 (160)
U7
S10 (160) S12 (270) S14 (270) S4 (100)
U8
S13 (250) S5 (200)
U6 U2
S3 (180)
S2 (150) S1 (300)
U1
FIGURE 6.2. Simplified hydrodealkylation of toluene. Adapted from Douglas, 1988.

U4
S8 (250) S7 (230)
S9 (130)
U5 U3
S11(150) S6 (160)
U7
S10 (160) S12 (270) S14 (270) S4 (100)
U8
S13 (250) S5 (200)
U6 U2
S3 (180)
S2 (150) S1 (300)
U1
FIGURE 6.3. Minimal network for the HDA process.
matrix is the following:
S1 S5 S7 S8 S12 S13 S14 S4 S9 S10 S11 S2 S6 S3 

1 0 0 0 0 0 0 0 0 0 0 −1 0 −1
 
0 1 0 0 0 0 0 −1 0 0 0 1 0 1
 
 0 0 1 0 0 0 0 −1 0 0 0 0 −1 0
C0 = 
0 0 0 1 0 0 0 −1 −1 0 0 0 −1 0

 
0 0 0 0 1 0 0 −1 −1 1 1 0 −1 0
 
0 0 0 0 0 1 0 0 0 1 0 −1 0 0
0 0 0 0 0 0 1 0 −1 0 1 0 −1 0
Subminimal Networks
When only certain variables are of interest, at least estimability of order one
should be requested for these variables, whereas the rest of the variables can
remain unobservable. These types of networks are here called subminimal be-
cause less than the minimum number of sensors will be needed. Note, however,
that this may not be true if estimability of order larger than one is requested in
a subset of variables.
The term minimal is reserved to relate to networks in which the minimum
number of sensors is installed to achieve the lowest possible estimability of all
variables. Unfortunately, no equivalent methodology based on a concept similar
to the spanning tree was yet developed. Thus, mathematical programming seems
to be the only existing choice to this date. Bagajewicz (1997) proposed an

efficient tree-searching algorithm.
Tree-Searching Algorithm
We now concentrate again in the mathematical programming form of this

design problem given by Figure 6.3.
Because the constraints on estimability cannot be put in explicit analytic
form, a relaxation of the integer constraints is not possible, and, consequently,
lower bounds cannot be generated. Therefore, Branch and Bound procedures
cannot be implemented. Although implicit tree-type enumeration is impractical
for fairly large systems, in the case of this problem its special characteristics
allow a fairly efficient search. Consider the tree depicted in Figure 6.4. This
tree has some important properties:
(1) Node q = 0 is trivially infeasible.
(2) As measurements are added to form a branch the nodes are infeasible be-
cause the key variables are unobservable, or not enough estimability has
been achieved in them. Eventually, as measured variables are added a node
will become feasible. The node level at which feasibility is attained varies
from branch to branch.
(3) Cost increases from one level to the next. Therefore, the first feasible node
in each branch is the one with lowest cost in the tree its spans.
(4) Finally, retrofit problems, where instrumentation is simply added, can be
handled easily by freezing the corresponding elements of q. In future
chapters, the notion of reallocation will be included, and solution proce-
dures will be presented.
LEVEL 0 q = (0,0,0,...)
LEVEL 1
q = (1,0,0,..) q = (0,1,0,...)
q = (1,1,0,..) q = (1,0,1,...)
LEVEL 2
FIGURE 6.4. Tree of solutions.

Consider now the following tree enumeration scheme:

r Start with a root node containing no variables being measured (q = 0).
r Use the branch first rule, that is, construct the tree by developing each branch
(making one element of q active) until a feasible node is found. Then back
up one level and develop the next branch.
Constraint Checking
Assume that the vector q is given. The following procedure checks the fea-
sibility of constraint E k (q) ≥ E k∗ :
(1) Identify all the estimable cutsets for variable Sk . Assume the number of
these cutsets is s and the total number of variables involved in all these
cutsets is b, excluding Sk .
(2) Construct the (s × b) matrix Mk by including all the cutsets, one per row.
The entries of these rows are binary values (1, 0), 1 if a stream belongs to
the cutset and zero otherwise. Variable Sk is excluded.
(3) Determine the minimum number of measurements n k whose deletion makes
Sk an unobservable variable. This can be accomplished by determining the
minimum number of columns that add up to a column of ones, that is,
solving the following problem:

Xb

Min α pk 



p=1 

s.t. (6.5)
X 

b 

α pk m pk ≥ 1


p=1
where
α pk = binary variable (0, 1); α pk = 1 implies that measurement p

participates in the set of measurements which deletion makes
Sk unobservable; α pk = 0 otherwise.
m pk = pth column of the (s × b) estimable cutset matrix, Mk .
1 = (s × 1) vector of ones.
The problem is solved by inspecting combinations of increasing number of

measurements until the constraint of Equation (6.5) is satisfied.
(4) Determine the degree of estimability of variable Sk . If Sk is unmeasured, its
degree of estimability is n k , but if it is measured its degree of estimability
is E k = n k + 1.
FIGURE 6.5. Simplified ammonia plant network.
Example 6.3
We now illustrate this procedure to check feasibility using the simplified am-
monia network included in Figure 6.5. All cutsets for this network are extracted
from Ali and Narasimhan (1993) and presented in Table 6.1.
Consider that E 3 ≥ 1 and E 5 ≥ 1 are the constraints of Equation (6.3). As-
sume now that q = (1, 2, 4) is under feasibility analysis. The estimable cutsets
that contain S3 are K 3 and K 4 . Thus, matrix M3 is
S1 S2 S4 S5 S6 S7 S8
· ¸
1 0 0 0 0 0 0
M3 = (6.6)
0 1 0 0 0 0 0
Similarly, the estimable cutsets that contain S5 are K 10 and K 13 . Thus, matrix
M3 is
S1 S2 S3 S4 S6 S7 S8
· ¸
0 1 0 1 0 0 0
M5 = (6.7)
1 0 0 1 0 0 0
TABLE 6.1. All Cutsets of Ammonia Plant Network.
No. Streams No. Streams No. Streams
K1 S5 S6 S7 K7 S1 S6 S8 K 13 S1 S4 S5
K2 S1 S2 K8 S1 S4 S6 S7 K 14 S2 S4 S6 S7
K3 S1 S3 K9 S1 S5 S7 S8 K 15 S3 S4 S6 S7
K4 S2 S3 K 10 S2 S4 S5 K 16 S3 S4 S5
K5 S2 S6 S8 K 11 S4 S7 S8 K 17 S2 S5 S7 S8
K6 S3 S6 S8 K 12 S4 S5 S6 S8 K 18 S3 S5 S7 S8
It can be seen from matrix M3 that the deletion of only one measure-
ment at a time does not make S3 unobservable (because the other cutset re-
mains estimable) and, therefore, does not satisfy the constraint of Equation
(6.5).
However, the simultaneous deletion of measurements 1 and 2, that is α13 = 1,
α23 = 1, αu3 = 0 ∀u = 3, . . . , 7), satisfy the constraint of Equation (6.5), that
is,
X
b · ¸
1
α p3 m p3 = (6.8)
p=1
1
so the degree of estimability of variable S3 is two for this set of instru-

ments.
For variable S5 , the deletion of measurement S4 only allows the constraint
in Equation (6.5) to be satisfied and renders this variable unobservable, so the
degree of estimability of variable S5 is one. Consequently, the set of instruments,
represented by vector q, is feasible.
Although the procedure is combinatorial, the nature of the design problem
avoids a significant increase in the number of combinations. Because the design
follows a minimum cost criterion, the number of b measurements involved in
the s estimable cutsets for a variable Sk are low.
Example of a Subminimal Network Design
Consider again the hydrodealkylation process (Figure 6.2) with the instru-
mentation costs provided earlier. Assume that a degree of estimability of one
is first required for streams S1 , S8 and S9 . For this case the minimum cost
solution is 590, which corresponds to the installation of sensors in streams
[S4 S5 S6 S9 ], shown in Figure 6.6.
Matrices M1 , M8 and M9 are the following:
S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S13 S14 

M1 = [ 0 0 1 1 0 0 0 0 0 0 0 0 0 ]


S1 S2 S3 S4 S5 S6 S7 S9 S10 S11 S12 S13 S14 


M8 = [ 0 0 0 1 0 1 0 1 0 0 0 0 0]
(6.9)
M9 = ø
U4
S8 (250) S7 (230)
S9 (130)
U5 U3
S11(150) S6 (160)
U7
S10 (160) S12 (270) S14 (270) S4 (100)
U8
S13 (250) S5 (200)
U6 U2
S3 (180)
S2 (150) S1 (300)
U1

FIGURE 6.6. Subminimal solution for the HDA process.
It is easy to see that the constraint of Equation (6.5) is satisfied. In addition,

the reader can verify that the measured streams are not a subset of the set of
measured streams for the minimal network.
General Networks
When the required degrees of estimability of some variables in a network

are greater than one, the networks are called general. Estimability constraints
may be imposed on some or all variables. The general sensor network design
problem is stated by Equation (6.3.) The feasibility of constraint E k (q) ≥ E k∗
is checked by using the procedure described previously.
Example 6.4
If a larger degree of estimability is required for streams S1 , S8 and S9 of

the hydrodealkylation process, higher instrumentation cost will result. For ex-
ample, if the degree of estimability lower bounds are E 1∗ = 2, E 9∗ = 1 the
feasible set of instruments corresponding to the minimum cost solution is
x M = [S1 S5 S7 S8 S9 S11 S14 ], which has a cost of 1530 (Figure 6.7).
U4
S8 (250) S7 (230)
S9 (130)
U5 U3
S11(150) S6 (160)
U7
S10 (160) S12 (270) S14 (270) S4 (100)
U8
S13 (250) S5 (200)
U6 U2
S3 (180)
S2 (150) S1 (300)
U1

FIGURE 6.7. Subminimal solution for the HDA process.
The corresponding matrices of estimable cutsets for this example are:
S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S13 S14

· ¸
0 0 0 1 0 0 1 0 0 1 0 0 1  
M1 = 
0 0 0 1 0 1 0 1 0 1 0 0 1  






S1 S2 S3 S4 S5 S6 S7 S9 S10 S11 S12 S13 S14 

· ¸

0 0 0 0 0 0 1 1 0 0 0 0 0
M8 = (6.10)
1 0 0 0 1 0 0 0 0 1 0 0 1  






S1 S2 S3 S4 S5 S6 S7 S8 S10 S11 S12 S13 S14 
· ¸

0 0 0 0 0 0 1 1 0 0 0 0 0  

M9 = 
1 0 0 0 1 0 1 0 0 1 0 0 1
These matrices show that constraints for variables S1 and S8 are satisfied as
equalities. In contrast, variable S9 has a degree of estimability of E 9 = 2, which
is higher than its lower bound E 9∗ = 1.
Design for Estimability Efficiency 139
Other Approaches
Luong et al. (1994) presented a sensor network design procedure for a re-
stricted class of problems. They consider the case where observability of at
least degree one is required for variables that are of interest for control pur-
poses. In addition, they propose to add the requirement that a certain subset of
these variables has degree of redundancy k. Their solution strategy is based on
two steps. First, they propose to determine the subminimal network that will
guarantee observability of order one. In a second step, they propose to choose
supplementary variables to be measured to achieve the degree of redundancy of
the desired subset of variables of interest. The method is based on the exhaus-
tive investigation of cycles of the graph. However, cycles are directly related
to cutsets, because a cutset that contains a certain variable also contains one
variable from each cycle that includes this variable. Therefore, the procedure
is closely related to the above-presented algorithm based on cutsets. However,
an exhaustive enumeration of cycles is required.
Meyer et al. (1994) proposed a slightly more general strategy where an ex-
haustive enumeration of cycles is also proposed to achieve estimability of order
one. To reduce the search, they propose a branch and bound procedure that
appears to be effective to solve the problem. The method is suitable to design
subminimal networks and it can be generalized to other general estimable net-
works. Finally, Maquin et al. (1995) used a cost minimization problem where
the concept of cycles is used to build proper constraints. This method is a
particular case of Figure (6.3).
DESIGN FOR ESTIMABILITY EFFICIENCY
If instrumentation costs are not available, a good goal for mass sensor network
design is the selection of the minimum number of sensors that fulfill estimability
constraints for key variables. This kind of sensor network design problem may
be posed as follows:

Min Ns 


s.t.
Ek ≥ Ek ∗
∀k ∈ M E  (6.11)


qi ∈ {0,1} ∀i ∈ M1
where Ns is the number of sensors of the network.

The tree type enumeration strategy with stopping criteria proposed by
Bagajewicz (1997) and described above is applied to solve the MINLP
U5
S8 S4
S7
S6 S5
U1 U6 U4
S1 S3
U2 S2 U3

FIGURE 6.8. One estimable efficient solution.
problem. Although the stopping criterion avoids the enumeration to be ex-

haustive, this procedure is still not efficient for large-scale systems. The im-
plementation of numerical efficient strategies is an area of current active
research.
Example 6.5
Consider the sensor network design for the simplified process flowsheet of
ammonia production (Figure 6.5). Assume that the estimability constraints are
the following E 2∗ = 2, E 5∗ = 3. These constraints are satisfied when a minimum
number of five instruments are installed. Three alternative sets of instruments
fulfill the estimability requirements: [S1 S4 S5 S6 S7 ], [S2 S4 S5 S6 S7 ] and
[S3 S4 S5 S6 S7 ]. The first one is shown in Figure 6.8.
Consider now the hydrodealkylation process and assume various different
estimability constraints (Table 6.2). For each case, the lower estimability bounds
on streams, the minimum number of sensors and the solution set of instruments
are presented.
It can be seen from the results of cases 1 and 2 that a lower number of sensors
are required by decreasing the estimability bounds for the same sets of streams.
Obviously the tendency is that the minimum number of instruments increases
when greater requirements of estimability are imposed.
COMPULSORY MEASUREMENTS AND THE UPGRADE CASE
It is usually mandatory to install instruments on some streams to satisfy

control, balance accounting or safety requirements. This situation is taken into
Compulsory Measurements and the Upgrade Case 141
TABLE 6.2. HDA Process—Results for the Minimum Number

of Sensors Problem.
Case Streams Ei∗ Ns Solution(s)
S1 2
1 S8 3 7 S1 S5 S7 S8 S9 S11 S14
S9 1
S1 2
2 S8 2 5 S1 S5 S6 S8 S9
S9 2
S7 2
3 S13 3 9 S1 S3 S5 S6 S9 S10 S11 S12 S13
S11 2 S2 S3 S5 S6 S9 S10 S11 S12 S13
account by setting a priori the corresponding binary variable qi to one. Thus, the
design of minimum-cost sensor networks subject to estimability and location
constraints is stated as follows:
X 
Min ci qi 



∀i 

s.t.
E k (q) ≥ E k∗ ∀k ∈ M E (6.12)


qj = 1 ∀ j ∈ MU 



qt = {0,1} ∀t ∈ (M1 − MU )
where MU contains all variables that should be measured.

A similar problem arises when the currently installed set of instruments does
not fulfill the estimability requirements, so it is necessary to incorporate others.
The formulation of the optimization problem is the same. The solution involves
the same set of instruments, but the objective function value is lower, because
the cost of the already existing instruments is zero.
Example 6.6
Consider for example where two flowmeters are installed on streams [S1 S4 ]
of the ammonia process flowsheet (Figure 6.5). Assume again that the estima-
bility constraints are E 2∗ = 2, E 5∗ = 3. These requirements are not satisfied with
the initial set of instruments, so the location of new instruments is obtained by
solving the problem in Equation (6.12). Instrumentation costs are given by the
vector c = [0 300 300 0 220 280 250 250], where the already located sensors
have a zero cost. The solution indicates that constraints may be fulfilled by
incorporating sensors on streams [S5 S6 S7 ]. The optimal cost is 750.
TABLE 6.3. HDA Process—Results for the Upgrading Design Problem.
MU (Streams with Streams with New

Case Streams Ei∗ Cost Installed Sensors) Sensors
S1 2
1 S8 3 1080 S2 S11 S12 S3 S5 S6 S8 S9 S10
S9 1
S1 2
2 S8 3 1260 S2 S11 S3 S5 S7 S8 S9 S14
S9 1
S7 2
3 S13 3 1350 S2 S11 S3 S5 S6 S9 S10 S12 S13
S11 2
Examples of upgrading of the hydrodealkylation process (Figure 6.2) are

shown in Table 6.3.
Cases 1 and 2 have the same estimability constraints, but the number of
sensors already installed is different. The flow rate on stream S12 is considered
unmeasured for case 2. Even though this flowmeter has a cost of 270, the optimal
solution of case 2 is only 180 more expensive than the optimum value of case
1. The optimal solution of case 3 is the same as the one obtained for the design
of minimum number of sensors in the previous section (case 3 in Table 6.2).
SENSOR NETWORKS FOR BILINEAR SYSTEMS
The difficulty in designing sensor networks for bilinear system stems from
the complicated procedures needed to classify these systems. Extensions of the
concept of observable and redundant variables are straightforward, and the issue
of degree of estimability can be handled in the same way as for linear systems.
However, the extension of the connection between observability and spanning
trees and cutsets is not straightforward. As shown in Chapter 5, concentrations
or temperatures, can be sometimes used to make some flow rates observable.
Therefore, there is an interaction between these two types of variable that needs
to be taken into account to determine estimability. Thus, problem (6.3) has to be
solved by using the tree-searching procedure presented in this chapter, together
with an algorithm for bilinear system variable classification for node feasibility.
Unfortunately, no alternative method exists, especially for large systems where
the tree-searching algorithm may be time consuming. We now concentrate
on special types of networks, for which relatively simple design procedures
exist.
Linearly Estimable Sensor Networks
Consider first the case where flow rates are estimable based on flow measure-
ment only. We call these Linearly estimable sensor networks. In particular, if
the degree of estimability is one, then the flowmeters are located in the chords
of a spanning tree. We will now prove that any spanning tree solution for the
concentration sensors will render complete system observability.
Lemma 6.1: The minimum number of sensors needed to make all concentra-
tions estimable in a bilinear sensor network that is at the same time a linearly
estimable sensor network, corresponds to flowmeters located on a chord of a
spanning tree and concentration sensors located in the chord of another, not
necessarily different, spanning tree.
Proof: Under the conditions of linear estimability of order one, the flowme-
ters are positioned in the chords of a spanning tree. Then, the component balance
equations become linear. That is
¡ f¢ f f
M f ⊗ c j = (M ∗ f )c j = M f c j (6.13)
Consider now the canonical form of matrix M f :
G f = [I G ∗f ] (6.14)
If concentration sensors are located in the streams that correspond to G ∗f , all the
rest of the concentrations are observable. In other words, locating the sensor
measurements in the chords of any spanning tree renders all concentrations
observable. Q.E.D.
Thus, if cost is to be considered, the task at hand is to select the pair of
spanning trees that will render the system estimable at the minimum cost.
Because one spanning tree has no influence in the other, the choice is to select
the most expensive spanning trees for flow rates and concentration sensors
independently and locate the sensors in their chords.
Minimally Estimable Sensor Networks
We now consider the class of sensor networks for bilinear systems featuring
a minimum number of sensors and nonzero minimum degree of estimability
for all the variables. Linearly estimable sensor networks featuring a minimum
number of sensors are a subset of these networks.
Because the number of sensors remains the same, we can start from the
design based on spanning trees for both flow rates and concentration sensors.
Assume that one wants to exchange a flowmeter by a concentration sensor.

Eliminating one flowmeter and adding one concentration sensor in an appro-
priate position will lead to an observable system. The unmeasured chord of the
flow rate spanning tree will be estimated through the use of concentration mea-
surement. The converse is not true, that is, the elimination of a concentration
sensor and the addition of a flowmeter do not make the concentration sensor
observable.
Networks without Splitters or Heat Exchangers
In what follows we present a methodology to design bilinear systems with-

out splitters or heat exchangers. Later, extensions to cover these cases will be
presented. We now define the following:
Definition: A set of flowmeters and concentration sensors that makes all
variables in a bilinear system estimable is k linearly estimable deficient if the
addition of k flowmeters, but not less, makes the system linearly estimable.
f
Definition: Consider a set of measured flowrate variables x M , a set of con-
f
centration sensors y M and a stream Si with unobservable flowrate f i . A con-
centration sensor y j in stream S j that makes the flowrate of Si observable is
f f
a flowrate observability enabler of f i with respect to x M and y M , which are
called the reference sets. Vice versa, a flowmeter in stream Si is the concentra-
f f
tion observability enabler of y j with respect to the reference sets x M and y M .
First note that a flowmeter can have many enablers. In addition, two unob-
servable flow rates can have the same enabler. In other words, one concentration
sensor makes both observable. Finally, an unmeasured flow rate that is observ-
able by using only flow rate measurements does not have an enabler, because
it is already observable. Consider the example of a simple flash (Figure 6.9),
where only the inlet flow rate and the compositions of the feed and vapor are
measured
F3, y3
T, P, h
F1, y1
F2, y2

( ):Measured compositions.
FIGURE 6.9. A flash unit.
The flow rates F2 and F3 are unobservable if only total material balances
are used. However, if one measures the concentration y2 , these two flow rates
can be calculated by using the total material and component balance equations.
Thus, y3 is a flow rate observability enabler of F2 and F3 . Similarly, F2 and F3
are each concentration observability enablers of y2 .
Definition: Given a stream Si with unobservable flow rate and its concen-
tration enabler, the difference between the costs of the flowmeter and the con-
centration enabler is called flowmeter enabling cost.
Note that the enabling cost can be negative. Indeed, the flowmeter enabling
cost of the example of Figure 6.9 is given by the difference between the cost of
the flowmeter for F3 and the cost of a concentration measurement for y2 . This
value can be positive or negative.
Definition: Given a stream Si with observable flow rate, the flowmeter ex-
change cost of this sensor is given by the smallest cost of all the enabling costs
of this flow rate sensor. The concentration sensor exchange cost is just the
opposite.
In the example in Figure 6.10 there are two enablers of F3 , a measurement of
y2 or y5 . Therefore, the flowmeter exchange cost of F3 is given by the smaller
of these two enabling costs.
Consider now the following Lemmas.
Lemma 6.2: A necessary condition of optimality of a sensor network that
makes bilinear systems completely observable is that all unmeasured flow rates
have positive exchange cost.
Proof: The property is fairly obvious. If the flowmeter exchange cost of the
other streams with unmeasured flow rates is positive, the flowmeter can be
introduced and its enabler eliminated by lowering the overall cost, condition
that contradicts the assumption of optimality. Q.E.D.
F3, y3
T, P, h
F1, y1 F4, y4
F2, y2 F5, y5

FIGURE 6.10. A flash unit and a mixer.
y
Lemma 6.3: Let Tle be a spanning tree corresponding to the concentration
sensor network of a linearly estimable network. A necessary condition of op-
timality for a k linearly estimable deficient system is that the set of measured
y
concentrations sensors contains all chords of Tle .
y
Proof: The linearly estimable spanning tree Tle has the property that any
exchange of a measurement from chords to branches to obtain a new spanning
tree is such that the cost increases. Any other solution to the k-deficient es-
timable network corresponds to a spanning tree T y where the concentrations of
all chords are measured and the concentrations of some branches are measured.
y
First, consider that the measured branches of T y contain a chord of Tle . In such
y
case, one can exchange a chord of T y that is not a chord of Tle by this measured
branch of T y without affecting observability of the flow rates (both are mea-
sured) or changing the cost. Thus, we can assume without loss of generality that
y
the measured branches of T y , the enablers, do not contain chords of Tle . In other
y
words, the chords of Tle are either chords of T y , or its unmeasured branches.
y
Therefore, to transform one solution Tle to another T y , one has to exchange
y
the measured chords of Tle with unmeasured branches of T y , which implies an
increase in cost. To complete the proof, we will now show that the unmeasured
y
branches of Tle cannot be used to enable the unmeasured chords of T f . Assume
the measured and unmeasured flowrates allow the following decomposition
variables:
fO = G1 f M (6.15)
f UO,1 = G 2 f UO,2 + G 3 f M (6.16)
where f UO,1 and where f UO,2 correspond to the unobservable flows represented
by the chords and branches of T f , respectively. We can also write the same
component balance equations
f O ⊗ cO = G 1 f M ⊗ cM (6.17)
f UO,1 ⊗ cUO,1 = G 2 ( f UO,2 ⊗ cUO,2 ) + G 3 ( f M ⊗ c M ) (6.18)
where c M , c O , cUO,1 and cUO,2 correspond to the same streams as f M , f O ,

f UO,1 and f UO,2 . Substitution of Equations (6.16) into (6.18) followed by some
manipulations gives:
[G 2 ∗ cUO,1 − G 2 ∗ cUO,2 ] f UO,2 = (G 3 ∗ f M )c M − (G 3 ∗ f M )cUO,1 (6.19)
Thus, all elements of cUO,1 or cUO,2 need to be measured (we assume that
the columns of G 2 and G 3 are not zero). They are either chords or measured
branches. A similar argument can be made for those elements of c M for which
G 3 does not have a zero column. Thus, the exchange between chords and
y
unmeasured branches of Tle needs to take place between elements of c O and
the elements of c M that are not enablers. Thus, the enablers are not touched.
Q.E.D.
Lemma 6.4: A necessary condition of optimality of a system that is k linearly

estimable deficient is that the measured flow rates are chords of the linearly
f
estimable solution Tle of minimum cost.
Proof: Suppose that f M , f O , f UO,1 and f UO,2 correspond to the linearly es-
timable solution. Consider any other solution with measured chords not being
f
chords of Tle . Assume now that one exchanges certain number of measured
f
chords by the same number of unmeasured branches of Tle . Exchanging with
elements of f O will only increase cost and no changes are needed in the concen-
tration measurements. Consider now an exchange with elements of f UO,1 . This
is possible but will again lead to an increased cost and no needed changes in
the concentration measurements. From Equation (6.19), one can conclude that
an exchange with elements of f UO,2 does not require changes in concentration
measurements. Therefore, an exchange with elements of f UO,2 will again lead
to an increase in cost. Q.E.D.
Lemma 6.3 and the proof of Lemma 6.4 suggest the following constructive
step-by-step procedure for an optimal estimable sensor network.
r Step 1: Locate flowmeters and concentration sensors on the chords of the

most expensive flow and concentration spanning trees.
r Step 2: Determine the exchange cost of all flowmeters. If the cost is neg-
ative, then eliminate the flowmeter and place a concentration sensor in its
enabler.
Example 6.7
We illustrate the algorithm by using the ammonia plant network (Figure 6.5).
Assume the costs to be c f = [100 300 300 200 220 280 250 250] and c y =
[70 70 90 80 10 30 5 3]. The linearly estimable solution has the following
f y
chords: x M = [ f 1 f 4 f 7 ], x M = [y5 y7 y8 ] and are shown in Figure 6.11(a).
Its cost is 568 (550 from the flowmeters). The optimal solution is shown in
Figure 6.11(b) and its cost is 388. The large reduction in cost comes from
the elimination of the flowmeter in f 7 and the introduction of an enabler
in y1 .
U5 U5
S8 S4 S8 S4
S7 S7
S6 S5 S6 S5
U1 U6 U4 U1 U6 U4
S1 S3 S1 S3
U2 S2 U3 U2 S2 U3
(a) (b)

FIGURE 6.11. Bilinear ammonia plant sensor network. (a) Linearly estimable solution.
(b) Minimal solution.
Sensor Networks for Bilinear Systems with Splitters
Splitter units require special considerations in bilinear processes, because

a component balance around a splitter has to be replaced by an equality of
all compositions (temperatures) of all streams incident to the node. Thus, if
the compositions of some stream incident on a splitter are observable, then the
compositions of all other streams incident on the splitter are also observable.
Obviously, if splitter nodes are present in a process network, then the minimum
number of composition sensors that are required to observe compositions in all
streams is less than n S − n U + 1. Instead, the minimum number of composition
sensors is given by (Ali and Narasimhan, 1996):
X
n sp
n y = (n S − n U + 1) − m s,i + 2n sp (6.20)
i=1
where rs,i is the number of edges that are incident on splitter Ui and n sp is
the number of splitter nodes in a process. This equation can be obtained easily
by observing that for splitter Ui , there are (rs,i − 1) composition equality con-
straints instead of one component flow balance. This gives rs,i − 2 additional
equations relating the compositions, which may prompt a corresponding reduc-
tion in the number of composition sensors. The preceding equation is applicable
to process networks that do not contain splitters in series. However, such con-
figurations seldom, if at all, occur in practical processes. In addition, a single
splitter can perform the same function as splitters in series, so the exception is
a moot one.
FIGURE 6.12. Splitter with one branch of the spanning tree.
When splitters are present, the streams with unmeasured compositions do not
form a spanning tree, because less than n S − n U + 1 composition sensors are
required to observe all mass fractions. However, we make some observations
that enable us to take advantage of the notion of a spanning tree, and hence, use
the same concepts described above to obtain a sensor network design. These
observations were made by Ali and Narasimhan (1996). We assume that only
one splitter is present in the process; the extension to multiple splitters follows
naturally from the arguments presented below.
To maintain the minimum number of composition sensors when splitters are
present, some of the chords of T y should also be unmeasured. The following
three cases are possible:
r Case (a). Only one of the streams incident on the splitter is a branch of T y .
r Case (b). Exactly two of the streams incident on the splitter are branches of
T y.
r Case (c). More than two streams incident on the splitter are branches of T y .
For case (a), the fundamental cutset corresponding to the splitter branch, is
identical to the cutset that separates the splitter unit from the rest of the process.
This is shown in Figure 6.12. In these figures, the node marked S represents the
splitter unit. The solid line is the branch of T y , and the dotted lines are the chords.
Ordinarily, for a unit that is not a splitter, all the chords would need to be
measured. However, in the case of a splitter, only one chord is needed. Thus, a
composition sensor must be placed on only one of the splitter streams in this
fundamental cutset, while the composition of the remaining streams can be
unmeasured.
Case (b) is depicted in Figure 6.13. In this case both the fundamental cutsets,
K S1 and K S2 , contain the same chords c1 , c2 , . . . , cT not connected to the
splitter.
For all practical purposes the pairs (S1 , S3 ) and (S2 , S4 ) can be replaced by
just the branches S1 and S2 because knowledge of these two immediately gives
the other two. Moreover, by the same argument, one can even eliminate the
splitter and replace it by a single stream connecting U1 and U2 . This process is
U1
S1 S3
KS2 S KS1
c1
S4 c2
cT
S2
U2
FIGURE 6.13. Splitter connected to two branches of the spanning tree.
Thus, the measurement of the chords c1 , c2 , . . . , cT provide complete ob-

servability of all the branches and chords connected to the splitter. Incidentally,
note that an exchange of any chord ci connecting U1 and U2 with one of the
branches S1 or S2 leads to a spanning tree that falls into case (a).
We now show that the spanning tree corresponding to case (c) can be analyzed
in the same way as case (b). First, the chords connected to the splitter can be
eliminated, and later the splitter can be merged successively with U2 and U3 .
This is shown in Figure 6.15. The net result is the same situation as in case (b).
Therefore, the evaluation of a cost of a spanning tree containing splitters can
be performed by distinguishing two cases. If only one branch of the spanning
tree is connected to the splitter, then a concentration sensor should be located
in the cheapest chord. Otherwise, if more than one branch connects the splitter,
no chord should be measured.
Example 6.8
We return to the example of the graph corresponding to the ammonia network.

Consider, the same costs, but now assume that unit U5 is a splitter. Then, the
linearly estimable solution does not contain more than one concentration sensor
FIGURE 6.14. Splitter connected to two branches of the spanning tree.

FIGURE 6.15. Splitter relation with branches of the spanning tree.
around the splitter (Figure 6.16). Indeed, the concentration measurement in

stream S7 can be eliminated [case (b)].
Energy Sensor Networks
Two different graphs can be used to represent energy networks, so that sensor
network design can be performed. A mass flow graph is used to represent the
flow balances of the process and is derived from the flowsheet by replacing
each exchanger by two disjoint nodes representing the tube and shell sides of
the exchangers and adding an environment node to which all inputs and outputs
of the process are connected. The energy network remains the same as the
flowsheet. As an example, a heat exchanger network is shown in Figure 6.17(a).
All the algorithms developed for bilinear networks can now be applied to
obtain the optimal placement of flow and temperature sensors. The streams with
unmeasured flows will form a spanning tree of the flow rate graphs, whereas the
streams with unmeasured temperatures will form a spanning tree of the energy
network.
FIGURE 6.16. Linearly estimable solution. Splitter case.

S1 S2 S3
U1 U2
ENV
S4 S6 (b)
S1 S2 S3
U1 U2
ENV
S5 S7 S4 S6
(a) U1 U2
S5 S7
ENV
(c)
FIGURE 6.17. Energy networks decomposition (a) flowsheet and energy graph. (b) Hot
streams. (c) Cold streams.
Upgrade of Bilinear Sensor Networks
If one wants to upgrade a system to make it linearly estimable, all one needs
to do is to determine the spanning trees whose chords contain all the existing
measurements. Once this is performed, one can locate instruments in the rest
of the chords. Although this will render a minimum number of sensors, the
new instrumentation added might be more expensive than other solutions that
feature a larger number of added instruments but at a smaller cost. For such an
alternative, as well as for more general ones, the model given in Figure (6.3)
can be used to perform this upgrade.
We now concentrate on a series of rules developed by Ragot et al. (1992) for
the upgrade of bilinear sensor networks. This method is based on the following
observations of the overall balance and component balance equation around a
node. For component j and unit Uk , these equations have the following form:
¡ f¢
Dk f ⊗ c j = 0 (6.21)
Dk f = 0 (6.22)
where Dk is a row of the incidence matrix. For the time being we will ex-
clude heat exchangers and any other units in which indirect heat is transferred.
Consider the following situations where three variables are unobservable and
one-sensor placement actions are proposed to make these variables observable.
r Situation A: A couple of equations with two unobservable flow rates and one
unobservable composition.
—Placement action 1: Place a sensor in a flow rate. This leaves a system of
two equations with two unknowns, a concentration and a flow rate.
—Placement action 2: Place a sensor in a concentration. This leaves a system
of two equations with two unknowns, the two flow rates.
r Situation B: A couple of equations with two unobservable concentrations
and all observable or measured flowrates.
—Placement action: Place a sensor in one of the concentrations. This leaves
a system where a concentration can be calculated.
r Situation C: A couple of equations with three unobservable flow rates and
all observable or measured concentrations.
—Placement action: Place a sensor in a flow rate. This leaves a system of
two equations with two unknowns, the two flow rates.
Note that for a situation where two concentrations and one flow rate are
unobservable or when three concentrations are unobservable, no placement of
one sensor will make all the variables observable. For situations with more than
three unobservable variables, there is no one-sensor placement action that will
render all variables in that pair of equations observable.
We now concentrate on the upgrade algorithm suggested by Ragot et al.
(1992). The steps are:
r Step 1: Determine all observable variables. For this purpose, use the method
presented in Chapter 5. If all variables are observable, stop.
r Step 2: Search through all pairs of node balance equations and determine
which sensor location, if realized, makes the larger number of variables ob-
servable. If no sensor makes this happen, that is, when no pair of equations
falls in situations A, B or C above, then place a sensor at random in the
cheapest location, but avoid making a flow rate redundant. Go to step 1.
Ragot et al. (1992) used an observability analysis based on the construction of

two canonical forms of the incidence matrix, one for the flow rates and the other
for the concentrations. The former is the standard form presented in Chapter 5,
whereas the latter is based on the same procedure but using the measured
concentrations instead of the measured flow rates. For step 2, they propose the
application of the above placement actions for all pairs simultaneously. Finally,
because they realize that there are different options, they suggest that the choice
be driven by cost.
S12 S25
U4 U9
S11 S16
S10 S4 S17 S22
S1 S3 S15
U3 U1 U5 U8
S24
S5 S2 S13 S 14 S23
S9 S6 S27
U2 U6 U 10
S18
S7
S8 S19
U7
S21 S26
S20
U11 S28

FIGURE 6.18. Steam metering network.
Example 6.9
We now show the application of this algorithm [as shown by Ragot et al.
(1992)] to the steam system of the methanol production system (Serth and
Heenan, 1986) depicted in Figure 6.18. Assume that one desires to determine
flow rates and the composition of one possible contaminant in the steam. The
measured streams are:
f
x M = [ f 4 f 9 f 12 f 15 f 18 f 19 f 21 f 22 f 24 f 27 f 28 ] (6.23)
y
xM = [y1 y5 y7 y10 y14 y15 y17 y18 y19 y20 y22 y24 y25 ] (6.24)
Assume now that sensors can be positioned only in the following sets of
candidate variables z f and z y .
z f = [ f 1 f 6 f 7 f 10 f 14 f 16 f 17 f 23 f 26 ] (6.25)
z = [y4 y6 y16 y23 y27 ]
y
(6.26)
The classification algorithm renders flow rates f 23 and f 26 , as well as con-

centration y23 , observable. The flow rates f 23 and f 26 would be observable if
only flow-rate measurements were used in the observability analysis. We con-
sider them as “measured” to construct new canonical matrices corresponding
to flows and concentrations. The structure of these matrices is shown next.
I GU
UO
GRO,1 GNR,1
D= I GM GRO,2 GNR,2 (6.27)
UO
GR
Unmeasurable Variables Measurable Variables
Because the observable variables have been lumped with the measured vari-
ables, these two matrices contain “unmeasured” variables that are unobservable.
In addition, the unobservable columns are partitioned and further rearranged
so that measurable variables and unmeasurable variables are separated. These
correspond to the superscripts M and U , respectively. Finally, rows with unob-
servable variables are put first.
The two resulting canonical matrices are shown next:
S2 S3 S13 S8 S11S5 S20 S25S1S16 S10 S6 S7 S14 S17 S4 S9 S12 S15 S18 S19 S21 S22 S24 S27 S28 S23 S26
1 1 -1
1 1 1 -1 1 1 -1 1 1 -1 -1
1 1-1 -1 1 1 -1 -1
1 1 -1
1 1 -1 1 -1 1 1 -1 -1
Df = 1 -1 1 1 -1 -1
1 1 1 1 1 -1 1 1 1 -1 -1
1 1 1 -1 1 -1 1 1 -1 -1
1 1 1 1 1 1 -1 1 1 -1 -1
1 1 1 -1
1 -1 1 -1 -1
(6.28)
S2 S3 S9 S28 S13 S8 S12 S21S26 S11 S4 S16 S6 S27 S1 S5 S7 S10 S14 S15 S17 S18 S19 S20 S22 S24 S25 S23
1 1 -1
1 1 1-1 -1 -1 -1 -1
1 -1 1 -1 1- 1 -1 1 -1
1 -1 -1 1 -1
1 1 -1 -1
y
D = 1 1 -1 -1 1 -1
1 -1 -1 -1
1 1 1 -1
1 -1 -1 -1 1 1 1 -1 1
1 -1 1 -1 -1
-1 1 -1
(6.29)
Measurable unobservable, and unmeasurable unobservable flow rates are

given by f UO U
= { f 2 , f 3 , f 13 }, f UO
M
= { f 1 , f 5 , f 8 , f 11 , f 16 , f 20 , f 25 }, respec-
tively. Similarly, measurable unobservable, and unmeasurable unobservable
concentrations are given by yUO U
= { f 2 , f 3 , f 13 }, yUO
M
= { f 1 , f 5 , f 8 , f 11 , f 16 ,
f 20 , f 25 }, respectively.
The possibility of locating new instruments is explored next. The last two
rows of D f are candidates for situation B. Of these:
r The first row has variables f 19 , f 27 , f 23 and f 26 . Of these, y19 and y23 are
observable, y26 can be measured, but y27 cannot. Thus, the possible action is
to place a concentration sensor in S27 . This will make y26 observable.
r The second row has variables f 15 , f 18 , f 22 , f 24 and f 23 . Of these, all concen-
trations are observable.
We now search for equations with two unknown flow rates. These are can-
didates for situation A. These equations are given by line 4 and 9 of D f but,
unfortunately, two concentrations are unobservable.
Finally, we look for an equation with three unmeasured flow rates for which
all concentrations are measured. This is given by line 11 of D y . The measured
variables are y1 , y5 and y10 . One can locate a flowmeter in either S1 or S10 , but
not on S5 because it is unmeasurable.
Thus, situation A and C were applied and after action is taken to put a
flowmeter in either S1 or S10 and a concentration sensor in S27; an observability
analysis reveals that no further action is possible. The following variables remain
unobservable: f 2 , f 3 , f 13 , y2 , y3 , y8 , y9 , y13 , y28 . Unfortunately, these are all
unmeasurable and the system cannot be upgraded any further.
NOMENCLATURE
cic : concentration of component i in all streams

ci : chord i
C(.): cost function
ci : cost of sensor i
D: incidence matrix for linear dynamic systems
Dk : row incidence matrix
E k∗ : threshold of estimability
M: incidence matrix in bilinear systems
M1 : set of variables where sensors can be placed
ME : set of key variables for estimability
References 157
MK : matrix of all cutsets

m pk : pth column of MK
nk : minimum number of measurements whose deletion makes
stream Sk unobservable
Ns : minimum number of measurements in a network
n min : minimum number of stream that need to be measured to obtain
observability of all variables of interest
nS: number of streams or edges in a graph
n sp : number of splitter nodes
nU : number of units or nodes in a graph
ny: minimum number of composition sensors
q: binary vector indicating whether sensors are located (1) or not
(0) in variables
rs,i : number of edges incident on splitter Ui
Si : stream i
T: spanning tree; temperature
T f: spanning tree corresponding to flow rate measurements
in its chords
T y: spanning tree corresponding to concentration measurements
in its chords
y
Tle : spanning tree of a concentration sensor network for a linearly
estimable network
f
Tle : spanning tree corresponding to flow rate sensors of a linearly
estimable network
Ui : unit i
f
xM : set of measured flow rates
f
yM : set of streams with concentration measurement
z :
f
candidate sets for new measured flow rates
zf: candidate sets for new measured concentration
Greek Letter
α pk : binary variable indicating if sensor p is in the set that makes

Sk unobservable
REFERENCES
AIChE J., 39, 5, pp. 2237–2249 (1993).
Ali Y. and S. Narasimhan. Sensor Network Design for Maximizing Reliability of Bilinear Processes.
AIChE J., 42, 9, pp. 2563–2575 (1996).
Bagajewicz M. Design and Retrofit of Sensor Networks in Process Plants. AIChE J., 43, 9, pp. 2300–
2306 (1997).
Bagajewicz M. and M. Sánchez. Design and Upgrade of Non-Redundant Linear Sensor Networks.
AIChE J., 45, 9, pp. 1927–1939 (1999).
Chmielewski D., T. E. Palmer and V. Manousiouthakis. Cost Optimal Retrofit of Sensor Networks
with Loss Estimation Accuracy. AIChE Annual Meeting, Dallas (1999).
Douglas J. Conceptual Design of Chemical Process. McGraw Hill (1988).
Even S. Graph Algorithms. Computer Science Press (1979).
Luong M., D. Maquin, C. T. Huynh and J. Ragot. Observability, Redundancy, Reliability and
Integrated Design of Measurement Systems. 2nd IFAC Symposium on Intelligent Components
and Instrument Control Applications, Budapest (1994).
Maquin D., M. Luong and J. Paris. Dependability and Analytical Redundancy. IFAC Symposium
on On-Line Fault Detection in the Chemical Process Industries, Newcastle, UK (1995).
Meyer M. J. M. Le Lann, B. Koehret and M. Enjalbert. Optimal Selection of Sensor Location
on a Complex Plant Using a Graph Oriented Approach. Comp. & Chem. Eng., 18, Suppl.,
pp. S535–S540 (1994).
Ragot J., D. Maquin and G Bloch. Sensor Positioning for Processes Described by Bilinear Equa-
tions. Revue Diagnostic et Surete de Fonctionnement, 2, 2, pp. 115–132 (1992).
AIChE J., 32,733 (1986).
CHAPTER 7
Data Reconciliation
DATA RECONCILIATION
The purpose of this chapter is to present the basic concepts of data reconcil-
iation. Attention is given to those methods that provide analytical background
for future chapters. However, for the purpose of completeness, other methods
for performing data reconciliation are nonetheless briefly discussed. The reader
is also referred to three good books in the field (Madron, 1992; Narasimhan
and Jordache, 2000; Sánchez and Romagnoli, 2000).
BACKGROUND
Because measurements contain errors, they need to be adjusted. Data recon-

ciliation is a term used to refer to the problem of correcting plant measurements
(flow rates, temperatures, pressures, concentrations, etc.) so that they conform
to a certain chosen model. Usually, this model is given by a mathematical rela-
tion, typically a set of differential algebraic equations (DAE), as the one shown
in Chapter 5 [Equations (5.1) and (5.2)]:
d x1
= g1 (x1 , x2 ) (7.1)
dt
g2 (x1 , x2 ) = 0 (7.2)
The general data reconciliation problem is stated as follows: Given a set of

measurement values of a subset of state variables zM = (zM,1 , zM,2 ) it is desired
159
160 DATA RECONCILIATION
to obtain the best estimate of these measured state variables exM and as many
of the unmeasured variables e xU as possible.
When a statistical approach is taken, the problem of reconciliation consists
of minimizing the weighted square of the difference between the measurements
z M and the estimates e
xM at the N instances of time at which the measurements
were made, using as weight the variance of the measurements Q. That is, it
consists of solving the following optimization problem:

XN

Min T −1
xM (tk ) − z M,k ] Q [e
[e xM (tk ) − z M,k ]




k=0 

s.t.
(7.3)
dex1 

= g1 (ex1 ,e
x2 ) 

dt 



x1 ,e
g2 (e x2 ) = 0
This least square problem can be derived from Bayesian theory by using
the assumption that the distribution of errors is normal. The reader can re-
fer to the work done by Johnston and Kramer (1995) for a maximum likeli-
hood derivation of the steady-state linear reconciliation model. In addition,
Crowe (1996) showed that the same result can be derived by using infor-
mation theory. For the case where steady state is assumed, only one mea-
surement is used, usually an average of several measurements. In addition, in
some cases, accumulation terms, especially hold-up changes in tanks, are still
included.
The model represented by Equations (7.1) and (7.2) is based on first prin-
ciples. However, there are some dangerous exceptions that have permeated
into practice. Notoriously, the petroleum-refining industry is used to mea-
sure production in volumetric units (barrels). This has prompted practition-
ers to believe that a “volumetric balance” is the proper model. The pres-
sure has been so intense that some data reconciliation software vendors, like
DATACONTM (from Simulation Sciences, Brea, CA), agreed to introduce such
models.
Hardware Redundancy
We consider now the case where more than one set of measurements is
performed for each measured variable, that is, more than one instrument is
used. In such case, Equations (7.1) and (7.2) are not altered, and the objective
function includes the additional terms. Without loss of generality, assume that
each variable has m measurements. Let Q s be the variance of each set of
Background 161
measurements s. Then, Equation (7.3) becomes:
( )
X
N X
m 

Min x M 0 (ti ) − z M,s,k ]
[e T
Q −1 
− z M,s,k ] 
s [e
x M 0 (ti ) 



k=0 s=1

s.t.
(7.4)
dex1 

= g1 (e
x1 ,e
x2 ) 

dt 




x1 ,e
g2 (e x2 ) = 0
It will be shown next that this can also be accomplished by assuming that
one measurement per variable is made. The new “measurement” that needs to
be used is
X
m
z M,k = Q −1
s z M,s,k (7.5)
s=1
and its new variance should be

( )−1
X
m
Q= Q −1
s (7.6)
s=1
Indeed,
( )
X
m X
m
xM (ti ) − z M,s,k ]
[e T
Q −1
s [e
xM (ti ) − z M,s,k ] = [e
xM (ti )] T
Q −1
s e
xM (ti )
s=1 s=1
( )
X
m X
m
T −1
−2 [z M,s,k ] Q s e xM (ti ) + [z M,s,k ]T Q −1
s z M,s,k (7.7)
s=1 s=1
In turn, the proposed alternative objective function
−1
Z = {[e
xM (ti ) − z M,k ]T Q xM (ti ) − z M,k ]}
[e (7.8)
can be rewritten in terms of the definition of Q and z M,k as follows:
−1 −1
xM (ti ) − z M,k ]T Q
Z = [e xM (ti )]T Q e
xM (ti ) − z M,k ] = [e
[e xM (ti )
(7.9)
−1 −1
− 2[e
xM (ti )]T Q z M,k + [z M,k ]T Q z M,k
FIGURE 7.1. Ghost units.
Thus, according to the assumption, the first two terms of Equations (7.7) and
(7.9) are the same and the third term is a constant, both problems have objective
functions that differ only by a constant and, therefore, can be considered equiv-
alent. Alternatively, one might introduce a ghost unit (Figure 7.1) and work
within the framework of Equation (7.4) and a new equation representing just
the equality of the properties of both streams.
LINEAR DATA RECONCILIATION
In commercial applications data reconciliation is mostly performed involving

flow rates and tank inventories, that is, using material balances. The dynamic
model of a material balance in a process plant can be represented by the fol-
lowing differential algebraic system of equations (DAE):
dw
= Af (7.10)
dt
Cf = 0 (7.11)
where f are the flows of the different streams connecting the units and w are
the hold-ups of the respective units.
We saw in Chapter 5 that observable unmeasured variables can only be
calculated by using redundant and nonredundant measured variables through
Equation (5.20). Thus, the only measurements that are in conflict are the redun-
dant measurements, and, therefore, data reconciliation only makes sense for
redundant variables, which should satisfy (see Chapter 5):
GR x R = 0 (7.12)
When there is redundant measured holdup, matrix GR has the following

structure
· ¸
AR −B R
GR = (7.13)
CR 0
Therefore, Equations (7.10) and (7.11) are rewritten as follows:
dw R
BR = AR fR (7.14)
dt
CR f R = 0 (7.15)
In addition vector x M is now partitioned into redundant and nonredundant

components, that is, x M = [x R x N R ].
As for the variance matrix we have:
· ¸
QR 0
Q= (7.16)
0 Q NR
Finally, we partition Q R and Q N R into its flow and holdup parts, that is:
· ¸
Q R,F 0
QR =
0 Q R,W
· ¸ (7.17)
Q NR,F 0
QN R =
0 Q NR,W
It is a common practical (but not theoretical) assumption in steady-state data

reconciliation that measurements are independent, that is, there is no covariance
between different variables. In those cases, nonredundant variables have to be
accepted at measured face value. When covariances are not zero, adjustments
of nonredundant variables can be made. In practice, however, there is no good
information about the values of these variances and covariances so that the
usual assumption is that these matrices are diagonal. In the dynamic case,
because several measurements are available, it is said that they present temporal
redundancy. This redundancy allows certain fitting or filtering to be performed.
These variables are called self-redundant.
STEADY-STATE LINEAR DATA RECONCILIATION
We now analyze the solution of the linear data reconciliation problem. Con-
sider the case where no holdup change takes place or negligible holdup is
assumed. In such case only flow rates are estimated. Therefore, we write the
problem as follows

f M − f M+ ]T Q −1 [ e
Min [ e f M − f M+ ]

s.t. (7.18)


De f =0
where e fMe
f = [e f U ]T . Once D has been put into its canonical form one can
write:

Min [ e
f R − fR+ ]T Q −1 e + 
R,F [ f R − f R ]
s.t. (7.19)


CR efR = 0
This problem and its solution was the object of the seminal article of this field
written by Kuehn and Davidson (1961). It is a quadratic programming problem
with linear equality constraints. The solution can be obtained by determining
the stationary point of the Lagrangian function:
L = [e
f R − f R+ ]T Q −1 e + e
R,F [ f R − f R ] + λ CR f R
T
(7.20)
This stationary point is:

¥ ¡ ¢−1 ¦ +
e
f R = I − Q R,F CRT CR Q R,F CRT CR f R (7.21)
After reconciliation is performed, the following estimates, similar to Equa-

tion (5.20), are obtained:
e
f N R = f N+R (7.22)
f O = CRO e
e f R + CNRO e
f NR (7.23)
The steady-state model has posed some practical challenges:

(1) Plants are never truly at a steady state. Thus, as practitioners have resorted
to perform averages of several measurements to obtain one single number
per stream to use in the objective function of Equation (7.19), they are aver-
aging process variations as well. Therefore, these variations are considered
random errors.
(2) Tank holdup measurements cannot be used in steady-state data reconcil-
iation models. To add redundancy, holdup changes are modeled as pseu-
dostreams. This will be discussed later in this chapter.
(3) It has always been a problem to pick the values of the variance matrix Q.
Typically, covariances are ignored and a diagonal form of Q is used. In the
absence of hard data, these variances are chosen by using vendor informa-
tion or in some cases, the standard deviation of the signal.
(4) Because process variations are included in the measurements, the vari-
ance Q should also be adjusted to take them into account. This issue is also
discussed later in this chapter.
Precision of the Estimates
Once the steady-state data reconciliation problem is solved, it is desired to

know the precision Xof the estimates obtained. The covariance of linear combi-
nation of variables γik xk , is defined as follows:
k
Z Z Ã !Ã !
∞ ∞ X X
σi j =
e .... γik xk − z i γ js x j − z j p(x) d x1 d x2 . . . d xn
−∞ −∞ k s
(7.24)
X
where z i = γik x k and x k is the mean value of the measurements of variable
k
xk . After some bookkeeping one can write
XX Z ∞ Z ∞ XX
σi j =
e γik γ js (xk − x k )(xs − x s ) p(x) d xk d xs = γik γ js σi j
k s −∞ −∞ k s
(7.25)
Thus, in general, if z = 0x , then the variance of z is given by:
e = 0 Q0 T
Q (7.26)
and consequently, the variances of the flow rate estimates obtained from recon-
ciliation are given by:
¡ ¢
eR,F = Q R,F − Q R,F CRT CR Q R,F CRT −1 CR Q R,F
Q (7.27)
Note that:
r CR Q
eR = QeR C T = 0, that is, they are orthogonal.
R
rQeR is singular.
r Diag Q
eR < diag QR , that is, diag [QR C T (CR QR C T )−1 CR QR ] > 0. In other
R R
words, the standard deviations of the estimates are always smaller than those
of the measurements. This is the confirmation of a somehow intuitive knowl-
edge that the more measurements one uses to estimate a value the more precise
one gets.
Variance of Observable Quantities
Using Equation (7.26) one obtains:

· ¸
e
eO = [CRO CSRO ] QR [CRO CSRO ]T
Q (7.28)
Q NR
FIGURE 7.2. Flowsheet for Example 7.1.
Example 7.1
Consider first the system of Figure 7.2 and the corresponding set of measured
data given in Table 7.1.
The system matrix for this case is:
S1 S2 S3 S4 S5 S6 S7
 
1 −1 −1
 1 −1 
  (7.29)
C =
 1 −1 

 1 1 −1 
1 −1
The reader can verify that the canonical form of this matrix is:
S7 S2 S3 S4 S5 S1 S6
1 -1
1 1 -1
(7.30)
C= 1 -1
1 1 -1
1 -1
From this canonical form one can conclude that all measured streams are
redundant, that nonredundant variables are not present, and that S2 , S3 , S4 and S5
are unobservable variables, whereas S7 is observable. Once data reconciliation
is performed, the following results are obtained (Table 7.2):
TABLE 7.1. Measurements

for Figure 7.2.
Standard
Stream Measurement Deviation
S1 101.3 2.1
S6 102.7 1.9
TABLE 7.2. Data Reconciliation Results for Figure 7.2.
Reconciled/ Standard
Standard Estimated Deviation of
Stream Measurement Deviation Value Estimate
S1 101.3 2.1 100.61 1.985

S6 102.7 1.9 100.61 1.985
S7 — — 100.61 1.985
Assume now that a measurement is added in stream S5 . Then the new canon-
ical form of the system matrix is:
S7 S2 S3 S4 S1 S6 S5
1 -1
1 -1 1
C= (7.31)
1 -1
1 -1 1
1 -1
As a result of the addition of this single measurement, which is nonredun-

dant, all the unmeasured variables become observable. Reconciled values are
the same, because the redundant system has not changed. Finally, the previ-
ously observable variable S7 is not calculated by using the new nonredundant
measurement. Thus, its variance does not change.
Presence of Tanks
Tanks are very often an important part of a chemical plant. It has been,
therefore, of great interest to practitioners to be able to include all the transfers
of raw material between tanks, from tanks to processes or vice versa (called
transactions or custody transfer when done at battery limits) as part of data
reconciliation. However, tank holdup changes do not fit in the description of
a steady state. Thus, when steady state is assumed over a period of time, the
changes in holdup are usually divided by the time elapsed and considered
as a pseudostream leaving (or entering) the system (Figure 7.3). If the level is
measured, then the pseudostream is considered a measured stream. Real streams
entering or leaving the tank are considered separately.
Because transactions between tanks are reported, this conversion has been
very useful in adding substantial amount of redundant streams to refinery instal-
lations and has proven to be valuable to perform refinery-wide oil accounting
and oil loss assessment.
FIGURE 7.3. Use of pseudostreams in tanks.
Variance Estimation
If measurements are independent and are not correlated, the variance-

covariance matrix Q is diagonal. In such case, the elements of the diagonal
are the variances of the individual measurements. In the case of steady-state
data reconciliation, estimates of these values can be obtained by calculating the
variance of the distribution of data around the mean value, that is:
1X n
xi = xi,k (7.32)
n k=1
1 X n
cov (xi , x j ) = (xi,k − x i )(x j,k − x j ) (7.33)
n − 1 k=1
This is called the direct method. This is a correct procedure if the two afore-
mentioned assumptions, independence and steady state, hold. In addition, out-
liers must not be present.
Because the system is never at a true steady state, the above formulas of the di-
rect method incorporate the process variations, that is, the variance of the natural
process oscillations or changes as part of the measurement variance. In a simple
case like a ramp function, for example, the variance will be a composite of one
half the change in true value of the measured value during the sampling interval
and the true variance. To ameliorate this problem and to assess the existence
of variable interdependence (nondiagonal variance matrix), an indirect method
was proposed. This method was originally proposed by Almasy and Mah (1984)
who used the covariance matrix of the constraint residuals (r = C f + ) and min-
imized the sum of the squares of the off-diagonal elements. The method was
later slightly modified by Darouach et al. (1989) who proposed an iterative
procedure based on the solution of a nonlinear optimization resulting from
using a maximum likelihood estimator. Finally, Keller et al. (1992) extended
this work to nondiagonal covariance matrices. All these approaches still suffer
from the problem that they do not consider the possible presence of outliers.
Nonlinear Steady-State Data Reconciliation 169
Chen et al. (1997) proposed the application of an M-estimator that applies a

weight to each data based on its distance to the mean. They called this robust
indirect method. The discussion of the estimation using dynamic data is omitted.
Although all these methods for variance estimation have been tested using
computer simulations and have shown their power in these controlled experi-
ments, there is no assessment of how they behave in practice. In particular, data
reconciliation software notoriously lacks of any module to perform such esti-
mation, and there are no published results about the efficiency of these methods
in practice.
NONLINEAR STEADY-STATE DATA RECONCILIATION
Nonlinear data reconciliation refers to the solution of the problem outlined

by Equations (7.3) through (7.7). Its steady-state version is:

xM − z M ]T Q −1 [e
Min [e xM − z M ]
s.t. (7.34)

x) = 0
g(e
For process plants z M includes the typical state variables, flow rates, concen-
trations, temperatures and pressures, and the model g(e x) can include any type
of unit operations and equipment. In addition, e x usually contains parameters
that are not measured directly.
Several methods have been proposed to solve this problem, especially when
g(•) is bilinear. Because the solution is supposed to be close to the measure-
ment (unless gross errors are present), then one can linearize g(e x), perform a
classification of variables and extract the redundant system of equations. Once
this is done, one can solve successively updating the Jacobian in each iteration
until convergence is achieved. Other approaches using nonlinear programming,
such as the popular sequential quadratic programming codes (SQP), can also
be used.
Commercially available software performs nonlinear steady-state data rec-
onciliation to a good extent. For example, DATACON (version 3.0) offered by
Simulation Sciences, Brea, CA, is capable of performing material, energy and
component balances, all of them simultaneously, and it reconciles temperature,
pressure, concentration and flow rates. In addition, it performs certain param-
eter estimations, such as heat transfer coefficients of heat exchangers. In the
absence of systematic errors and leaks in the system, there is no reason why
these models cannot be made as sophisticated and complex as the optimization
techniques used to solve permit. However, software vendors are reluctant to in-
troduce such models. One of the reasons is a certain conviction that undetected
gross errors may still be largely amplified by the reconciliation, especially
when nonlinearities correspond to very nonideal systems. Bagajewicz and

Mullick (1995) discussed this issue in more detail.
DYNAMIC DATA RECONCILIATION
Early work in dynamic data reconciliation is rooted in the problem of process

state estimation using the concept of filtering. Lately, the problem has been
solved by using the concept of model-based data smoothing.
Consider the three types of state estimation problems that are illustrated in
Figure 7.4. Assume an estimation of the state of the system is desired at time t.
When only measurement values prior to the time of prediction t are used,
including the measurement at time t, the estimation is called filtering. When
time t is not included, the estimation is called prediction, and finally, when
data for times larger than t are used, the estimation process is called smoothing.
Finally, when discrete measurements are used, the estimators are called discrete
estimators.
The design of sensor networks for the specific purpose of using dynamic
data reconciliation is an area on incipient research, and there is very little work
performed, if any. Thus, we will limit the scope of this section to cite the different
work on dynamic data reconciliation performed.
For the case of linear systems one-step integration-reconciliation procedures,
many of them rooted on Kalman filtering (Kalman, 1960), have been pro-
posed. Stanley and Mah (1977) showed how Kalman filtering could be adapted
to take advantage of spatial and temporal redundancy in a quasi-steady-state
condition. Darouach and Zasadzinski (1991) proposed a backward difference
approximation and a recursive technique to solve the constrained least square
FIGURE 7.4. State estimation problems. Adapted from Gelb, 1974.

Nomenclature 171
optimization problem. Rollins and Devanathan (1993) improved on the esti-

mation accuracy by using a maximum likelihood function and proposing two
estimators that are later averaged.
The above-summarized line of work on local estimators is useful for control
and monitoring purposes, and in the absence of gross errors (biases and leaks),
they all perform rather well. However, the detection of gross errors shares some
of the difficulties of tests based on steady-state models which are reviewed in
Chapter 15.
Other methods have been presented for dynamic data reconciliation.
Ramamurthi et al. (1993) proposed a technique based on a successively lin-
earized horizon, Liebman et al. (1992) used orthogonal collocation and
Albuquerque and Biegler (1995) on a discretization of the system of DAE
using Runge-Kutta methods.
Work has also been done by using neural networks. Terry and Himmelblau
(1993) proposed their use in data rectification, a concept they claimed to be more
general than data reconciliation. A comparison with traditional methods was
performed later by Karjala and Himmelblau (1994). Later, Himmelblau (1994)
and Karjala and Himmelblau (1996) showed more examples of how data can
be rectified without the need of models. Later, Du et al. (1997) proposed the
use of auto-associative neural network for steady-state systems. Finally Reddy
and Mavrovouniotis (1998) proposed the use of input-training neural network
for gross error detection and size estimation.
Setting aside the computational volume challenges, these methods, as any
one-step procedures, will at best lessen the fluctuations of measurement data,
but rarely produce a smooth profile. For plant day-to-day management and
economics, such fluctuating patterns are not satisfactory. Indeed, a fluctuating,
rather than a smooth reconciled pattern does not provide a simple description
of data so that, for example, daily balances cannot be performed without further
processing. Smooth stream profiles, based on a few parameters, are amenable
for integration and thus easy to use and inexpensive to store. To ameliorate this
deficiency of the aforementioned methods, Bagajewicz and Jiang (1997) pro-
posed a method based on the formal integration of the system [Equation (7.10)],
a polynomial representation and a reconciliation using the polynomial param-
eters as unknowns.
NOMENCLATURE
A: incidence matrix
AR: incidence matrix for redundant systems
B: incidence matrix
BR : incidence matrix for redundant systems
C: incidence matrix
CR : incidence matrix for redundant systems
ef: estimates of flow rates
ef +M : measured values of redundant flow rates
ef o: estimates of observable unmeasured flow rates
ef R: estimates of redundant flow rates
ef NR : estimates of nonredundant flow rates
Q: variance matrix of measurements
Qs : variance matrix of measurements of set s
Q R: variance matrix of redundant measurements
eR :
Q variance matrix of estimates
Q R,F : variance matrix of redundant flow rate measurements
Q R,W : variance matrix of redundant holdup measurements
Si : stream i
Ui : unit i
w: vector of holdup of units
wR : vector of redundant holdup of units
e
z M,k : set of measurements at time k
Greek Letters
σ̂ : precision of estimates
σi j : covariance of measurements of variables xi and x j
REFERENCES
Albuquerque J. S. and L. T. Biegler. Decomposition Algorithms for On-line Estimation with Non-
linear DAE Models. Comp. & Chem. Eng, 19, pp. 1031 (1995).
Almasy G. A. and R. S. H. Mah. Estimation of Measurement Error Variances from Process Data.
Ind. Eng. Chem. Process Des. Dev., 23, pp. 779 (1984).
Bagajewicz M. and Q. Jiang. An Integral Approach to Dynamic Data Reconciliation. AIChE J.,
43, pp. 2546 (1997).
Bagajewicz M. and S. Mullick. Reconciliation of Plant Data. Applications and Future Trends.
AIChE Spring Meeting and First International Plant Operations and Design Conference, Houston,
Texas, March (1995).
Chen J., A. Bandoni and J. A. Romagnoli. Robust Estimation of Measurement Error Vari-
ance/Covariance from Process Sampling Data. Comp. Chem. Eng., 21, 6, pp. 593–600 (1997).
Crowe C. M. Formulation of Linear Data Reconciliation using Information Theory. Comp. Chem.
Eng., 51, 12, pp. 3359–3366 (1996).
References 173
Darouach M., R. Ragot, M. Zasadzinski and G. Krzakala. Maximum Likelihood Estimator of

Measurement Error Variances in Data Reconciliation. IFAC. AIPAC Symp. 2, pp. 135–139
(1989).
Darouach M. and M. Zasadzinski. Data Reconciliation in Generalized Linear Dynamic Systems.
AIChE J., 37, 2, pp. 193 (1991).
Du Y., D. Hodouin and J. Thibault. Use of a Novel Auto Associative Neural Network for Nonlinear
Steady-State Data Reconciliation. AIChE J., 43, 7, pp. 1785 (1997).
Gelb A., editor. Applied Optimal Estimation. The M.I.T. Press, Cambridge, Massachusetts (1974).
Himmelblau D. M. Rectification of Data in a Dynamic Process Using Artificial Neural Networks.
Proceedings of the Process Systems Engineering (PSE) International Symposium, Kyongju,
Korea (1994).
Johnston L. P. M. and M. A. Kramer. Maximum Likelihood Data Rectification. Steady state Systems.
AIChE J., 41, 11 (1995).
Kalman R. E. New Approach to Linear Filtering and Prediction Problems. J. Basic Eng., ASME,
82D, 35 (1960).
Karjala T. W. and D. M. Himmelblau. Dynamic Data Rectification by Recurrent Neural Networks
vs. Traditional Methods. AIChE J., 40, 11, pp. 1865 (1994).
Karjala T. W. and D. M. Himmelblau. Dynamic Rectification of Data via Recurrent Neural Nets
and the Extended Kalman Filter. AIChE J., 42, pp. 2225 (1996).
Keller J. Y, M. Zasadzinski and M. Darouach. Analytical Estimator of Measurement Error Vari-
ances in Data Reconciliation. Comp. & Chem. Eng., 16, pp. 185 (1992).
Kuehn D. R. and H. Davidson. Computer Control. II. Mathematics of Control. Chem. Eng. Prog.,
57, 44 (1961).
Liebman M. J., T. F. Edgar and L. S. Lasdon. Efficient Data Reconciliation and Estimation for
Dynamic Process Using Nonlinear Programming Techniques. Chem. Eng. Sci., Vol. 16, No.
10/11, pp. 963 (1992).
Narasimhan S. and C. Jordache. Data Reconciliation & Gross Error Detection. Gulf Publishing
Company, Houston, Texas (2000).
Ramamurthi Y., P. B. Situ and B. W. Bequette. Control-Relevant Dynamic Data Reconciliation and
Parameter Estimation. Comp. & Chem. Eng., 17, 1, pp. 41–59 (1993).
Reddy V. N. and M. L. Mavrovouniotis. An Input-Training Neural Network Approach for Gross
Error Detection and Sensor Replacement. Inst. of Chem. Eng. Trans IchemE, Vol. 76, part A,
May (1998).
Rollins D. K. and S. Devanathan. Unbiased Estimation in Dynamic Data Reconciliation. AIChE
J., 39, 8, pp. 1330 (1993).
Sánchez M. and J. Romagnoli. Data Processing and Reconciliation for Chemical Process Opera-
tions. Academic Press, San Diego, California (2000).
Stanley G. M. and R. H. S. Mah. Estimation of Flows and Temperatures in Process Networks.
AIChE J., 23, 5, pp. 642 (1977).
Terry P. A. and D. M. Himmelblau. Data Rectification and Gross Error Detection in a Steady state
Process via Artificial Neural Networks. Ind. Eng. Chem. Res., 32, pp. 3020–3028 (1993).
CHAPTER 8
Design of Precise Sensor Networks
INTRODUCTION
In Chapter 6, the design of sensors networks to guarantee estimability was

covered. In this chapter, a methodology for the grassroots design of a sensor
network in a process plant with the goal of achieving certain degree of precision
in key variables is presented.
COST-OPTIMAL DESIGN
The cost-optimal design of a precise sensor network is obtained by solving

the following problem (Bagajewicz, 1997):
X 
Min ci qi 



i∈M1 
s.t. (8.1)
σ j (q) ≤ σ j∗ ∀ j ∈ M P 




qi ∈ {0, 1} ∀i ∈ M1
In the formulation of the objective function, it is assumed that there is only one
potential measuring device with associated cost ci for each variable. Further-
more σ j (q) represents the variance of the estimated value of variable x j obtained
after data reconciliation using Equation (7.27) or (7.28). This should be kept
lower than the threshold value σ j∗ for each variable or parameter in the set MP .
In Chapter 5, the degree of observability of unmeasured variables and the
degree of redundancy of measured variables were introduced. These two con-
cepts were merged into the concept of degree of estimability. In Chapter 6, these
175
176 DESIGN OF PRECISE SENSOR NETWORKS
properties were used as goals to design and upgrade sensor networks. Estima-
bility can of course be added as a separate constraint, as in the model below.
X 
Min ci qi 



i∈M1 

s.t. 

∗ (8.2)
σ j (q) ≤ σ j ∀ j ∈ M P


E j (q) ≥ E j ∀ j ∈ M E 
∗ 




qi ∈ {0, 1} ∀i ∈ M1
To obtain a smaller value of σ j∗ , it may be required to increase the estimabil-

ity of a variable. Therefore, when both constraints, precision and estimability
are included, very rarely both will be binding, that is, at their limiting values
simultaneously. However, estimability constraints can be a way of requesting a
level of reliability that cannot be quantified otherwise.
These models can be solved by using the tree enumeration procedure
described in Chapter 6. Recall that the constraint on precision (σ j (q) ≤ σ j∗ )
cannot be explicitly written in an analytical form because the matrices involved
change in dimension with the choices for q. Thus, this tree enumeration proce-
dure requires the evaluation of the constraints at each node. In a recent work,
Chmielewski et al. (1999) showed that the problem can be in principle re-
duced to a traditional MINLP formulation. Because details of this work are un-
published as this time, they are omitted.
MULTIPLE INSTRUMENTS AND HARDWARE REDUNDANCY
If more than one device is being considered as a potential candidate to be

used in each variable measurement, the objective function requires the use of
additional binary variables and additional constraints. Indeed, let n im be the
number of different alternative candidates of measurement devices and let the
cost of each of these candidates be given by ci,k (k = 1, . . . , n im ). Finally for
each variable xi , introduce binary variables pi,k (k = 1, . . . , n im ) to determine
which candidate will be used, that is,
½
1 if device k is used to measure variable i
pi,k = (8.3)
0 otherwise
Then, the total cost is now a function of p and is given by:
m
XX
ni
C( p) = ci,k pi,k (8.4)
∀i k=1
Multiple Instruments and Hardware Redundancy 177
Therefore, the optimization problem is written as:

m 
XX
ni


Min ci,k pi,k 



∀i k=1 



s.t. 


σ j ( p) < σ j∗ ∀ j ∈ MP (8.5)


Xn im 



p j,k ≤ m j ∀ j ∈ M1 



k=1 

m
pi,k ∈ {0, 1} ∀ i ∈ M1 , ∀k = 1, n i
where the constraint

m
X
ni
p j,k ≤ m j (8.6)
k=1
guarantees that at the most, m j devices are assigned to each variable.

When m j = 1, constraint Equation (8.6) restricts the model by not allowing
more than one measurement per variable, that is, only systems with spatial
redundancy and no hardware redundancy are considered. The standard deviation
σ j ( p) can be easily obtained for m j > 1 through slight modifications of the
variance estimation Equation (7.27). In Chapter 7, two ways of performing
data reconciliation, by appropriately lumping the measurements into one and
reformulating its variance, or by adding a ghost stream, were presented. Variance
estimation formulas were also included.
Example 8.1
Consider the process flow diagram of Figure 8.1. Flow rates are given by:
x = (150.1, 52.3, 97.8, 97.8). Assume that for each rate, flowmeters of preci-
sion 3%, 2% and 1% are available at costs 800, 1500 and 2500, respectively,
regardless of size. Precision is only required for variables f 1 and f 4 , that is,
M P = {S1 , S4 }, with σ1∗ = 1.5% and σ4∗ = 2.0%.
FIGURE 8.1. Example 8.1.

TABLE 8.1. Solutions of the Precision

Constrained Problem.
Solution S1 S2 S3 S4
A — 2% 2% —
B — 2% — 2%
Two solutions are obtained featuring a cost of C = 3000. The corresponding

meters are shown in Table 8.1.
Assume now that the cost of the 3% sensors drops to 700. Then the optimal
solution is no longer the one shown in Table 8.1. In this case two solutions of
equal cost (C = 2900) are shown in Table 8.2. This solution is redundant.
Although the solutions obtained with this method are low in cost, some feature
nonredundant sensor networks. Though precision is achieved, gross errors are
impossible to detect. Therefore, the feasible region should contain redundant
networks. If at least one degree of redundancy is requested, that is, the threshold
for the estimability of key variables is two, then there are two solutions with a
cost of C = 3100 (Table 8.3).
MAXIMUM PRECISION MODELS
Maximum precision models include all those models developed for sensor
network design that contain a measure of the estimation quality of parameters or
state variables in the objective function. In most models, precision is minimized,
sometimes taking into account the cost as a constraint.
Madron and Veverka (1992) proposed to design sensor networks to mini-
mize the mean square error of the required quantities, by solving the following
problem:
1 X 2
Min σ (8.7)
n M i=1 i

Constrained Problem (New Cost
of Instruments).
C 3% 3% 2% —
D 3% 3% — 2%
Maximum Precision Models 179

Constrained Problem (Redundancy Required).
E 3% 3% 2% —
F 3% 3% — 2%
where σi is the standard deviation of the ith required quantity and n M is the total
number of required measurements. This model was later efficiently solved by
Madron (1992), who used the concept of minimum spanning tree. In essence,
the method is the same as the one presented in Chapter 6 for minimal sensor
networks. Indeed, all is needed is to use σi2 /n M as a cost of each stream. If all
variables need to be observable, then the same method applies. The algorithm
stops when all the spanning trees of distance one are more expensive than the
current solution. However, the costs are not fixed, and, therefore, optimality is
not guaranteed.
Alhéritière et al. (1997) proposed the following nonlinear optimization prob-
lem to design a system that will maximize the precision of one variable.

Min σθ 



s.t. 

X 
c j (σ ) = cT (8.8)


j∈M1 



c j ≤ cUj ∀ j ∈ M1 
where cT is the total resource allocated to all sensors and σ is the vector of
measurement standard deviations. The standard deviation of the parameter
estimated, σθ , can be expressed in terms of σ through:
X
σθ2 ≈ si σi2 (8.9)
i
where the coefficients si take into account the redundancy of the system and
the instrument location is assumed a priori.
This type on nonlinear programming formulation has some deficiencies:
(1) The continuous representation of variables leads to nondiscrete values for
the number of sensors.
(2) If set of measurements leading to the estimation of the parameter is redun-
dant, a smaller variance for each variable can be obtained, and, therefore,
fewer measurements can accomplish the same parameter variance.
(3) Different sets of measurements can lead to the estimation of the same
parameter. The procedure chooses one a priori. In other words, the coeffi-
cients si in Equation (8.9) are not fixed, but rather vary with the choice of
instrumentation.
GENERALIZED MAXIMUM PRECISION MODEL
In this section, a generalized model for maximum precision is presented. The

model considers the minimization of a weighted sum of the precision of the
parameters.
X 
Min a j σ j2 (q) 



j∈M p 

s.t. 
X (8.10)
ci qi ≤ cT 



i∈M1 


qi ∈ {0, 1} ∀i ∈ M1
When hardware redundancy is used, then an upper bound on the cost for
each instrument (or on number of instruments, if the costs are all equal) can be
imposed. Lower bounds on cost are not needed.
The generalized model
r Can provide a design for multiple parameter estimation, which Equation (8.8)
cannot.
r Because binary variables are used, more realistic results in accordance with
the discrete nature of sensors can be obtained.
r Takes into account redundancy and all possible forms of obtaining the
parameters.
r The minimum overall variance model (Madron and Veverka, 1992; Madron,
1992) is a particular case where cT is a large value, that is, the constraint on
total cost is dropped.
Equation (8.8) is also a particular case of Equation (8.10). Indeed, the objec-
tive function in Equation (8.8) can be obtained by assuming one parameter in
Equation (8.10). In addition, when q is a priori selected, as in Equation (8.8),
then the problem becomes NLP; the only difference between Equation (8.10)
and Equation (8.8) is that the cost constraint is an equality in Equation (8.8).
However, it can easily be proved that both problems have the same solution.
Indeed, assume that a solution e σ of Equation (8.10) is such that the constraint
on cost is not binding, that is,
X
σ ) < cT
ci (e (8.11)
i∈M1
Relation Between Sensor Network Models 181
Because the functions c j (σ ) are continuous functions of the variance σ , then

there exists a solution σ̂ in the neighborhood of σ̂ , such that
X X
cT > ci (σ̂ ) > σ)
ci (e (8.12)
i∈M1 i∈M1
Because the cost is monotonically decreasing with σ j , then the values of pre-
cision can be lowered until the cost constraint is binding, producing a reduction
in objective function. Q.E.D
RELATION BETWEEN SENSOR NETWORK MODELS
In this section a mathematical connection between the maximum preci-

sion model, given by Equation (8.10) and the minimum cost model given by
Equation (8.5) is presented. We first modify the generalized maximum precision
model by adding upper bounds on precision.
X 
Min a j σ j2 (q) 



j∈M p 

s.t. 

X 

ci qi ≤ cT (8.13)


i∈M1 


σ j (q) ≤ σ j∗ (q) ∀ j ∈ M p  



qi ∈ {0, 1} ∀i ∈ M1
The added constraints are trivial if these upper bounds are properly se-
lected. Similarly, we modify the minimum cost model by adding a trivial con-
straint consisting of the weighted average sum of existing precision constraints
(σ j (q) ≤ σ j∗ ).
X 
Min ci qi 



i∈M1 

s.t. 

X X 

a j σ j2 ≤ a j (σ j∗ ) 2
(8.14)


j∈M P j∈M P


σ j (q) ≤ σ j∗ ∀ j ∈ MP 





qi ∈ {0, 1} ∀i ∈ M1
We are now in a position to show that the minimum cost model is the dual of
the maximum precision model in the Tuy sense. This duality was established
in general by Tuy (1987) and a short version of it is reproduced next.
Tuy Duality Theorem
Consider the problems Pβ : Inf { f (x): x ∈ V, g(x) ≥ β} and Q α : Sup{g(x):

x ∈ V, f (x) ≤ α} where V is an arbitrary set in R n , f : R n → R and g: R n → R
are two arbitrary functions, and α and β are two real numbers. Then:
If the solutions of both Q α and Pβ are bounded, then
α ≤ min Pβ ⇔ β ≥ max Q α (8.15)
Duality in Sensor Network Design
Applying Tuy Duality Theorem we obtain (Bagajewicz and Sánchez, 1999)

 X 

cT = α ≤ Min f (x) = Min ci qi 


 



i∈M1 


 s.t. 


 X X 

2 ∗ 2
g(x) = − a j σ ji ≥ − a j (σ ji ) = β

 j∈M p j ∈ Mp 


 


 ∗ 


 σ j (q) ≤ σ ∀ j ∈ M p 



j


qi = {0, 1} ∀i ∈ M1
 X X 

 β=− a j (σ j∗ )2 ≥ Max g(x) = −Min a j σ j2 


 



j∈M p j∈M p 


s.t. 


 X 

⇔ f (x) = ci qi ≤ cT = α (8.16)

 



i∈M1 


 


σ j (q) ≤ σ j∗ ∀ j ∈ M p 


 

qi = {0, 1} ∀i ∈ M1
This implies, in simple terms that,

r If the constraint on cost (cT ) for the maximum precision model is smaller
than the optimum cost obtained from the minimum cost model, then the
weighted sum of variances in the maximum precision model does not reach
its maximum possible value given by the weighted sum of bounds.
r If the optimum value of the weighted sum of variances obtained in the max-
imum precision model is smaller than the weighted sum of bounds, then the
solution of the minimum cost model is larger than the maximum cost used in
the maximum precision models.
r When the minimum cost obtained from the minimum cost model is used as
an upper bound on cost in the maximum precision model, then the weighted
sum of variances is equal to the weighted sum of bounds. This also implies
Solution Procedures for Linear Systems 183
then that the constraint on weighted averaged is binding in the minimum cost
model. In other words, the solution of one problem is one solution of the other
and vice versa.
SOLUTION PROCEDURES FOR LINEAR SYSTEMS
Consider now the tree-searching algorithm outlined in Chapter 6 (Figure 6.4).

It was stated that once a node becomes feasible, the enumeration of the rest
of the branches of that node are not necessary because the addition of new
instrumentation will only increase cost. However, cost increases from one level
to the next if only one type of instrument is used (n im = 1). In such case, the first
feasible node in each branch is the one with lowest cost in the tree its spans.
If more than one type of instrument is used (n im ≥ 1), then the cost of each
node has a lower bound of
X
C(q) = c i qi (8.17)
∀i
where ci = Min {ci,k } and qi = Max { pi,k }. This property is used below in the
∀k ∀k
branching stopping criterion.
Stopping Criteria
In the case where more than one instrument is available for each stream
(n im ≥ 1), it is possible that some nodes belonging to the tree spanned by a
feasible node can be cheaper. If the solution of corresponding to the feasible
node has larger cost than its lower bound, a cheaper solution can only exist
in a few levels below this node. This number of levels is given by the nearest
Cb − C(q)
lower integer of , whereec is the cost of the cheapest instrument and
ec
Cb is the cost of the current node. Thus, after a feasible node is obtained, the
maximum number of additional levels to explore is determined. The stopping
criterion becomes, therefore, to stop after the maximum level of each feasible
branch is reached and the node is evaluated.
In the case where n im > 1 the enumeration of the tree can also be based on
developing a tree for p, rather than for q. In such case the simple criteria of
stopping the tree at the first feasible node should be used. However, the tree
will contain more nodes at each level, and, therefore, it will be larger.
This stopping criteria scheme prevents the enumeration procedure from be-
ing exhaustive and impractical. However, it is still not an efficient procedure
for large systems. Because the focus of early articles in cost models was the
conceptual development of the problem, investigations of numerical efficient
schemes were not addressed, and this aspect is still a matter of research.
PARAMETER ESTIMATION IN NONLINEAR SYSTEMS
Considerable attention is given today to the issue of parameter estimation,

especially in the context of the increasing popularity of on-line optimization,
which requires these parameter values to tune a model.
Several authors have addressed the problem of parameter estimation. One
of the early articles (Britt and Luecke, 1973) proposed the use of maximum
likelihood principle to obtain parameters in implicit models. Reilly and Patino-
Leal (1981) initiated a line of work that bases parameter estimation on lin-
earization, and Kim et al. (1997) proposed the use nonlinear programming. The
effect of data reconciliation and gross error detection in parameter estimation
was analyzed by MacDonald and Howat (1988), Serth et al. (1993) and Pages
et al. (1994). The optimization of the existing resources allocated to the sen-
sors for improving the accuracy of a parameter was attempted by Alhéritière
et al. (1997). It was reported that this method was successfully applied to a crude
distillation unit (Alheritiere et al., 1998a,b). Among the approaches based on
linear algebra, Kretsovalis and Mah (1987) proposed a combinatorial search
based on the effect of the variance of measurements on the precision of rec-
onciled values. Tjoa and Biegler (1991) explored methods for the estimation
of parameters in differential-algebraic equation systems. However, no method
to select measurements for such systems has been proposed. Finally, Krishnan
et al. (1992a,b) presented a strategy of sensor network design to satisfy the
needs of on-line optimization, and Loeblein and Perkins (1998) discussed the
economics of this issue in light of the needs of on-line optimization systems.
The method proposed by Krishnan et al. (1992a) relies on a screening
procedure that involves three steps.
(1) A first step performs a structural analysis (singular value decomposition)

that disregards measurements with little or no effect on the parameters.
(2) A second step disregards measurements that have insignificant effect on the
axis length of the confidence region of the parameter estimates.
(3) The last step determines the interaction between the parameter estimates by
means of calculating a covariance matrix (Beck and Arnold, 1977). If the
off-diagonal elements are too large, then the parameters are highly inter-
active, and, therefore, any problem with the set of measurements that affects
one parameter will also affect the other. The “best” set of measurements
will have a small confidence region and lead to low interaction between the
parameters.
Unfortunately, this method does not take into account cost and does not offer a
systematic procedure to make a final selection of the “best” set. Notwithstanding
its importance, the method is not discussed further in this book. Instead, attention
is turned towards the minimum cost model.
Parameter Estimation in Nonlinear Systems 185
Consider a given process whose steady-state operation is described by the

nonlinear algebraic system of equations f (z) = 0, where z contains the vector x
of state variables and the vector θ of process parameters
· ¸
x
z= (8.18)
θ
In the design stage of a given process, nominal values x0 are available such
that they are consistent with mass and energy balances. As it was pointed out
in Chapter 5, the model can be linearized around its expected operation point
so that the following linear system is obtained:
Jz ∼
=d (8.19)
where matrix J represents the Jacobian of f (z) around z 0 and d is the corre-
sponding constant. Matrix J may be partitioned in submatrices related to the
vectors of measured variables and unmeasured variables.
Precision of Estimates
The observability of unmeasured variables is determined through a variable

classification procedure that allows categorizing them in observable and un-
observable variables. Several strategies to do this classification have been out-
lined in Chapter 4. If the observability analysis indicates that all parameters in
M P are observable, then the set of proposed instruments represented by q is
feasible and the following step, the accuracy estimation of the parameters, can
be undertaken.
Expressions that allow the estimation of unmeasured state variables and pa-
rameters in terms of reconciled measurements are obtained from the observ-
ability analysis, as it was described in Chapter 5.
If data reconciliation procedures are not applied, then, in the presence of
redundant equations, there are several different sets of nonredundant measure-
ments that can be used to obtain the parameter. In the absence of information
about how good each of these sets are, one can only resort to using the set that
will provide the best precision or resort to averaging. No option is, however,
better than straight data reconciliation, especially because data reconciliation
is always accompanied with gross error detection.
Example 8.2
We now return to the example of a single flash unit introduced in Chapter 5

through Equations (5.108) through (5.111) and Table 5.3, for which it is desired
to estimate the vaporization efficiency ηi . For simplicity, we will assume that all
components have the same vaporization efficiency η, that is, Equation (5.111)
is rewritten as follows:
yi3 = ηyi2 Pi (sat)/P (8.20)
Assume the vector z is given by:
z T = [F1 y11 y21 y31 F2 y12 y22 y32 F3 y13 y23 y33 P η] (8.21)
Then, the corresponding matrices (J, z and d) of the linearized model around
the nominal operation point are the following:
F1 y11 y21 y31 F2 y12 y22 y32 F3 y13 y23 y33 P h
0.2 100 0 0 - 0.167 - 50.5 0 - 0.233 - 49.5
0.5 0 100 0 - 0.563 - 50.5 - 0.436 - 49.5
0.3 0 0 100 - 0.270 - 50.5 - 0.331 - 49.5
1 1 1
J= 1 1 1
1 1 1 (8.22)
- 1.395 1 1.e - 4 - 0.245
- 0.77 1 1.e - 4 - 0.458
- 1.227 1 1.e - 4 - 0.348
d T = [0.033 −0.0135 −0.0195 1. 1. 1. 0. 0. 0.] (8.23)

Assuming all flow rates, all compositions and the pressure are measured,
the best precision will be obtained. Using Q-R orthogonal factorizations for
variable classification (Chapter 5), the precision of the estimation of parameter
η is obtained as a function of measured variables.
Table 8.4 shows the results of the minimum cost model when it is solved
using different bounds for the standard deviation of the parameter.
Table 8.5 presents the results for the minimum cost model, and Table 8.6
presents results for the maximum precision model for different bounds on
standard deviation of the parameter and total cost. As the results indicate, instru-
mentation cost increases with precision requirements. Sometimes, no feasible
set is available when cost constraints are imposed. Two examples (cases 1 and 2)
are included where the minimum cost for the minimum cost model is consid-
ered as a bound for the maximum precision model; both models were run with
the same bound for the parameter standard deviation and they produce the same
results.
We now illustrate the case of multiple parameter estimation. Consider the
same flash example, with a different equilibrium correction parameter for each
component, that is, yi3 = ηi yi2 Pi (sat)/P. The minimum cost model was run
using the maximum variances indicated in Table 8.7, where optimal solutions
are presented. The optimal cost of instrumentation increases with the precision
TABLE 8.4. Set of Possible Instruments.
Instrument Instrument
Measured Instrument Standard Measured Instrument Standard
Variable Cost Deviation Variable Cost Deviation
F1 250 3.0 y32 700 0.01

y11 700 0.015 F3 300 1.418
y21 700 0.015 y13 800 0.01
y31 700 0.015 y23 800 0.01
F2 250 1.515 y33 800 0.01
y12 700 0.01 P 100 14.0
y22 700 0.01
TABLE 8.5. Results for the Minimum Cost Model.
Case No. σ∗ σ Cost Optimal Set
1 0.05 0.00865 1600 y12 , y33 , P

2 0.007 0.005896 1750 F1 , y12 , y22 , P
3 0.006 0.005896 1750 F1 , y12 , y22 , P
4 0.0058 0.00551 2300 y22 , y32 , y13 , P
5 0.005 0.004982 2900 F1 , F2 , y22 , y13 , y23 , P
TABLE 8.6. Results for the Maximum Precision Model.
Case No. σ∗ cT∗ σ Cost Set
1 0.05 1600 0.00865 1600 y12 , y33 , P

2 0.007 1750 0.005896 1750 F1 , y12 , y22 , P
3 0.0058 3600 0.0048 3600 F1 y21 F2 y22 y13 y23 P
4 0.0055 1750 — — —
5 0.005 2500 — — —
6 0.005 3000 0.004982 2950 F2 y22 F3 y13 y23 P
Case
No. ση∗1 ση∗2 ση∗3 ση1 ση2 ση3 Cost Optimal Set
1 0.1 0.2 0.09 0.0990 0.0265 0.0520 3100 Y22 , y32 , y23 , y33 , P
2 0.07 0.09 0.09 0.0660 0.0288 0.0486 3900 y12 , y32 , y13 , y23 , y33 , P
3 0.06 0.06 0.06 0.0572 0.0218 0.0425 4600 y12 , y22 , y32 , y13 , y23 , y33 , P
4 0.04 0.05 0.05 — — — — —
187
Case No. ση∗1 ση∗2 ση∗3 cT ση1 ση2 ση 3
1 — — — 3100 0.0990 0.0265 0.05209

2 0.07 0.09 0.09 3900 0.0660 0.0288 0.0486
3 0.06 0.06 0.06 4600 0.0573 0.02180 0.04258
4 0.06 0.06 0.06 7000 0.0571 0.02179 0.04254
5 0.5 0.5 0.5 4000 0.07008 0.026572 0.073572
6 0.5 0.5 0.5 3100 0.0990 0.0265 0.05209
7 0.06 0.06 0.06 4600 0.0573 0.02180 0.04258
8 0.06 0.03 0.05 6500 0.05713 0.02178 0.04256
9 0.05 0.03 0.05 9000 — — —
of the estimates of the parameters, but, for the last case, no set of available
instruments can fulfill precision requirements.
Optimal designs obtained by applying maximum precision models are pre-
sented in Tables 8.8 and 8.9. For cases 1–4, all weights in the objective function
are considered equal to one. In case 1, bounds on the standard deviation are
removed. In cases 2–4, higher costs and higher precision of the parameters are
allowed. This leads to lower weighted sum of parameter variances.
For cases 5–8, the weights in the objective function values are [5 1000 1].
The same optimal set is obtained for cases 1–6 and 3–7 of Table 8.8. They have
the same feasible region and differ only in weight values. Higher precision is
required in case 8 and an unfeasible situation is presented in case 9. Case 2 and
cases 3–7 of Tables 8.7 and 8.8 are examples of duality.
From the results of the above examples, some conclusions and recommen-
dations can be made. If bounds on standard deviation of the parameters are
available, then minimum cost models are a better alternative because the selec-
tion of weights for the objective function can be avoided. If, in turn, bounds

X
ση2i
Case No. c i Optimal Set
1 3100 0.01326 y22 , y32 , y23 , y33 , P

2 3900 0.00750 y12 , y32 , y13 , y23 , y33 , P
3 4600 0.00557 y12 , y22 , y32 , y13 , y23 , y33 , P
4 6800 0.00555 F1 , y11 , y31 , F2 , y12 , y22 , y32 , F3 , y13 , y23 , y33 , P
5 3600 0.736044 F1 , F2 , y12 , y22 , y13 , y23 , P
6 3100 0.757837 y22 , y32 , y23 , y33 , P
7 4600 0.493668 y12 , y22 , y32 , y13 , y23 , y33 , P
8 6500 0.492814 F1 , y11 , y21 , F2 , y12 , y22 , y32 , y13 , y23 , y33 , P
9 — — —
Nomenclature 189
are not available, maximum precision models constrained only by cost can
be used, using for example all weights equal to one. Different weights may
be selected to reflect the relative importance of the precision of the para-
meters. However, instead of this, after the maximum precision model is run
with proposed weights, one can use the minimum cost model and perform a
more meaningful sensitivity analysis in terms of cost, by using values suggested
by the result of the maximum precision model as bounds on precision.
NOMENCLATURE
C(.): cost function

C(.): lower bound on cost
e
c: cost of the cheapest instrument
cT : bound on total cost
mj : maximum number of sensors to measure variable i
Ml : set of variables where sensors can be placed
MP : set of key variables for precision
nM: number of measurements
n im : number of instruments available to measure variable i
P: pressure
pi,k : binary vector indicating whether sensor k is located (1) or not (0) in
variable i
q: binary vector indicating whether sensors are located (1) or not (0) in
variables
Si : stream i
Ui : unit i
e
y: molar fractions
z: vector of state variables, and parameters
Greek Letters
η: vaporization efficiency
σ: precision
σ ∗: precision threshold
e
σ: precision of estimates
θ: vector of parameters
REFERENCES
2306 (1997).
Bagajewicz M. and M. Sánchez. Duality of Sensor Network Design Models for Parameter Estima-
tion. AIChE J., 45, 3, pp. 661–664 (1999).
Beck J. V. and K. J. Arnold. Parameter Estimation for Engineering and Science. Wiley, New York
(1977).
Britt H. I. and R. H. Luecke. The Estimation of Parameters in Nonlinear Implicit Models. Techno-
metrics, 15, 2, pp. 233–247 (1973).
Kim I., M. S. Kang, S. Park and T. F. Edgar. Robust Data Reconciliation and Gross Error Detection:
The Modified MIMT Using NLP. Comp. & Chem. Eng., 21, 7, pp. 775–782 (1997).
Process Networks. AIChE J., 33, pp. 70–82 (1987).
Krishnan S., G. Barton and J. Perkins. Robust Parameter Estimation in On-line Optimization. Part I.
Methodology and Simulated Case Study. Comp. & Chem. Eng., 16, pp. 545–562 (1992a).
Krishnan S., G. Barton and J. Perkins. Robust Parameter Estimation in On-line Optimization.
Part II. Application to an Industrial Process. Comp. & Chem. Eng., 17, pp. 663–669 (1992b).
Loeblein C. and J. D. Perkins. Economic Analysis of Different Structures of On-Line Process
Optimization Systems. Comp. & Chem. Eng., 22, 9, pp. 1257–1269 (1998).
MacDonald R. and C. Howat. Data Reconciliation and Parameter Estimation in Plant Performance
Analysis. AIChE J., 34, 1 (1988).
Madron F. and V. Veverka. Optimal Selection of Measuring Points in Complex Plants by Linear
Models. AIChE J., 38, 2, pp. 227 (1992).
Pages A., H. Pingaud, M. Meyer, X. Joulia. A Strategy for Simultaneous Data Reconciliation and
Parameter Estimation on Process Flowsheets. Comp. & Chem. Eng., 18, Suppl., pp. S223–S227,
(1994).
Reilly P. M. and H. Patino-Leal. A Bayesian Study of the Error-in-Variables Model. Technometrics,
23, 3, pp. 221 (1981).
Serth R., B. Srinkanth, S. Maronga. Gross Error Detection and Stage Efficiency Estimation in a
Separation Process. AIChE J., 39, pp. 1726 (1993).
Tjoa I. B. and L. T. Biegler. Simultaneous Solution and Optimization Strategies for Parameter
Estimation of Differential-Algebraic Equation Systems. Ind. Eng. Chem. Res., 30, pp. 376–385
(1991).
Tuy H. Convex Programs with an Additional Reverse Convex Constraint. JOTA, 52, pp. 463–486
(1987).
CHAPTER 9
Precision Upgrade of Sensor Networks
INTRODUCTION
In this chapter the issues related to the upgrade of instrumentation to improve

the accuracy of the estimation of key variables are discussed and specific models
are presented.
As pressure mounts to obtain more reliable and accurate estimates of variables
and parameters, the use of data reconciliation techniques is the first level of
response. In many cases, the existing redundancy is not enough to guarantee
the level of accuracy required. Thus, one is faced with an existing set of sensors,
to which an addition of new sensors will provide the desired accuracy.
UPGRADE OPTIONS
There are three possible ways of performing the upgrade of a sensor network.
They are:
(1) addition of new instruments
(2) substitution of existing instruments by new ones
(3) relocation of existing instruments
Typically, addition of new instruments has been the response first consid-
ered. However, substitution and/or relocation are options that are sometimes
substantially cheaper. For example, the substitution of thermocouples by ther-
moresistances or their relocation is comparatively inexpensive. However, it
is in the case of laboratory analysis where the options of substitution and/or
relocation should be strongly considered.
191
192 PRECISION UPGRADE OF SENSOR NETWORKS
COST BENEFIT ANALYSIS
Because upgrading requires capital expenditure, it must be done on the basis

of a cost benefit analysis. The costs of the instrumentation are straightforward
to obtain. However, the benefits need to be somehow quantified. In the case of
data accuracy needs for accounting purposes, the benefit can be quantified as the
decrease of lost revenue due to imprecise data. Describing it in simple terms,
the larger the uncertainty in the assessment of the amount of raw materials
purchased and/or the products sold, the larger the probability of loss revenue.
Thus, every percent accuracy of these flows can be assigned a monetary value.
In the case of monitoring and parameter estimation for on-line optimization,
an economic measure can also be developed. First, in the case of monitor-
ing, one can associate a revenue loss for product that is not in specification.
Finally, in the case of on-line optimization Loeblein and Perkins (1998) discuss
the use of measures of loss of economic performance due to offset from the
correct optimum due to plant-model mismatches.
In other cases, industry is looking less at the benefit in monetary terms and
simply plots the increased precision as a function of investment (Alhéritière
et al., 1998b). Although this approach is intuitive, its importance relies on the
possibility of visualizing the effect of instrumentation cost.
UPGRADE MODELS BASED ON ADDITION OF SENSORS
The upgrading of a sensor network by the simple addition of instrumentation

is considered first. This has to be done at minimum cost while reaching the
goals of precision in key variables. This task can be accomplished by using the
following minimum cost model (Bagajewicz and Sánchez, 2000):
XX 
Min cik qikN 



i∈M1 k∈K i 

s.t. 



σ j (q) ≤ σ j∗ ∀ j ∈ Mp (9.1)
X 

qikN + Ni ≤ Ni∗ ∀i ∈ M1 



k∈K i 



qikN ∈ {0, 1} ∀i ∈ M1 , ∀k ∈ K i
where
M p = set of all variables for which precision constraints are imposed.
M1 = set of all variables that can be measured.
K i = set of new sensors available to measure variable i.
Ni = number of existing sensors measuring variable i.
cik = cost of the k-new sensor for measuring variable i.

Ni∗ = maximum number of sensors that are allowed to be used to
measure variable i.
q N = vector of binary variables for new sensors.
The first constraint establishes a bound on precision, which is a function of

the set of fixed existing sensors and new instrumentation. The second constraint
establishes an upper bound on the number of sensors used to measure each vari-
able. This number is usually one for the case of flowrates, but it can be larger
in the case of laboratory measurements of concentrations. In this case, these
constraints can be lumped in one constraint for all concentration measurements
to express the overall limitation of the laboratory on a daily (or fraction thereof )
basis.
When maximum precision is requested and cost is a constraint, then a max-
imum precision model of the following type can be used:
X 
Min a j σ j2 (q N ) 



j 

s.t. 

XX 

cik qikN ≤ cT 

i∈M1 k∈K i (9.2)
X 



qikN + Ni ≤ Ni∗ ∀i ∈ M1 



k∈K i 



qikN ∈ {0, 1} ∀i ∈ M1 , ∀k ∈ K i
where cT is the bound on the capital expenditure. The two models are equivalent
in the Tuy sense (Chapter 8).
Alhéritière et al. (1997, 1998a) presented a model without binary variables
that can provide a first good approximation to the upgrade. However, this model
has some limitations, because it can render fractional instruments and relates the
cost to the precision in a continuous manner. These limitations were discussed
in Chapter 8. Later, Alhéritière et al. (1998b) addressed these limitations and
presented an MINLP model.
Example 9.1
In Chapter 5 a simplified flash tank model and nominal variable values from
Van Winkle (1967) were presented. The model has been used to illustrate the
design of accurate sensor networks in Chapter 8.
In this case, a set of three initial instruments is installed on the unit. The mea-
sured variables are [y12 , y33 , P]. New instruments are necessary to satisfy pre-
cision requirements on the vaporization efficiency coefficient η. Measurements
TABLE 9.1. Set of Possible Instruments

for the Flash.
Instrument
Measured Instrument Standard
Variable Cost Deviation
F1 250 3.0
y11 700 0.015
y21 700 0.015
y31 700 0.015
F2 250 1.515
y12 700 0.01
y22 700 0.01
y32 700 0.01
F3 300 1.418
y13 800 0.01
y23 800 0.01
y33 800 0.01
P 100 14.0
are obtained by installing instruments chosen from the set presented in

Table 9.1. In this case, hardware redundancy is not considered, and only one type
of instrument is available to measure a variable. The optimal set of additional
instruments is determined by solving both the minimum cost model and the
maximum precision model. Results from each model are provided in Table 9.2
and 9.3, respectively. The standard deviation of the parameter is calculated by
using Equation (7.28).
In some examples, higher requirements in precision are satisfied by incor-
porating a large set of instruments and increasing the cost. But in other cases,
there is no a feasible set of instruments that fulfils precision constraints.
Example 9.2
Consider the industrial heat exchanger network of Figure 9.1. It corresponds

to a set of heat exchangers where crude is heated up by using hot gas-oil
TABLE 9.2. Minimum Cost Optimal Solutions for Flash Sensor Upgrading.
Case σ∗ σ Cost Additional Optimal Set
1 0.2 0.00866 0 —
2 0.006 0.00574 1200 F1 , F2 , y22
3 0.005 0.00480 2550 F1 , y22 , y13 , y23
4 0.0046 0.00459 4900 F1 , y11 , y21 , F2 , y22 , y32 , y13 , y23
5 0.004 — — —
TABLE 9.3. Maximum Precision Optimal Solutions for Flash Sensor Upgrading.
Case cT σ∗ σ Cost Additional Optimal Set
1 1700 ∞ 0.0087 0 —
2 1200 0.006 0.00574 1200 F1 , F2 , y22
3 2550 0.005 0.00480 2550 F1 , y22 , y13 , y23
4 3000 0.005 0.00474 3000 y22 , y32 , y13 , y23
5 5500 0.0046 0.00459 5200 F1 , y11 , y21 , F2 , y22
y32 , F3 , y13 , y23
6 7000 0.004 — — —
coming from a column. In this case, the heat transfer coefficients for the heat
exchangers are estimated using temperature and flow rate measurements. The
existing instrumentation is given in Table 9.4, and data for the streams and the
units are given in Tables 9.5 and 9.6.
The standard deviations of heat transfer coefficients calculated by using the
installed set of instruments are [12.27 2.96 3.06]. To obtain these values,
all redundant measurements have been used. To enhance the precision of the
parameter, new instruments should be added. In this example, hardware re-
dundancy is considered. Furthermore, different types of new instruments are
available to measure some temperatures. Data for new instrumentation are pre-
sented in Table 9.7, where costs and standard deviations are shown.
The maximum number of allowed instruments for measuring each variable
is given in Table 9.8; a zero value indicates there is a restriction for measuring
the corresponding variable.
Tables 9.9 and 9.10 present results for the upgrade problem by using both
types of models. When there are two possible instruments to measure a variable,
the type of instrument is indicated between parenthesis in the optimal solution
set. The weights for the maximum precision model are assumed equal to one.
Case 1 from Table 9.9 and case 3 from Table 9.10 illustrate the equivalency of
the models.
FIGURE 9.1. Industrial heat exchanger network.

TABLE 9.4. Existing Instrumentation for the Heat Exchanger Network.
Standard Standard
No. of Deviation of No. of Deviation of
Streams Flowmeters Flowmeters Thermocouples Thermocouples
S1 1 3% 1 2◦ F
S3 — — 1 2◦ F
S4 — — 1 2◦ F
S5 1 3% 1 2◦ F
S7 1 3% 1 2◦ F
S9 — — 1 2◦ F
TABLE 9.5. Flow Rate Information for

Figure 9.1.
Flowrate Temperature
Stream (lb/hr) (◦ F)
S1 224,677 542.8
S2 224,677 516.3
S3 224,677 448.3
S4 224,677 402.2
S5 217,019 307.6
S6 217,019 339.8
S7 398,008 191.2
S8 398,008 221.6
S9 398,008 266.9
TABLE 9.6. Unit Information for Figure 9.1.
Heat Exchanger Area (ft 2 ) FT Cph (BTU/lb◦ F) Cpc (BTU/lb◦ F)
U1 500 0.997 0.6656 0.5689

U2 1100 0.991 0.6380 0.5415
U3 700 0.995 0.6095 0.52
TABLE 9.7. Availability of New Instrumentation.
Flowmeters Temperature Sensors

Standard Standard
Stream Deviation Cost Deviation Cost
1 3% 2250 2F/0.2◦ F 500/1500

2 3% 2250 2◦ F 500
3 3% 2250 2◦ F 500
4 3% 2250 2F/0.2◦ F 500/1500
5 3% 2250 2◦ F 500
6 3% 2250 2◦ F 500
7 3% 2250 2◦ F 500
8 3% 2250 2◦ F 500
9 3% 2250 2◦ F/0.2◦ F 500/1500
196
TABLE 9.8. Maximum Number of

Instruments for the Heat Exchanger Network.
Variable Ni∗ Variable Ni∗
F1 1 T1 2
F2 1 T2 1
F3 1 T3 1
F4 1 T4 2
F5 1 T5 1
F6 0 T6 1
F7 1 T7 1
F8 0 T8 0
F9 1 T9 2
Combinatorial Analysis
Even though the above presented model is general enough to cover all possi-
bilities, we briefly review some early attempts to solve the problem for steady-
state systems with only random measurement noise. Kretsovalis and Mah (1987)
developed a combinatorial strategy to incorporate measurements to an observ-
able system. Each time a new instrument i is added, the following objective
function is minimized:
¡ ¢ ¡ ¢
Ri = a c0 + cik + H0 + 1Hik (9.3)
In this function, a is a weighting factor, c0 the cost of the existing instru-

mentation, cik the cost associated with the instrument i placed at position k, H0
measure of the estimation error corresponding to the original structure and 1Hik
the change in the estimation error when an instrument i is placed at position k.
The estimation error H0 is calculated as the trace of the variance matrix, that
e
is, H0 = tr( Q).
This model is equivalent to solving a slightly modified version of model (9.2),
using a penalty function approach and in a sequential manner, that is, for the

Case σU∗1 σU∗2 σU∗3 σU1 σU2 σU3 Cost Optimal Set
1 4.0 4.0 4.0 3.6160 1.9681 2.7112 500 T6 (1)
2 3.5 2.0 2.5 2.7746 1.6892 2.3833 1500 T2 (1)T4 (1)T6 (1)
3 3.0 1.5 2.5 2.7230 1.4972 2.2844 6500 F2 F3 T2 (1)T4 (1)T6 (1)T9 (1)
F2 F4 T2 (1)T4 (1)T6 (1)T9 (1)
F3 F4 T2 (1)T4 (1)T6 (1)T9 (1)
4 3.5 2.0 2.0 — — — — —

X
∗
ση2i
Case σ ∗
U1 σ∗
U2
σ
U3 Cost σU1 σU2 σU3 i Optimal Set(s)
1 — — — 6250 2.5379 1.5962 2.1341 13.4578 F2 T1 (2)T2 (1)T4 (2)T6 (1)

2 — — — 3000 2.6635 1.6854 2.2068 14.8048 T1 (1)T2 (1)T4 (2)T6 (1)
3 4.0 4.0 4.0 500 3.6160 1.9681 2.7112 24.2999 T6 (1)
½
F2 F3 T2 (1)T4 (1)T6 (1)T9 (1)
4 3.0 1.5 2.5 6500 2.7230 1.4972 2.2844 14.8749 Alternatives F2 F4 T2 (1)T4 (1)T6 (1)T9 (1)
F3 F4 T2 (1)T4 (1)T6 (1)T9 (1)
5 3.0 1.5 2.5 4500 — — —
addition of one instrument at a time. Indeed, consider the addition to model (9.2)
of a constraint stating that the total number of instruments added is only one, with
no upper bound on cost, that is, cT → ∞. Then the second term of model (9.3)
is the objective function of this modified version of model (9.2) with all weights
equal to one. In turn, the first term is the penalty function consisting of the cost.
We will not elaborate on this method further, because it is obvious it cannot
guarantee optimality. In fact, at any given step, it may lead to the addition of a
sensor that is not part of the optimal solution.
An improved combinatorial method for linear systems is to first consider
the addition of one sensor. The result of this procedure is a sensor to be added
and a cost c1 . The minimum cost model, model (9.1), is recommended, but
if bounds on precision are not known, model (9.2) can be used. In the next
step, one can consider the addition of two instruments, resulting in a pair of
new instruments (which does not necessarily have to include the one found
when only one instrument was added) and a cost c2 . In this manner, one can
continue considering an increasing number of sensors and construct a table of
cost benefit. In this way, one can choose how many instruments one wants to
add. It is apparent that this procedure is less efficient than the use of MINLP
procedures to solve models (9.1) or (9.2) directly. However, one can readily take
advantage of this approach when a quick evaluation is needed and the solvers
for models (9.1) or (9.2) are not available.
MODEL FOR RESOURCE REALLOCATION
In many cases, measurements can be easily transferred at no cost from

one stream to another. This is the case of concentration measurements that
are performed in the laboratory. Pressure gauges and thermocouples can also
be transferred from one place to another at a relatively small cost. However,
flowmeters are probably an exception. Because one can consider that the re-
source reallocation does not involve cost, or eventually neglect it, the minimum
cost model would reduce to a set of MINLP algebraic inequalities. However,
Model for Resource Reallocation 199
even in the case where cost is not a matter of consideration, one would like
to minimize the number of changes. The derivation of a model presented by
Bagajewicz and Sánchez (2000) is now presented.
Let us introduce the binary variable u t,k,r that indicates sensor k from vari-
able t is relocated to variable r . This sensor is the k-th element of the set K t of
originally allocated sensors for measuring variable t. Following, a model that
minimizes the number of reallocations is presented.
X XX 
Min u t,k,r 



t∈MT k∈K i r ∈Mr 

s.t. 



σ j (u) ≤ σ j∗ ∀ j ∈ MP 



X X 

u t,k,r ≤ u tT ∀t ∈ MT 



r ∈M R k∈K t 
X X (9.4)
u t,k,r ≤ u rR ∀r ∈ M R 



t∈MT k∈K t Ã ! 

X X X X 



u t,k,i + Ni − u i,k,r ≤ Ni∗ ∀i ∈ M1 



t∈MT k∈K t r ∈M R k∈K i 


u t,k,r ∈ {0, 1} ∀t ∈ MT , ∀r ∈ M R 



∀k ∈ K t
where
Ni = number of originally allocated sensors for measuring variable i, that
is, the cardinality of K i .
M R = set of streams where instruments from other place can be reallocated.
MT = set of streams whose instruments can be reallocated.
In this model, bounds are imposed on:
(1) The maximum number of instruments that can be reallocated from variable t
to other variables (u tT ).
(2) The maximum number of instruments that can be reallocated from other
variables to variable r (u rR ).
(3) The total number of instruments allocated for each variable (Ni∗ ). This
is needed because one may allow a large number of instruments to be
reallocated to a certain variable r , but this change has to be accompanied
with the reallocation to some of the instruments from K r to some other
variables to maintain the total number of sensors limited.
The value of the parameter standard deviation σ j (t) is obtained by using
Equation (7.28).
As in the case of instrumentation upgrading, a maximum precision problem
that is the dual (in the Tuy sense) of the minimum cost model can be defined
as follows:
X 
Min a j σ j2 (u) 



j∈M P 

s.t. 

X X X 



u t,k,r ≤ Nr 



t∈MT k∈K t r ∈M R 



σ j (u) ≤ σ j∗ ∀ j ∈ MP 



X X 

u t,k,r ≤ u tT ∀t ∈ MT 
r ∈M R k∈K t
X X 






t∈MT k∈K t
Ã ! 



X X X X 

u t,k,i + Ni − ≤ Ni∗ ∀i ∈ M1 

u i,k,r 

r ∈M R k∈K i 

t∈MT k∈K t 


u t,k,r ∈ {0, 1} ∀t ∈ MT , ∀r ∈ M R 



∀k ∈ K t
(9.5)
This model:
r Includes binary variables to handle hardware redundancy; therefore, it does
not reallocate fractions of sensors.
r Considers software redundancy through data reconciliation.
r Does not commit to a specific set of sensors.
r Does not include cost as a bound.
The last issue of cost is important to be addressed in more detail. It makes
sense that the constraint on cost is superfluous, because the problem is the
reallocation of existing resources, which already have a fixed cost.
When the bounds are too tight, the set of equations becomes infeasible and the
reallocation is not possible. When precision bounds are not known, model (9.5)
can be used setting the precision bounds to a large value. However, proper
weights need to be used. A satisfactory criteria is not yet available to choose
these weights.
Example 9.3
Let us consider the reallocation of duplicate existing thermocouples as an

attempt to fulfill precision requirements for the heat exchanger network pre-
sented in Example 9.2. The existing instrumentation is shown in Table 9.11;
the standard deviation of the heat transfer coefficients with these sensors is
[11.65 2.87 2.46]. The corresponding bounds are shown in Table 9.12
(flowmeters are not reallocated).
TABLE 9.11. Existing Instrumentation for the Heat Exchanger Network.
Standard
No. of Deviation of No. of Standard Deviation
Streams Flowmeters Flowmeters Thermocouples of Thermocouples
S1 1 3% 2 2◦ F/0.2◦ F
S3 — — 1 2◦ F
S4 — — 2 2◦ F/0.2◦ F
S5 1 3% 2 2◦ F/2◦ F
S7 1 3% 1 2◦ F
S9 — — 1 2◦ F
Results are given in Tables 9.13, 9.14a and 9.14b. Note first that several al-
ternative solutions exist for each case. The example shows that higher require-
ments in precision may be fulfilled by increasing the number of reallocations
of existing thermocouples. Furthermore, three examples are included where the
minimum number of reallocations for the minimum number of sensor models
is considered as a bound for the maximum precision model, and both models
were run with the same bounds for the standard deviation of the parameters. The
same solution is obtained for both models in terms of the set of reallocations
and standard deviation of the parameters.
GENERALIZED MODEL FOR RESOURCE REALLOCATION

AND UPGRADE
In this section a resource reallocation model based on cost minimizing is

presented. Switching from one sample to another in the laboratory may prompt
changes of reagents, changes in sampling costs etc. The change of positions
of thermocouples and pressure gauges have also rewiring costs and recalibra-
tion costs. Finally, one may want to seriously consider reallocating flowme-
ters, although this may not be a likely outcome because of the relatively
higher cost. In addition, these reallocation costs may overcome the simple
addition of new instrumentation. Therefore, any reallocation and upgrade pro-
gram should consider the trade-off between all these decisions. This trade-off
TABLE 9.12. Bounds for the Reallocation

Problem.
T1 T2 T3 T4 T5 T6 T7 T8 T9
u tT 1 0 0 1 1 0 0 0 0
u rR 0 2 1 0 0 1 1 1 1
∗
N 2 2 2 2 2 1 2 1 2
202
TABLE 9.13. Reallocation Model—Minimum Number of Reallocations Results for the Heat Exchanger Network.
No. of Optimal Reallocations

Case σU∗1 σU∗2 σU∗3 σU1 σU2 σU3 Reallocations (t , k , r )
1 4.00 1.80 2.40 3.1683 1.5458 2.2124 1 (T1 , 2, T2 )

1.9019 1.5458 2.2124 1 (T5 , 2, T2 )
2 1.35 1.53 2.70 1.3487 1.5236 2.6623 2 (T1 , 2, T6 ) and (T4 , 2, T2 )
or
(T1 , 2, T2 ) and (T4 , 2, T6 )
1.3455 1.5188 2.6610 2 (T4 , 2, T6 ) and (T5 , 2, T2 )
or
(T4 , 2, T2 ) and (T5 , 2, T6 )
3 2.00 1.40 2.00 1.9057 1.3904 1.9411 2 (T4 , 2, T2 ) and (T5 , 2, T8 )
or
(T4 , 2, T8 ) and (T5 , 2, T2 )
4 1.80 1.40 2.00 1.7901 1.3730 1.9297 3 (T1 , 2, T2 )(T4 , 2, T6 ) and (T5 , 2, T8 )
or
(T1 , 2, T2 )(T4 , 2, T8 ) and (T5 , 2, T6 )
or
(T1 , 2, T6 )(T4 , 2, T2 ) and (T5 , 2, T8 )
or
(T1 , 2, T6 )(T4 , 2, T8 ) and (T5 , 2, T2 )
or
(T1 , 2, T8 )(T4 , 2, T2 ) and (T5 , 2, T6 )
or
(T1 , 2, T8 )(T4 , 2, T6 ) and (T5 , 2, T8 )
TABLE 9.14a. Reallocation Model—Maximum Precision Model Results

for the Heat Exchanger Network.
Case σU∗1 σU∗2 σU∗3 Nr∗ σU1 σU2 σU3
1 1.50 1.50 2.00 1 — — —

2 1.80 1.40 2.00 3 1.7901 1.3730 1.9297
3 4.00 1.80 2.40 1 1.9019 1.5458 2.2124
4 4.00 4.00 4.00 2 1.6973 1.5694 1.9573
5 4.00 4.00 4.00 3 1.7901 1.3730 1.9297
6 2.00 1.40 2.00 2 1.9057 1.3904 1.9411
is taken into account if the following upgrade cost function is used (Bagajewicz
and Sánchez, 2000):
XX X X X
cik qikN + h t,k,r u t,k,r (9.6)
i∈M1 k∈K i t∈MT k∈K t r ∈M R
where h t,k,r represents the cost of reallocation of the type k instrument from
variable t to variable r . The complete generalized reallocation and upgrading
TABLE 9.14b. Reallocation Model—Maximum Precision Model Results

for the Heat Exchanger Network.
P Optimal Reallocations
ση2i
Case Nr i
(t , k , r )
1 — — —
2 3 8.8130 (T1 , 2, T2 )(T4 , 2, T6 ) and (T5 , 2, T8 )
or
(T1 , 2, T2 )(T4 , 2, T8 ) and (T5 , 2, T6 )
or
(T1 , 2, T6 )(T4 , 2, T2 ) and (T5 , 2, T8 )
or
(T1 , 2, T6 )(T4 , 2, T8 ) and (T5 , 2, T2 )
or
(T1 , 2, T8 )(T4 , 2, T2 ) and (T5 , 2, T6 )
or
(T1 , 2, T8 )(T4 , 2, T6 ) and (T5 , 2, T8 )
3 1 10.9016 (T5 , 2, T2 )
4 1 9.1751 (T1 , 2, T6 ) and (T4 , 2, T8 )
or
(T1 , 2, T8 ) and (T4 , 2, T2 )
5 3 8.8130 Same as in case 2
6 2 9.3325 (T4 , 2, T2 ) and (T5 , 2, T8 )
or
(T4 , 2, T8 ) and (T5 , 2, T2 )
model, which is a generalization of model (9.4) is the following (Bagajewicz

and Sánchez, 2000):
XX X X X 
Min cik qikN + h t,k,r u t,k,r 



i∈M1 k∈K i t∈MT k∈K t r ∈M R 

s.t. 



∗ 

σ j (q , u) ≤ σ j
N
∀ j∈M P 

X X 



u t,k,r ≤ u t T
∀t∈MT 



r ∈M R k∈K t 

X X 

u t,k,r ≤ u r R
∀r ∈M R


t∈MT k∈K t
Ã ! 



X X X X X 

qik +
N
u t,k,i + Ni − ∗
u i,k,r ≤ Ni ∀i∈M1 



r ∈M R k∈K i 

k∈K i t∈M R k∈K t 



u t,k,r ∈ {0, 1} ∀t ∈ MT , ∀r ∈M R 



∀k∈K t 



qik ∈ {0, 1}
N
∀i ∈ M1 , ∀k∈K i
(9.7)
An equivalent maximum precision model can be obtained by finding the dual

in the Tuy sense:
X 
Min a j σ j2 (q N , u) 



j∈M p 

s.t. 



XX X X X 

cik qikN + h t,k,r u t,k,r ≤ cT 



t∈MT k∈K t r ∈M R 

i∈M1 k∈K i




σX
j (q X
N
, u) ≤ σ j∗ ∀ j ∈ MP 



u t,k,r ≤ u tT ∀t ∈ MT 


r ∈M R k∈K t
X X 





t∈M R k∈K t Ã ! 

X X X X X 

qikN + u t,k,i + Ni − u i,k,r ≤ Ni∗ ∀i ∈ M1 





k∈K i t∈MT k∈K t r ∈M R k∈K i 



u t,k,r ∈ {0, 1} ∀t ∈ MT , ∀r ∈ M R 




∀k ∈ K t 


qikN ∈ {0, 1} ∀i ∈ M1 , ∀k ∈ K i
(9.8)
TABLE 9.15. Installed and New Instrumentation Data.
Variable
Index Variable Ni σi ,k ci ,k σi ,k
1 F1 2 2.5 2.5 350 2.

2 y1 2 0.015 0.01 2700 0.01
3 F2 1 1.515 — 350 1.48
4 y2 1 0.01 — 2700 0.01
5 F3 1 1.418 — 400 1.38
6 y3 1 0.01 — 2700 0.01
7 P 1 14. — 100 14.
Example 9.4
Consider the flash tank Example 9.1. The number of existing instruments
to measure each variable (Ni ) and the corresponding standard deviations (σi,k )
for different types of installed instruments are included in Table 9.15. In this
example the mass fractions of all the components of a stream are measured
on-line; a laboratory analysis may be done as a second alternative to know their
values.
The standard deviation of the vaporization efficiency coefficient is estimated
using the existing instrumentation is 0.00434. Because this value is not sat-
isfactory, a reallocation and possibly an incorporation of new flowmeters and
laboratory composition analysis are proposed. The constraint bounds are pre-
sented in Table 9.16. The sets of transferred and received measurements are
MT = {1, 2, 3} and M R = {1, 3, 4, 6} respectively. The costs of feasible reallo-
cations between sets MT and M R are given in Table 9.17. For example, the cost
of relocating a sensor from F1 to F2 is 80. Infinite costs are used for forbidden
reallocations. Typically, this relocation pattern is constructed using engineering
judgment.
Table 9.18 shows the results for the minimum cost model. The first row
represents the case for the existing instrumentation. It is interesting to notice
that a reduction of the standard deviation from 0.00438 to 0.00347 results if
TABLE 9.16. Flash Drum: Constraints Bounds

for Sensor Reallocation and Upgrade.
Variable F1 y1 F2 y2 F3 y3 P
N∗ 2 2 2 2 2 2 2
u iT 2 1 1 0 0 0 0
u iR 1 0 2 1 0 1 0
TABLE 9.17. Flash Drum: Costs
of Relocation.
t /r F1 F2 y2 y3
F1 ∞ 80 ∞ ∞
y1 ∞ ∞ 0 50
F2 80 ∞ ∞ ∞
TABLE 9.18. Flash Drum: Minimum Model Results for Sensor Reallocation
and Upgrade.
Reallocations New
Case σ∗ σ Cost (t , k, r ) Instruments
1 — 0.00438 —
2 0.0038 0.00352 100 — P
0.00347 100 (2,2,4) P
3 0.0033 0.00329 2800 (2,2,4) y3 P
4 0.0031 — — — —
TABLE 9.19. Maximum Precision Model Results for Sensor

Reallocation and Upgrade.
New
Case σ∗ cT σ2 σ Cost Reallocations Instruments
1 — — 1.918e-5 0.00438 — — —
2 0.0038 100 1.202e-5 0.00347 100 (y1 , 2, y2 ) P
3 0.0033 2050 — — — — —
4 0.0033 2800 1.080e-5 0.00329 2800 (y1 , 2, y2 ) y3 –P
5 0.0033 3550 1.0797e-5 0.00329 3550 (y1 , 2, y2 ) F2 F3 y3 P
TABLE 9.20. Heat Exchanger Network: Bounds for Reallocation

and Upgrade Problems.
Variable Ni utT urR Variable Ni uiT urR
F1 1 1 0 T1 2 1 0
F2 1 0 0 T2 2 0 1
F3 1 0 1 T3 2 0 0
F4 1 0 0 T4 2 1 0
F5 1 0 0 T5 2 1 0
F6 0 0 0 T6 2 0 1
F7 1 1 0 T7 2 0 0
F8 0 0 0 T8 0 0 0
F9 1 0 1 T9 2 0 0
206
TABLE 9.21. Heat Exchanger Network:
Relocation Cost.
MT /M R F3 F9 T2 T6
F1 1000 ∞ ∞ ∞
F7 ∞ 1000 ∞ ∞
T1 ∞ ∞ 100 100
T4 ∞ ∞ 100 100
T5 ∞ ∞ 100 100
the laboratory analysis for the feed stream is relocated to the liquid stream and
a pressure sensor is added. The cost of this case is 100. Higher precision is
obtained by means of the reallocation and addition of instruments.
In Table 9.19 the results from the application of the maximum precision model
are presented. Here two examples are included to show the duality between the
minimum cost model for reallocation and upgrade and the maximum precision
model.
Example 9.5
The reallocation and upgrading of instrumentation is illustrated by using

the heat exchanger network presented in Example 9.2. The existing process
instrumentation was provided in Table 9.11. The relocation bounds and the
maximum number of instruments per stream are included in Table 9.20. In
this example, two flowmeters and three thermocouples may be reallocated. The
allowable new positions for these instruments and the reallocation costs are
indicated in Table 9.21. The cost and standard deviation of new instrumentation
is given in Table 9.22.
Results are given in Tables 9.23, 9.24 and 9.25. For some cases, reallocation
is sufficient to fulfill precision requirements, and the solution contains only new
instruments for higher precision requirements. A mixed alternative of medium
cost (reallocation and new instrumentation) is achieved in other cases. Two
examples are provided to show the duality between both models.
TABLE 9.22. Heat Exchanger

Network: Available New
Instrumentation.
M1 Standard Deviation Cost
F4 2.5% 2250
F9 2.5% 2250
T2 0.2 1500
T6 0.2 1500
207
TABLE 9.23. Heat Exchanger Network: Minimum Cost Model Results
for the Reallocation and Upgrade Problem.
New
Case σU∗1 σU∗2 σU∗3 σU1 σU2 σU3 c Reallocations Instruments
1 4.00 4.00 4.00 3.2826 1.9254 2.2168 100 (T1 , 2, T6 ) —

2.3947 1.5482 2.6690 100 (T4 , 2, T2 ) —
2.0620 1.8277 2.6698 100 (T4 , 2, T6 ) —
3.6035 1.9122 2.2140 100 (T5 , 2, T6 ) —
2 2.00 2.00 2.00 — — — — — —
3 2.00 2.00 2.20 1.3891 1.5148 2.1935 3000 — T2 , T6
4 1.50 1.50 2.20 1.3492 1.3664 2.1125 5250 — F4 , T2 , T6
1.3839 1.4995 2.1606 5250 — F9 , T2 , T6
5 2.40 2.30 2.20 2.0587 1.8174 2.1938 1500 — T6
6 2.20 1.80 2.40 1.7890 1.6827 2.2014 1600 (T1 , 2, T2 ) T6
2.1503 1.5306 2.2031 1600 (T1 , 2, T6 ) T2
2.0980 1.6816 2.1993 1600 (T5 , 2, T2 ) T6
1.8644 1.5295 2.2025 1600 (T5 , 2, T6 ) T2
TABLE 9.24. Heat Exchanger Network: Maximum Precision Model Results

for the Reallocation and Upgrade Problem.
Case No. σU∗1 σU∗2 σU∗3 c∗ σU1 σU2 σU3
6 2.20 1.80 2.40 1600 1.8644 1.5295 2.2025

7 2.80 1.80 2.30 3000 1.3891 1.5148 2.1935
4 1.50 1.50 2.20 5000 — — —
4 1.50 1.50 2.20 5250 1.3492 1.3664 2.1125
TABLE 9.25. Heat Exchanger Network: Maximum Precision Model Results

for the Reallocation and Upgrade Problem (Continued).
P 2
ση i
Case Cost i Reallocations New Instruments
6 1600 10.6663 (T5 , 2, T6 ) T2

7 3000 9.0354 — T2 , T6
4 — — — —
4 5250 8.1503 — F4 , T2 , T6
208
References 209
NOMENCLATURE
ci,k : cost of the kth sensor for measuriun variable i

C p: heat capacity
FT : configuration correction factor for heat exchangers
H0 : measure of the estimation error
Ki : set of new sensors available for variable i
MR : set of variables where sensors from other places can be reallocated
MT : set of variables whose sensors can be reallocated
MU : set of key variables compulsory measured
Ni : number of existing sensors for variable i
Ni∗ : maximum number of sensors for variable i
e
Q: covariance matrix of estimates
q: binary vector indicating whether sensors are located (1) or not (0) in
variables
qN: binary vector indicating whether new sensors are located (1) or not (0)
in variables
Si : stream i
T: temperature
Ui : unit i
u t,k,r : binary variable indicating if sensor k from variable t is relocated to
measure variable r
Greek Letters
σ: precision
REFERENCES
Bagajewicz M. and M. Sánchez. Reallocation and Upgrade of Instrumentation in Process Plants.
Comp. & Chem. Eng. (2000).

Process Networks. AIChE J., 33, pp. 70–82 (1987).
Loeblein C. and J. D. Perkins. Economic Analysis of Different Structures of On-Line Process
Optimization Systems. Comp. & Chem. Eng., 22, 9., pp. 1257–1269 (1998).
Van Winkle M. Distillation. McGraw-Hill, New York (1967).
CHAPTER 10
Reliability of Nonrepairable
Sensor Networks
INTRODUCTION
In this chapter, the concepts of availability and reliability of sensor networks

are explored. A review of the concepts of service availability and reliability of
individual sensors is presented first.
Sensors fail, and they do so with a certain frequency. Because sensors are
used to estimate other variables through data reconciliation, when they fail,
the observability of some important key variables may be compromised. When
the degree of observability of these key variables is larger than one, the fail-
ure of one sensor may not compromise observability, but the accuracy of the
estimates always changes. This change may make the new standard deviation
larger than the original threshold (σ ∗ ). In this context, availability and reliability
are concepts that relate to the probability of a variable being estimable and are
independent of the loss of accuracy that takes place when some sensors fail.
SENSOR SERVICE AVAILABILITY
The probability of failure of sensors is related to many factors. These factors

are inherent to the sensor and are in principle independent (or mildly indepen-
dent) of external factors, such as weather, accidents etc. They manifest during
time, that is, a sensor can start in a good condition and a probability of its failure
must be connected to the amount of time the sensor has been in service. These
concepts are formalized next through the following definitions:
Definition: Sensor Service Availability Ais (t) is the probability that at time t,
the sensor i is normal, given that is good as new at time t = 0.
211
212 RELIABILITY OF NONREPAIRABLE SENSOR NETWORKS
Definition: Sensor Service Unavailability Uis (t) is the probability that the
sensor i is in failed state at time t, given that it jumped into the normal state at
time t = 0.
From the definitions, it is obvious that:
Ais (t) + Uis (t) = 1 (10.1)
SENSOR SERVICE RELIABILITY
Sensor service availability gives an indication of what is the state of a sensor

at a given point in time t. One may, however, be interested in determining the
probability of a sensor staying at a normal state for a given period of time. This
is called reliability and is defined as follows:
Definition: Sensor Service Reliability Ris (t) at time t, is the probability that
the sensor i is in the normal state in the interval [0, t], given that it was on the
normal state at time t = 0.
Note first that service availability is a probability expressed in terms of the
state of the sensor at time t, regardless of whether the sensor has failed at any
given time prior to that or not. Thus,
Ais (t) ≥ Ris (t) (10.2)
This difference will become apparent when corrective and preventive main-
tenance is analyzed. However, when no maintenance is performed, availability
and reliability are equal. Indeed, if a sensor has failed, it will stay in its failed
state, because in the absence of maintenance, no corrective action is taken. Then,
the probability of a sensor being at its normal state at time t is equal to the prob-
ability of the sensor being at the normal state during the whole interval [0, t].
Thus, the rest of this chapter concentrates on the concept of reliability only.
A typical reliability curve Ris (t) is shown in Figure 10.1. It is also called
survival distribution. It is characterized by a relatively sharp decline at the
beginning, a stabilization period and a declining last period. These curves are
constructed by simply starting with a large population of devices and monitoring
the amount of devices surviving at each time.
FAILURE DENSITY AND FAILURE RATE
Assume that a certain large number of identical sensors are put in operation at
time t = 0. The number of sensors failing in the interval [t, t + dt) divided by
the number of sensors that are at normal state is a measure of an intrinsic rate of
Failure Density and Failure Rate 213
FIGURE 10.1. Reliability or survival distribution.
failure of the system at that given point in time. This thought experiment can be
now translated into the realistic condition that one is analyzing one instrument
and not many. Thus:
Definition: Failure density f i (t) at time t is the derivative of the survival
distribution.
dRis (t)
f i (t) = (10.3)
dt
The failure density is, therefore, related to reliability. As shown in Figure 10.2,
there is a period of early failures, followed by a fairly constant failure rate period
and finally by a wear-out failure period. The maximum at the end is attributed
to the depletion of the sample.
However, one can relate this absolute rate to the existing sensors in the normal
state. Thus, failure rate is defined as follows:
Definition: The failure rate ri (t) at time t is the probability that the sensor
fails per unit time at time t given that it has survived until time t.
FIGURE 10.2. Failure density function.

FIGURE 10.3. Failure rate.
This rate is expressed in conditional terms by requesting that the sensors be at

normal state in the whole interval [0, t). In other words, when looked in terms
of the thought experiment of a population of sensors at normal state at time
zero, the failure rate is none other than the portion of sensors that fail calculated
as a fraction of those that have survived. Indeed,
f i (t)
ri (t) = (10.4)
Ris (t)
A typical form of this curve as a function of time is shown in Figure 10.3 and
is known as a bathtub curve. This curve indicates a first period of early failure
(burn-in period) followed by a constant called prime of life period where failures
are random and then, a final wear-out or burnout period phase in which it is as-
sumed that the component has reached its usability limits or simply its lifetime.
It is in the phase where the instruments have constant failure rate where one
desires the systems to function. Thus, all desired properties established as goals
of a design/upgrade procedure are typically based on a constant failure rate.
MARKOVIAN MODEL
In this section, a Markovian model of a system that has no repairs is presented.

Assume now the model shown in Figure 10.4.
The basis for Markovian models is that the derivative of the probability that
the system is in state k, Pk (t), at time t is given by:
dPk (t)
= inflow to state k − outflow from state k (10.5)
dt
Markovian Model 215
FIGURE 10.4. State transition.
In turn, the inflow to state k can be written as the sum over all other states of
the rate of transition from these other states to state k, X i,k , multiplied by the
probability of the system being in these other states.
X
inflow to state k = X i,k (t)Pi (t) (10.6)
i
Similarly, the outflow from state k can be written as the sum over all other
states of the rate of transition from state k to these other states, X k,i multiplied
by the probability of the system being in state k.
X
outflow from state k = X k,i (t)Pk (t) (10.7)
i
In our case, there are only two states, normal and failed. In addition, the flow
from failed state to the normal state is zero because there are no repairs, whereas
the outflow from the normal state to the state of failure is related to the failure
rate. Finally, the probability of the sensor of being in normal state is its service
availability and the rate of transition from normal to failed state is related to the
failure rate. Then the following differential equation follows:
dRis (t)
= −ri (t)Ris (t) (10.8)
dt
which, taking into account that Ris (0) = 0, has the following solution:
Rt
Ris (t) = e− 0 ri (t) dt
(10.9)
In the case of constant failure rate, the reliability and the availability of a
system without repairs becomes:
Ris (t) = Ais (t) = e−ri t (10.10)
When the failure rate is constant, the assumption is that the failure rate
is proportional to the number of sensors. It is, however, an optimistic as-
sumption. Assuming that the sensor failure is only proportional to the num-
ber of sensors in the normal state is assuming that there is no deterioration
of their parts through time. This is by no means true, but it has been con-
sidered so far a sufficiently appropriate assumption for the design of sensor
networks. In addition, Equation (10.10) does not take into account the burn-in
period.
Other models to calculate reliability exist. The reader is referred for exam-
ple to Henley and Kumamoto (1981) or Dhillon (1983) for further details.
However, all these other models are far more sophisticated than what has
been used for sensor network design. Thus, they are outside the scope of this
book.
MEAN TIME TO FAILURE
Because reliability is a function of time, a number representative of the

reliability that is independent of time is the mean time to failure. This is defined
as follows:
Z ∞
τi = Ris (t) dt (10.11)
0
In the case of constant failure intensity this is given by:
1
τi = (10.12)
ri
For example, a sensor that has a failure intensity of two sensors/year, the
mean time to failure is half a year.
In Table 10.1 some reliability data for instruments is provided. Ranges ex-
tracted from the chemical and petroleum industries, as well as from the nuclear
industry, are shown.
TABLE 10.1. Reliability Values

for Instruments.
Failure Rate
Instrument (Failures per 106 Hours)
Flow 0.1–10
Temperature 0.03–7
Pressure 1–8
Level 2–80
Source: Bloch and Geltner (1999).
Estimation Availability and Reliability of Variables 217
ESTIMATION AVAILABILITY AND RELIABILITY OF VARIABLES
The degree of estimability of a variable was defined in Chapter 5, and methods

to design sensor networks featuring this property were outlined in Chapter 6.
The degree of estimability is in many ways connected to reliability, because it
is directly defined as the minimum number of sensors that have to fail to make
a variable unobservable. Thus, the more sensors have to fail to make a variable
unobservable, the larger will be the probability that this will not happen. These
concepts are now extended to include probabilities making the connection with
the reliability of individual sensors.
Definition: Variable estimation availability Aiv (t) is the probability that
variable i can be estimated at time t using its own measurement (if any) or
other measurements, through balance equations, given that this estimation is
possible at time t = 0.
Definition: Variable estimation reliability Riv (t) is the probability that
variable i can be estimated at time t using its own measurement (if any) or
other measurements, through balance equations in the interval [0, t], given
that this estimation is possible at time t = 0.
Example 10.1
Consider as an example a simple flowsheet (Figure 10.5). Consider first the

case where x M = {S1 , S2 , S3 }. Thus, the only redundant equation of this system
is:
F1 = F2 + F3 (10.13)
The probability of being able to estimate F1 at time t is the estimation avail-

ability of S1 at time t, and the probability of being able to estimate F1 in the
interval (0, t), assuming all sensors are working at time zero is called estimation
reliability. Using probability theory, one can write: Av1 (t) = Pr{S1 is in normal
state OR S2 and S3 are in their normal state simultaneously} which suggests
that there are two ways of estimating the flow rate of stream S1 , by measuring it

directly, or by using the balance Equation (10.13) and the measurements of S2

and S3 . Using sets and probability theory, the estimation availability can then
be expressed as follows:
Av1 (t) = Pr{S1 ∪ (S2 ∩ S3 )} (10.14)
where the union of both sets that can estimate F1 represents the event that either
one set or the other can be in normal state, or both, whereas the intersection
represents the fact that both sets are in normal state.
Because the events are not mutually exclusive, one can use a well-known
result from elementary probability theory to obtain:
Av1 (t) = Pr{S1 } + Pr{S2 ∩ S3 } − Pr{S1 ∩ (S2 ∩ S3 )} (10.15)
where the intersection of both sets that can estimate F1 now represents the event
that both sets are in their normal state.
In turn, the probability of sensors S2 and S3 being in their normal state is
related to the sensors service availability as follows:
Pr{S2 ∩ S3 } = As2 (t)As3 (t) (10.16)
because both sensors should be active and they are independent. Therefore:
Av1 (t) = As1 (t) + As2 (t)As3 (t) − As1 (t)As2 (t)As3 (t) (10.17)
DETERMINATION OF ESTIMATION RELIABILITY
One can generalize the procedure shown in the previous motivating exam-
ple and construct a framework to determine the estimation availability of a
variable in any network. First, the expression in Equation (10.14) includes the
measurement of the variable of interest and the rest of the measurements of
the cutsets that includes the variable. To generalize Equation (10.17), the jth
reduced cutset Z j (i) = K j (i) − Si , which is the set of all elements of the jth
cutset containing variable i, K j (i)-except variable i, is defined. Thus,
© ª
Riv (t) = Pr Si ∪ Z 1 (i) ∪ Z 2 (i) ∪ . . . . . ∪ Z n k (i) (i) (10.18)
where n K (i) is the number of cutsets that contain variable i. In other words, the
estimation availability/reliability of variable i is related to the service availabil-
ity/reliability of a sensor measuring this variable and all the sensors through
which independent balances can be made. The reliability of variable i when it
is unmeasured is now introduced:
© ª
Riu (t) = Z 1 (i) ∪ Z 2 (i) ∪ . . . . . ∪ Z n k (i) (i) (10.19)
Determination of Estimation Reliability 219
Applying elementary probability theory to the union of two sets, as in

Equations (10.15) through (10.17) one obtains:
© ª
Riv (t) = Ris (t) + Riu (t) 1 − Ris (t) (10.20)
In turn, the reliability of variable i when unmeasured Riu (t) can be obtained
by expanding Equation (10.19)
nX nX

K (i) X
K (i) n K (i)

Riu (t) = Pr{Z s (i)} − Pr{Z s (i) ∩ Z j (i)} 





s=1 s=1 j>s 

nX X
K (i) n X
K (i) n K (i) (10.21)
+ Pr{Z s (i) ∩ Z j (i) ∩ Z t (i)} 



s=1 j>s t> j 

© ª


· · · + (−1)n K (i)−1 Pr Z s (i) ∩ · · · ∩ Z n K (i) (i)
which can now be further expanded by using the expression for each cutset:
Pr{K j (i)} = Rms 1 (t)Rms 2 (t) . . . Rms m (t) (10.22)
where Rms s (t) is the service reliability of sensor m s , which defines cutset K j (i).
There is yet other ways to calculate Riv (t). Several algorithms have been pub-
lished to evaluate Equation (10.18) (Veeraraghavan and Trivedi, 1991; Fong and
Buzacott, 1987; Tewari and Verma, 1980). We review the method based on the
so-called sum-of-disjoint products (Ali, 1993). We start with the following
equality:
Z 1 (i) ∪ Z 2 (i) . . . . . ∪ Z n K (i) (i) = Z 1 (i) ∪ Z 2∗ (i) ∪ . . . . Z n∗K (i) (i) (10.23)
where
Z ∗j (i) = Z 1 (i) ∩ Z 2 (i) ∩ . . . . ∩ Z j−1 (i) ∩ Z j (i) (10.24)
This equality is illustrated with three sets (Figure 10.6). The subset in gray
is indicated immediately below.
Note that the sets Z ∗j (i) are disjoint. This is trivial, as the last set in Z ∗j (i) is
active while the same set is inactive in Z ∗j+1 (i). Therefore, the following holds:
© ª
Pr Z 1 (i) ∪ Z 2 (i) . . . . . ∪ Z n K (i) (i) = Pr{Z 1 (i)} + Pr{Z 2∗ (i)} + · · · .
© ª
+ Pr Z n∗K (i) (i) (10.25)
FIGURE 10.6. Illustration of identity 10.24.
In turn,
Pr{Z ∗j (i)} = Pr{Z 1 (i)} Pr{Z 2 (i)} . . . . Pr{Z j−1 } Pr{Z j (i)} (10.26)
Finally:
© ª© ª © ª
Pr{Z j (i)} = 1 − Rmt
s
1
(t) 1 − Rms 2 (t) . . . . 1 − Rms m (t) (10.27)
Ali and Narasimhan (1995) compared the speed of this algorithm to others and
found it competitive.
Example 10.2
Consider the system of Figure 10.5 and assume x M = {S1 , S2 , S3 , S4 }. Note

first that there are two cutsets containing S1 and only measured variables,
namely:
K 1 (1) = {S1 , S2 , S3 } (10.28)

K 2 (1) = {S1 , S3 , S4 } (10.29)
Equation (10.21) renders:

)
Riu (t) = Pr{Z 1 (i)} + Pr{Z 2 (i)} − Pr{Z 1 (i)} Pr{Z 2 (i)}
(10.30)
= R2s (t)R3s (t) + R3s (t)R4s (t) − R2s (t)R3s (t)R3s (t)R4s (t)
The same equation can be written by using the sum of disjoint products:
R1Z (t) = Pr{Z 1 (1)} + Pr{Z 2*(1)} = Pr{Z 1 (1)} + Pr{Z *1(1)} Pr{Z 2 (1)}
£ ¤ (10.31)
= R2s (t)R3s (t) + 1 − R2s (t)R3s (t)}R3s (t)R4s (t)
which contains three multiplications and two additions/subtractions, contrasting

with Equation (10.30), which contains one multiplication more. Let us now add
Estimation Reliability in Nonredundant Systems 221
one more measurement in S6 . There is one new cutset containing S1 and only
measured variables:
K 3 (1) = {S1 , S6 } (10.32)
Thus, Equation (10.21) renders:


R1Z (t) = Pr{Z 1 (i)} + Pr{Z 2 (i)} + Pr{Z 3 (i)} − Pr{Z 1 (i)} Pr{Z 2 (i)} 



− Pr{Z 1 (i)} Pr{Z 3 (i)} − Pr{Z 2 (i)} Pr{Z 3 (i)} 



+ Pr{Z 1 (i)} Pr{Z 2 (i)} Pr{Z 3 (i)}


= R2s (t)R3s (t) + R3s (t)R4s (t) + R6s (t) − R2s (t)R3s (t)R3s (t)R4s (t) 



© ª2 s 

− R2 (t)R3 (t)R6 (t) − R3 (t)R4 (t)R6 (t) + R2 (t) R3 (t) R4 (t)R6 (t)
s s s s s s s s s
(10.33)
whereas the sum of disjoint products gives:

R1Z (t) = Pr{Z 1 (1)} + Pr{Z 2*(1)} = Pr{Z 1 (1)} + Pr{Z *1(1)} Pr{Z 2 (1)}



+ Pr{Z *1(1)} Pr{Z *2(1)} Pr{Z 3 (1)} = R2s (t)R3s (t) 
© ª © ª
+ 1 − R2s (t)R3s (t) R3s (t)R4s (t) + 1 − R2s (t)R3s (t) 

© ª s 


× 1 − R3 (t)R4 (t) R6 (t)
s s
(10.34)
Equation (10.33) contains 6 additions/subtractions and 13 multiplications,
whereas Equation (10.34) contains only 5 additions/subtractions and 8 multi-
plications. It is clear that the sum of disjoint products renders fewer multiplica-
tions, which for systems with many cutsets can become a factor in computation
time.
ESTIMATION RELIABILITY IN NONREDUNDANT SYSTEMS
In nonredundant systems, the number of cutsets determining an unmeasured

variable is reduced to one. In other words
Riv (t) = Ris (t) (10.35)
for measured variables and
Riv (t) = Rms 1 (t)Rms 2 (t) . . . . Rms m (t) (10.36)
for unmeasured variables. The service reliabilities in Equation (10.36) corre-

spond to the sensors located in the chords of the spanning tree that are part of
the only cutset that contains variable i.
AVAILABILITY, RELIABILITY AND DEGREE OF ESTIMABILITY
The availability of estimating a variable encompasses the concepts of ob-

servability and redundancy as given by the following properties:
(1) A variable is observable if and only if its estimation availability/reliability

is greater than zero.
(2) A measured variable is redundant if and only if its estimation availabil-
ity/reliability is greater than the service availability/reliability of the sensor
measuring it.
Ali and Narasimhan (1993) chose to present the above connection to observ-
ability in terms of the reliability of variables. Because estimation availability and
reliability are the same for nonrepairable systems, the statements are equivalent.
However, for repairable systems, availability is still nonzero after a failure is
followed by a repair, whereas reliability remains zero. This difference changes
the goals and procedures for the design of repairable systems, as covered in
Chapter 14.
The difficulty in establishing numerical analogies between estimation avail-
ability of variables and degree of estimability stems from the fact that the
estimation availability is always smaller than one, and, therefore, even though
it is monotone with the number of cutsets contributing, the monotonicity is far
from being linear.
SYSTEM AVAILABILITY AND RELIABILITY
Hitherto, the service and estimation availability and reliability of variables

have been defined. There is, however, a need to define a unique measure that
will quantify in one single number the reliability of the different variables.
Using the notion that a chain is no stronger than its weakest link, Ali and
Narasimhan (1993) defined the system estimation reliability as the minimum
estimation reliability among all variables. Their definition is slightly modified
by restricting the minimization to a set of variables M R where reliability is of
interest. Thus, the reliability of the system is given by the smallest reliability
of their components.
R v (t) = Min Riv (t) (10.37)

∀i∈M R
In other words, the reliability of the system is defined by its weakest element.
This measure of the system reliability can be repeated exactly in the same way
to define the system estimation availability. Ali and Narasimhan (1993, 1995)
Nomenclature 223
use this measure as the basis for their sensor network design procedures
(Chapter 11).
Other definitions are possible. For example, one can use averages of relia-
bilities or even weighted averages of some sort. Instead of averages, the prod-
uct of all reliabilities of the components has been used for systems in series
(Dhillon, 1983)
Y
R v (t) = Riv (t) (10.38)
∀i∈M R
This expression reflects the fact that it is enough that one unit fails for the
system to fail. Systems in parallel follow other reliability expressions reflecting
the fact that all elements in parallel need to fail for the system to fail. The reader is
referred to Henley and Kumamoto (1981) or Dhillon (1983) for various forms of
defining the system reliability. Because they have not been applied to the field
of sensor design yet, its coverage is omitted. Measures based on mean-time
calculations of some sort have also been suggested for specific sensor design
case studies (Maquin et al., 1994), but no generalizations for generic systems
have been made.
NOMENCLATURE
Ais (t): sensor service availability

Aiv (t): variable estimation availability
f i (t): failure density
Pk (t): probability that the system is in state k, at time t
ri (t): failure rate
Riv (t): variable estimation reliability
Ris (t): sensor service reliability
R v (t): system reliability
Si : stream i
Ui : unit i
Uis (t): sensor service unavailability
X i,k (t): rate of transition from state i to state k
Z j (i): reduced cutset
Greek Letters
τi : mean time to failure of sensor i
REFERENCES
Ali Y. Sensor Network Design for Maximizing Reliability of Processes. Ph.D. Thesis, Indian Institute
of Technology. Kanpur, India (1993).
AIChE J., 39, 5, pp. 2237–2249 (1993).
Ali Y. and S. Narasimhan. Redundant Sensor Network Design for Linear Processes. AIChE J., 41,
10 (1995).
Bloch H. P. And F. K. Geltner. Use Equipment Failure Statistics Properly. Hydrocarbon Processing,
January (1999)
Dhillon B. S. Reliability Engineering in Systems Design and Operation. Van Nostrand Reinholdt
Company, New York (1983).
Fong C. and J. Buzacott. An Algorithm for Symbolic Reliability Computation with Pathsets and
Cutsets. IEEE Trans. Rel., R-36, 34 (1987).
Henley E. J. and H. Kumamoto. Reliability Engineering and Risk Assessment. Prentice Hall, New
Jersey (1981).
Maquin D., M. Luong and J. Ragot. Observability Analysis and Sensor Placement. SAFE
PROCESS ‘94 IFAC/IMACS Symposium on Fault Detection, Supervision and Safety for
Technical Process, June 13–15, Espoo, Finland, (1994).
Tewari R. K. and M. Verma. An Algebraic Technique for Reliability Evaluation. IEEE Trans. Rel.,
R-29, pp. 311 (1980).
Veeraraghavan M. and K. S. Trivedi. An Improved Algorithm for the Symbolic Reliability Analysis
of Networks. IEEE Trans. Rel., R-40, pp. 347 (1991).
CHAPTER 11
Design of Reliable Linear

Nonrepairable Sensor Networks
INTRODUCTION
This chapter is devoted to present methodologies for the design of reliable

sensor networks without considering requirements of accuracy. Estimability
goals are also omitted, but they are somehow connected to reliability. Indeed,
increasing the degree of estimability of a variable is equivalent to increasing
the number of ways the variable is calculated. The estimation reliability of
a variable has the same monotonic behavior, but, unfortunately, there is no
one-to-one relation between both so that they cannot be considered equivalent.
Therefore, although estimability goals can be useful when the service availabil-
ity of sensors is not known, the design using estimability goals can be more
costly than one based on hard instrument reliability data.
NONREDUNDANT NETWORKS FEATURING

MAXIMUM RELIABILITY
Ali and Narasimhan (1993) proposed to use the system reliability given by
Equation (10.37) as the objective function for the sensor network design. They
extended both the set of variables for which the reliability is of interest (M R )
and the set of streams where sensors can be placed (M1 ) to the whole system.
Thus, unless some other constraints are imposed, maximizing the reliability
of the system may result in a solution where all the variables are measured.
This will occur because the reliability of a variable increases monotonically as
measured cutsets containing it are added. Therefore, some restriction on the
number of measurements is needed to control the cost.
225
226 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
To address the issue of cost, Ali and Narasimhan (1993) proposed to limit
the number of sensors to the minimum possible that will still guarantee observ-
ability. As shown in Chapter 5, systems that feature all nonmeasured variables
being observable (degree of estimability one) and minimum number of sen-
sors, correspond to measurements located in the chords of a spanning tree of
the systems. Based on this fact, all that is needed is to obtain a spanning tree
conducive to maximized reliability.
Motivating Example 11.1
Consider the simplified ammonia network (Kretsovalis and Mah, 1988a) that
consists of six nodes and eight edges, with node U6 representing the environ-
mental node, as shown in Figure 11.1(a) and (b).
The minimum number of sensors required for this process is three, which
is the number of chords of any spanning tree of the graph in Figure 11.1(b).
Assume that sensors can be placed on any stream and each of them has a failure
probability of 0.1. Let us examine the following two cases:
a. Case 1. The mass flows of streams S1 , S4 and S7 are measured. The mass flow
of stream S6 can be estimated using the fundamental cutset {S1 , S4 , S6 , S7 },
which gives a reliability of 0.729.
b. Case 2. The mass flows of streams S4 , S5 and S7 are measured. The mass flow
of stream S6 is now estimated through the fundamental cutset {S5 , S6 , S7 }
giving a reliability of 0.81, which is higher than that for Case 1.
The above example illustrates the type of relation that exists between sensor
placement and reliability.
FIGURE 11.1. Simplified ammonia plant network.

Nonredundant Networks Featuring Maximum Reliability 227
Foundations of the Design Strategy
To solve the problem avoiding enumeration, Ali and Narasimhan (1993)

presented a strategy where a spanning tree is chosen first and, in a second step,
one sensor is removed at a time being replaced by another sensor, such that:
a. A new spanning tree solution is obtained with the new sensor.
b. The reliability of the new spanning tree solution is higher.
To switch from one spanning tree to another, a measurement placed in a chord
of a spanning tree that is being removed should be replaced by a measurement
in a branch such that both belong to the same fundamental cutset.
Before the actual algorithm is presented, a few definitions and lemmas ex-
tracted from Ali and Narasimhan (1993) are presented.
Definition: The ring sum of two cutsets K 1 and K 2 is the set of all edges
(streams) that belong to either cutset K 1 or K 2 but are not in both.
For example, consider the two cutsets defined in the previous motivating
example. K 1 = {S1 , S4 , S6 , S7 }, and K 2 = {S5 , S6 , S7 }. The rig sum of these
cutsets is K 1 ⊕ K 2 = {S1 , S4 , S5 }. We now present the following important
lemma (Ali and Narasimhan, 1993):
Lemma 11.1: The ring sum of two fundamental cutsets, which have at least
one common chord of the associated spanning tree gives a cutset of a graph.
Before proving the lemma we illustrate it using the above ammonia network.
Assume the spanning tree with chords {S2 , S5 , S6 }. Consider now the follow-
ing two cutsets: K 1 (S8 ) = {S2 , S6 , S8 } and K 1 (S4 ) = {S2 , S5 , S4 }. Figure 11.2
shows the ring sum of these two cutsets. The dashed lines correspond to the
branches of the spanning tree, and the curved lines indicate how the cutsets
separate the graph into two disjoint graphs. The lemma is now formally proved.
FIGURE 11.2. Ring sum of two cutsets with one chord in common.
Proof: Consider a spanning tree T with branches given by the set {b1 , . . . ,
bn−l } and chords given by the set {c1 , . . . , ce−n+l }. Without loss of generality,
let the fundamental cutsets K 1 (bx ) and K 1 (b y ) be defined as:
K 1 (bx ) = {bx , c1 , . . . , cr , cr +1 , . . . , cs } (11.1)

K 1 (b y ) = {b y , c1 , . . . , cr , cs+l , . . . , ct } (11.2)
where chords {c1 , . . . , cr } are common to both K 1 (bx ) and K 1 (b y ). The ring
sum of K 1 (bx ) and K 1 (b y ) is the set of all edges in K 1 (bx ) and K 1 (b y ) excluding
the common ones. Thus,
K 1 (bx ) ⊕ K 1 (b y ) = {bx , cr +1 , . . . , cs , b y , cs+1 , . . . , ct } (11.3)
To prove that the ring sum is a cutset, it is only required to show that if a
proper subset of edges from the above set is deleted, it does not disconnect the
graph and only the deletion of the whole set does. Deletion of the set of chords
{cr +l , . . . , cs , cs+1 , . . . , ct } or any subset of it will not disconnect the graph,
because the branches of T still exist and they maintain connectivity. Similarly,
deletion of the set {bx , cr +1 , . . . , ct } will not disconnect the graph because the
common edges {cl , . . . , cr } are subsets of K 1 (bx ) and K 1 (b y ); therefore, they
preserve connectivity between the nodes connected by the deleted set. To better
visualize this, consider the nodes of the graph divided into three subgraphs: G 1 ,
G 2 and G 3 , such that
r G 1 and G 2 are connected by the set {bx , cr +1 , . . . , cs }
r G 2 and G 3 are connected by the set {b y , cs+1 , . . . , ct }
r G 1 and G 3 are connected by the set {c1 , . . . , cr }
This is shown in Figure 11.3.
We illustrate this by using the ammonia network and the cutsets used previ-
ously, that is K 1 = {S1 , S9 , S6 , S7 } and K 2 = {S5 , S6 , S7 }. The common set of
chords is {cl , . . . , cr } = {S2 }, bx = S8 , bx = S4 and the chords not in common
FIGURE 11.3. Illustration of the proof of Lemma 11.1.

are given by {cr +1 , . . . , cs } = {S6 } and {cs+1 , . . . , ct } = {S5 }, respectively. As

illustrated in Figure 11.4, the system preserves connectivity through stream {S2 }
when the set {bx , cr +1 , . . . , ct } = {S8 , S5 , S6 } is deleted.
A similar argument can be made for the deletion of {b y , cs+1 , . . . , ct }. Thus,
it is only when bx and b y and all the chords are deleted that the graph is
disconnected. Hence, the ring sum forms another cutset and not a union of edge
disjoint cutsets. Q.E.D.
Remark: A ring sum of two fundamental cutsets does not always give a
cutset. When they do not have a chord in common, their ring sum renders a
union of edge-disjoint cutsets.
Remark: The property given by Lemma 11.1 is useful in generating the
fundamental cutsets of a new spanning tree obtained through an elementary
tree transformation.
We now present two other important lemmas, proposed and proved by Ali
and Narasimhan (1993).
Lemma 11.2: For any sensor network that has measured variables in the
chords of a spanning tree, the variable that has the lowest availability/reliability
is always an unmeasured variable.
Proof: Consider the measurement with lowest sensor service availability
Asm (t). This will be a chord of the spanning tree corresponding to the sensor
network design. Because every chord appears in some fundamental cutset, let
chord Sm be a member of fundamental cutset K 1 (Si ), which is a cutset that
includes branch Si of the spanning tree, and one or more additional chords.
Thus,
Y
Aiv (t) = Asj (t) ≤ Asm (t) (11.4)
j∈K 1 (i)
In other words, either Si has the lowest reliability or some other branch does.
In any case, the least reliability is attained for an unmeasured variable. Q.E.D.
Remark: This lemma shows that to improve the network reliability of a

given sensor network, the reliability of some unmeasured variable needs to be
improved.
Lemma 11.3: Let T be a spanning tree with branch bx having the minimum
reliability. If the failure probabilities of all sensors are equal, then the network
reliability can be improved by placing a sensor on branch bq (leaving vari-
able) and removing the sensor from chord c p (entering variable) provided the
following conditions hold:
(1) The chord c p is part of fundamental cutsets containing bx and bq . That is,
c p ∈ K 1 (bq ) and c p ∈ K 1 (bx ),
(2) The cardinality of the ring sum of both cutsets K 1 (bx ) ⊕ K 1 (bq ) is smaller
than the cardinality of K 1 (bx ), that is, 0{K 1 (bx ) ⊕ K 1 (bq )} < 0{K 1 (bx )}
(3) If c p ∈ K 1 (bs ), where K 1 (bs ) is any other fundamental cutset, then
0{K 1 (bs ) ⊕ K 1 (bq )} < 0{K 1 (bx )}.
Proof: By placing a sensor on bq and removing the sensor from chord c p

another spanning tree solution is obtained. Let the new spanning tree be T and
let K be the set of fundamental cutsets corresponding to T . The objective is to
prove that reliabilities of all variables in the new solution are larger than R(bx ).
Thus, if
K 1 (bx ) = {bx , c p , c1 , . . . , cr , cr +1 , . . . , cs } (11.5)

K 1 (bq ) = {bq , c p , c1 , . . . , cr , cs+l , . . . , ct } (11.6)
then
K 1 (bx ) ⊕ K 1 (bq ) = {bx , cr +1 , . . . , cs , bq , cs+1 , . . . , ct } (11.7)
The ring sum K 1 (bx ) ⊕ K 1 (bq ) is a cutset of the graph (Lemma 11.1). Further-
more, in the new spanning tree solution, the cutset K 1 (bx ) ⊕ K 1 (bq ) contains
only one unmeasured variable, which is bx (See Figure 11.5). Condition (1) of
this lemma ensures that the new unmeasured variable c p is not a member of
K 1 (bx ) ⊕ K 1 (bq ). Based on these observations K 1 (bx ) ⊕ K 1 (bq ) is a funda-
mental cutset with respect to spanning tree T containing branch bx . Thus,
K 1 (bx ) = K 1 (bx ) ⊕ K 1 (bq ) (11.8)

From condition 2 and the assumption that all sensors have same failure
probability it follows that:
v
R (bx ) > R v (bx ) (11.9)
v
where R (bx ) represents the reliability of bx with respect to the spanning tree
T.
The new unmeasured variable c p belongs to the same cutset as bq , that is
K 1 (bq ). Then,
v
K 1 (c p ) = K 1 (bq ) ⇒ R (c p ) = R v (bq ) > R v (bx ) (11.10)
The last inequality follows from the original assumption that bx is the variable
of minimum reliability. Finally, if cp is a member of some other fundamental
cutset K 1 (bs ), then
K 1 (bs ) = K 1 (bs ) ⊕ K 1 (bq ) (11.11)
From condition 3, it follows that:

v
R (bs ) > R v (bx ) (11.12)
On the other hand, if c p is not a member of cutset K 1 (bs ), then the reliability
of bs remains unchanged because K 1 (bs ) = K 1 (bs ). Thus, the reliability of each
of all variables corresponding to T is strictly larger than the minimum reliability
corresponding to T . Therefore,
v
Min R (bi ) > R(bx ) (11.13)
i
proving that the network reliability has increased. Q.E.D.

Remarks: The following observations with respect to Lemma 11.3 can be

made:
(1) The second condition ensures that the reliability of variable bx increases,
whereas the third condition ensures that the reliabilities of all other unmea-
sured variables remain greater than the current network reliability.
(2) The network reliability cannot be improved by placing a sensor on variable
bx . Indeed, to maintain a spanning tree solution, the sensor of some chord
of K 1 (bx ) should be removed. The reliability of this variable in the new
solution will then be equal to R(bx ), and thus the network reliability would
not change.
(3) If sensor failure probabilities are not equal, then the entering and leaving
variables can be chosen in a similar manner, except that instead of checking
for the cardinality conditions 2 and 3, one must explicitly evaluate the
reliabilities.
(4) The converse of Lemma 11.3 is not true, that is, if a branch bq and chord c p
satisfying the three conditions are not found, it does not imply that the global
optimum solution has been obtained. However, it can be viewed as a local
optimum. Moreover, if conditions 2 and 3 do not hold as strict inequalities
or if there are two or more variables with minimum reliability, then a
degenerate solution is obtained, that is, the network reliability remains the
same, and this may be true for other solutions as well. Both these problems
are handled using heuristic strategies.
The algorithm for the sensor network design procedure is described next.
Algorithm
The algorithm developed by Ali and Narasimhan (1993) is called SENNET.

A flowchart is given in Figure 11.6.
r Step 1. Generate a spanning tree of the process graph.
r Step 2. Generate all the fundamental cutsets corresponding to the spanning
tree.
r Step 3. Obtain K max the set of fundamental cutsets that have the maximum
cardinality. The branches corresponding to these fundamental cutsets are the
variables with minimum reliability.
r Step 4. Choose one element of K max for example K i (bx ), which has not been
examined before. Mark K i (bx ) as examined and go to step 5. If no unmarked
fundamental cutset exists, go to step 10.
r Step 5. Choose another fundamental cutset K j (bq ) that has not been exam-
ined before. Obtain the ring sum K i (bx ) ⊕ K j (bq ). Mark K j (bq ) as examined
and go to step 6. If no such K j (bq ) exists, go to step 4.
FIGURE 11.6. Flowchart of algorithm SENNET.
233
r Step 6. If 0{K i (bx ) ⊕ K j (bq )} ≤ 0{K i (bx )} go to step 7. Otherwise, go to

step 5.
r Step 7. Choose a chord c p . To do this, compute the set of all elements of
K i (bx ) not present in K i (bx ) ⊕ K j (bq ). This set is denoted as K i (bx ) −
{K i (bx ) ⊕ K j (bq )}. Mark one element (chord) c p as examined and go to
step 8. If no such chord exists, go to step 5.
r Step 8. If all fundamental cutsets K m (bs ) containing chord c p satisfy
0{K m (bs ) ⊕ K j (bq )} ≤ 0{K i (bx )}, then go to step 9. Otherwise, go to step 7.
r Step 9. Branch bq is selected as the new measured variable and chord c p is
chosen to enter the spanning tree, that is, to become unmeasured. Update all
fundamental cutsets to obtain the fundamental cutsets of the new tree, and go
to step 3.
r Step 10. Stop if this step, called “hill climbing,” has already been executed
10 times. Otherwise, store the current solution. Let branch bq correspond to
the most recently examined cutset K j (bq ) and c p any arbitrary element of
the set K i (bx ) − {K i (bx ) ⊕ K j (bq )}. Select bq as the leaving variable. Go to
step 3.
Handling Degeneracy
Note that in the algorithm, the inequalities are weak inequalities and may lead
to successive solutions that do not improve the network reliability (degeneracy).
In fact, it is theoretically possible to be caught in an infinite loop. Degeneracy
is, therefore, broken by random selection of a set K i (bx ) from set K max in step 4
and random selection of chords c p from set K i (bx ) − {K i (bx ) ⊕ K j (bq )} in
step 7. Ali and Narasimhan (1993) claim that this heuristic method has worked
well with the problems they tested.
Hill Climbing Procedure
Step 10 in the above algorithm is implemented as an attempt to get away

from a local minimum and reach the global optimum solution. If a situation in
which any choice of entering/leaving variable worsens the network reliability
is reached, then the current best solution is perturbed allowing the network
reliability to decrease and the algorithm is repeated. This is similar to “hill
climbing” strategies used in optimization.
Updating the Fundamental Cutsets
In step 9, the fundamental cutsets of the new spanning tree can simply be
obtained by updating the fundamental cutsets of the previous spanning tree as
follows. The fundamental cutset of the entering variable c p is given by K j (bq ).
If a fundamental cutset K m (bs ) contains c p as a member, then the updated
fundamental cutset is given by the ring sum of K m (bs ) and K j (bq ). All other
fundamental cutsets remain unaltered.
Example 11.2
The algorithm is illustrated by designing a sensor network for the simplified

ammonia plant shown in Figure 11.1. As established, the minimum number of
sensors required for this plant is three. Let the sensor failure probabilities for
all edges be 0.1.
r Step 1. Let the initial spanning tree solution consist of the unmeasured edges
(branches) S2 , S3 , S5 , S6 and S8 .
r Step 2. The fundamental cutsets are:
K 1 (S8 ) = {S4 , S7 , S8 }
K 1 (S6 ) = {S1 , S4 , S6 , S7 }
K 1 (S5 ) = {S1 , S4 , S5 }
K 1 (S2 ) = {S1 , S2 }
K 1 (S3 ) = {S1 , S3 }
r Steps 3 and 4. The set K max contains only one element, the maximum cardi-
nality cutset K 1 (S6 ) in which variable S6 is determined by using three mea-
surements. This is the variable that has the minimum reliability, and we at-
tempt to improve its reliability. Thus, bx = S6 , and K 1 (bx ) = {S1 , S4 , S6 , S7 }.
r Steps 5 and 6. The ring sum of K 1 (bx ) with other fundamental cutsets gives
the following cutsets:
K 1 (S8 ) ⊕ K 1 (bx ) = {S1 , S6 , S8 } Unmeasured: {S6 , S8 }

K 1 (S5 ) ⊕ K 1 (bx ) = {S5 , S6 , S7 } Unmeasured: {S5 , S6 }
K 1 (S2 ) ⊕ K 1 (bx ) = {S2 , S4 , S6 , S7 } Unmeasured: {S2 , S6 }
K 1 (S3 ) ⊕ K 1 (bx ) = {S3 , S4 , S6 , S7 } Unmeasured: {S3 , S6 }
Among the above cutsets, only the first two have cardinality smaller than
that of K 1 (bx ). The first is arbitrarily chosen. Thus, bq = S8 and K 1 (bq ) =
{S4 , S7 , S8 }. Therefore, K 1 (bx ) ⊕ K 1 (bq ) = {S1 , S6 , S8 }.
r Steps 7 and 8. We now compute the set of all elements of K 1 (bx ) not present
in {K 1 (bx ) ⊕ K 1 (bq )}, that is: K 1 (bx ) − {K 1 (bx ) ⊕ K 1 (bq )} = {S4 , S7 }. Any
of the chords in this set may be chosen. Thus, chord S4 is selected, that is,
c p = S4 .
r Step 9. The branch bq = S8 leaves the tree and becomes measured, whereas
chord S4 enters the tree and becomes unmeasured. The new spanning tree
obtained through the elementary tree transformation is {S2 , S3 , S4 , S5 , S6 }.
The following fundamental cutsets for this new spanning tree are
K 1 (S4 ) = {S4 , S7 , S8 }
K 2 (S6 ) = {S1 , S6 , S8 }
K 2 (S5 ) = {S1 , S5 , S7 , S8 }
K 2 (S2 ) = {S1 , S2 }
K 2 (S3 ) = {S1 , S3 }
Note that the reliability of variable S6 has improved. However, the net-
work reliability remains the same, because cutset K 2 (S5 ) contains three chords
(S1 , S7 , S8 ). This is due to the fact that for this cutset, condition 3 of Lemma 11.3
holds only as a weak inequality.
One can proceed with the next iteration in which S6 is the leaving variable and
S1 is the entering variable to give the spanning tree solution {S1 , S2 , S3 , S4 , S5 }.
In the subsequent iteration, there is no choice of entering variables and a local
optimum solution is obtained. The hill climbing technique is applied next.
The network reliability does not improve in 10 successive attempts, and the
solution {S1 , S2 , S3 , S4 , S5 } is adopted as optimal. This solution corresponds to
x M = {S6 , S7 , S8 }.
To compare this solution with the global optimum, Ali and Narasimhan
implemented an algorithm for explicit enumeration of all spanning trees and
obtained the best solutions. The algorithm used was taken from Nijenhuis and
Wilf (1978). The process graph has 32 spanning trees of which only 8 are
globally optimal, giving a network reliability of 0.81. The optimal solutions
generated by the algorithm for five different initial starting solutions are shown
in Table 11.1. All the solutions obtained by the algorithm are also globally
optimal, thus indicating that the algorithm is robust, converging for each choice
of the initial solution.
A larger example based on the steam-metering network of a methanol plant
(Serth and Heenan, 1986) was also presented by Ali and Narasimhan (1993).
TABLE 11.1. Initial Data and Results

of Ammonia Plant Design.
Initial Solution Optimal Solution
{S2 , S3 , S5 , S7 , S8 } {S1 , S2 , S3 , S4 , S5 }
{S2 , S3 , S5 , S6 , S7 } {S2 , S3 , S4 , S7 , S8 }
{S2 , S3 , S4 , S6 , S7 } {S1 , S2 , S4 , S7 , S8 }
{S1 , S3 , S4 , S5 , S6 } {S1 , S2 , S4 , S7 , S8 }
{S1 , S2 , S3 , S5 , S8 } {S1 , S2 , S3 , S6 , S8 }
This example consists of 12 nodes (units) and 28 edges (streams) and will be
used later in this chapter to illustrate other methods (Figure 11.9). In this case,
there are over 1.2 million spanning trees of which only 0.01% are globally
optimal. Ali and Narasimhan (1993) also attempted to solve the problem for
unequal sensor service reliabilities reporting the need for hill climbing.
REDUNDANT NETWORKS FEATURING MAXIMUM RELIABILITY
The procedure presented in the previous section provides solutions for nonre-
dundant networks with a degree of estimability of one for all variables, in other
words, for minimal networks (Chapter 5). However, redundant systems have
several desired properties. First, one can perform data reconciliation, and, there-
fore, the accuracy of the estimates obtained is improved. Second, the estimation
reliability of any redundant variable is higher, because more cutsets are instru-
mented. Finally, redundant systems are the only ones capable of detecting gross
errors (biased instruments and/or leaks).
To address redundant networks, Ali and Narasimhan (1995) presented a
model where a fixed number of sensors larger than the minimum is used. The
model relies on
a. Choosing first a spanning tree with the measured variables located in its
chords. The remaining measured variables are added to some of the branches
of this spanning tree.
b. An evolutionary strategy, based on choosing entering and leaving variables,
as in the case of nonredundant networks, is used. As in the case of nonredun-
dant networks, the leaving variables are unmeasured variables that become
measured and vice versa; the entering variables are measured variables that
become unmeasured.
We recall that in SENNET, the entering and leaving variables are chosen in
such way that a spanning tree structure is maintained. For a redundant network,
however, the only condition that needs to be met is that the estimability of
all variables is not zero. In other words, unmeasured variables should always
be observable. In addition, in SENNET, the reliability is computed using the
fundamental cutsets. It was shown in Chapter 5 that all the estimable cutsets
containing variable Si correspond to all the indirect ways that the variable can
be estimated. Thus, in redundant networks, all the estimable cutsets containing
the variable of interest should be used to compute the reliability. Finally, in
contrast with nonredundant networks, where the lowest reliability corresponds
to an unmeasured variable, the lowest reliability could correspond either to a
measured variable or to an unmeasured variable.
Determination of the Leaving Variable
Leaving variables are, as defined earlier, unmeasured variables that become

measured. Let Si be the variable having the least reliability in some iteration.
We consider two cases:
(1) Si is unmeasured. In this case, the estimable cutsets, which are the only
cutsets contributing to the reliability of Si , contain all measured variables.
Thus there are two ways of improving the reliability of Si : making Si the
leaving variable or finding a cutset that contains Si and only one other
unmeasured variable St so that St is chosen as leaving variable. The latter
is not possible in the case of the type of redundant networks analyzed
here because all variables are observable by construction and, therefore, all
cutsets are estimable.
(2) Si is measured. Then, the leaving variable should be such that by making
it measured, a new estimable cutset containing Si is formed. This is ac-
complished by choosing the leaving variable from a cutset that does not
contain any other unmeasured edge besides the leaving variable; otherwise,
the introduction of a measurement in the leaving variable would not have
any effect.
These observations prompt the introduction of some notation and suggest a
procedure for selecting the leaving variable. We first introduce the notation.
Definition: An almost redundant cutset of order one containing variable
Si , denoted K s(1) (St , Si ), is an estimable cutset containing Si and only one
unmeasured variable St .
Remark: Si can be unmeasured, in which case the cutset is denoted
K s(1) (Si , Si ).
Definition: The set of all almost redundant cutsets of order one containing
Si is denoted by 8(1) (Si ).
Thus, if Si is measured, variable St belonging to K (1) j (St , Si ) can be chosen
as a leaving variable candidate. The number of ways of indirectly estimating
Si may increase by more than one if the cardinality of 8(1) (Si ) is larger than
one. If there are many candidates for the leaving variable, the one that gives
the maximum number of additional ways of indirectly estimating Si should be
selected first.
If Si is an unmeasured variable, then it can itself be a leaving variable candi-
date. In the algorithm implementation presented later, if Si is unmeasured, it is
selected as the first choice of leaving variable. Thus, the set of leaving variable
candidates, Lc is defined by
© ª
Lc(Si ) = St | St ∈ K (1)
j (St , Si ) (11.14)
Determination of the Entering Variable
To keep the number of sensors fixed, once a leaving variable has been chosen,
a sensor from a measured variable has to be removed. The entering variable
should, therefore, be chosen so that the network reliability does not decrease.
Because there are multiple ways of estimating a variable, it is not possible to
predict which sensor removal will meet this condition. Explicit enumeration,
where each sensor is removed at a time and the reliabilities are computed is,
therefore, the first option. The following observations are used to reduce the
search space and, hence, decrease the computational burden.
(1) The objective is to improve the reliability of the least reliable variable Si .
Because the leaving variable is chosen with the objective of providing at
least one additional way for estimating Si , if the entering variable is a
member of any cutset belonging to 8(1) (Si ), then it would nullify this gain.
This ensures that there will be at least one additional new way of indirectly
estimating variable Si in the new sensor network. The set of entering variable
candidates is, therefore, given by
\
Le = x M − K s(1) (St , Si ) (11.15)
s∈8(1) (Si )
(2) As pointed out earlier, if the number of ways of estimating a variable in-
creases, then the reliability of the variable also increases. Furthermore, if
the cardinality of a cutset containing variable Si is high, then its contribution
to the reliability of Si is small. In other words, cutsets with large cardinality
should be avoided and variables belonging to such cutsets should be chosen
as entering variables, that is, to become unmeasured. However, variables
belong to different cutsets at the same time, a fact that makes the selection
difficult.
The above observations prompt a few definitions and the development of a
measure to rank the entering candidates. We present the definitions first:
Definition: An almost redundant cutset of order two containing variables
Si and measured variable S j denoted K s(2) (St , Si , S j ) is an estimable cutset
containing Si , S j and only one unmeasured variable St .
Remark: Variable Si could be unmeasured in which case St = Si .
Definition: The set of all almost redundant cutsets of order two containing
Si and S j is denoted by 8(2) (Si , S j ).
We are trying to define a quantity Ä(Si , S j ) representing the potential of
measured variable S j to increase the reliability of Si . To intuitively introduce
such measure, suppose two candidate variables, each participating in only one
cutset with Si . Suppose further that these cutsets have a different number of
elements. Removal of any such candidate lowers the reliability of Si , an effect
that should be countered by the leaving variable. Thus, because low cardinality
implies high reliability, we should choose as entering variable the one with
lower cardinality. Thus, a high value of Ä(Si , S j ) will mean that the removal of
S j , that is, making it unmeasured, will have a small effect on the reliability of
Si . However, variables can be part of more than one cutset. Thus, Ä(Si , S j ) is
defined as the ratio of the number of all the variables participating in all almost
redundant K s(2) (St , Si , S j ) divided by the number of the cutsets involved. In other
words, Ä(Si , S j ) is given by the sum of all the cardinalities of K s(2) (St , Si , S j )
divided by the cardinality of 8(2) (Si ,S j ) (Ali and Narasimhan, 1995).
X © ª
0 K s(2) (St , Si , S j )
∀K s(2) ∈8(2) (Si ,S j )
Ä(Si , S j ) = © ª (11.16)
0 8(2) (Si , S j )
Finally, to cover the case in which a variable is not present in any of the
cutsets of 8(2) (Si , S j ), its measure Ä(Si , S j ) is set to some high value, that
is, its effect on the reliability of Si is zero. A high value of Ä(Si , S j ) implies
that S j occurs in fewer cutsets through which Si can be estimated in the new
sensor network, or, conversely, the cardinality of such cutsets is large. Thus,
the variables that have the smallest measure should be chosen to be removed.
The measure developed is nonetheless only a guideline. Choosing the vari-
able with lowest entering score does not guarantee the increase of network
reliability. Then, the reliabilities of all variables are computed, and if the mini-
mum reliability over all variables is greater than the current network reliability,
then the choice holds. Otherwise, the next candidate is considered.
Algorithm
Based on the preceding discussion, the steps of the iterative improvement

algorithm are given below. Ali and Narasimhan (1995) called this algorithm
GSENNET. A flowchart describing the algorithm is also presented in
Figure 11.7
r Step 1. Obtain a spanning tree and generate an initial solution by placing
sensors on all the chords and remaining sensors on some of the branches of
the spanning tree. Compute the network, reliability.
r Step 2. Obtain all cutsets of the process graph.
r Step 3. Find the set of edges Rmin that have the least reliability.
r Step 4. Select an unmarked element Si ∈ Rmin and mark it. If no unmarked
edge exists, then stop.
FIGURE 11.7. Flowchart of algorithm GSENNET.
241
r Step 5. Obtain the set of leaving variable candidates Lc using Equation

(11.14). Arrange the variables in set Lc in decreasing order based on the
number of cutsets of 8(1) (Si ) in which they occur.
r Step 6. Pick the next unmarked edge Sq ∈ Lc and mark it. If an unmarked
element does not exist, go to step 4. Otherwise, select Sq as a leaving variable
and go to step 7.
r Step 7. Find the set Le of entering variable candidates as shown in Equa-
tion (11.15). Compute their measures using Equation (11.16) and arrange
them in order of decreasing measures.
r Step 8. If an unmarked element in Le does not exist, go to step 6. Otherwise,
choose the next unmarked element S p ∈ Le and mark it.
r Step 9. Evaluate the network reliability. If the network reliability improves,
accept S p as an entering variable and go to step 3. If the network reliability
does not improve, go to step 8.
The GSENNET algorithm does not guarantee that a globally optimal solution
will be reached because in any iteration it only searches for a better neighboring
solution that differs from the current solution in the placement of one sensor. If a
better neighboring solution cannot be obtained, then this is similar to obtaining
a locally optimal solution.
Example 11.3
We illustrate the algorithm for the sensor network design of a simplified

ammonia plant (Figure 11.1). For simplicity, the sensor failure probabilities for
all edges are 0.1.
Step 1. All the cutsets are given in Table 11.2.
Step 2. The initial solution is the spanning tree {S1 , S2 , S3 , S7 , S8 } in which
the measured edges are S4 , S5 and S6 . Because one more sensor needs to be
installed, edge S1 from this spanning tree is selected to be also measured. Thus,
the initial solution is edges x M = {S1 , S4 , S5 , S6 }. By computing the reliabilities
(Chapter 10), one obtains the values shown in Table 11.3:
TABLE 11.2. All Cutsets of Ammonia Plant Network.
No. Streams No. Streams No. Streams
1 S5 S6 S7 7 S1 S6 S8 13 S1 S4 S5
2 S1 S2 8 S1 S4 S6 S7 14 S2 S4 S6 S7
3 S1 S3 9 S1 S5 S7 S8 15 S3 S4 S6 S7
4 S2 S3 10 S2 S4 S5 16 S3 S4 S5
5 S2 S6 S8 11 S4 S7 S8 17 S2 S5 S7 S8
6 S3 S6 S8 12 S4 S5 S6 S8 18 S3 S5 S7 S8
TABLE 11.3. Reliabilities of the

Initial Solution.
Stream Riv (t ) Stream Riv (t )
S1 0.981 S5 0.981
S2 0.981 S6 0.900
S3 0.981 S7 0.972
S4 0.981 S8 0.972
Step 3. Because edge S6 has the smallest reliability, Rmin = {S6 }. Note that
although edge S6 is measured, it has the smallest reliability among all variables.
Steps 4 and 5. Rmin contains only one element. We mark it and go to step 6.
Steps 6 and 7. Using Table 11.2 we obtain the set 8(1) (S6 ), which contains
the following four cutsets.
K 1(1) (S7 , S6 ) = {S5 , S6 , S7 } Unmeasured: = {S7 }

K 2(1) (S8 , S6 ) = {S1 , S6 , S8 } Unmeasured: = {S8 }
K 3(1) (S7 , S6 ) = {S1 , S4 , S6 , S7 } Unmeasured: = {S7 }
K 4(1) (S8 , S6 ) = {S4 , S5 , S6 , S8 } Unmeasured: = {S8 }
The set of leaving variable candidates is given by the set of unmeasured

streams belonging to the above cutset elements of 8(1) (S6 ), that is, Lc =
{S7 , S8 }. Note that each of these variables occur in only two cutsets of 8(1) (S6 ).
We choose edge S7 as the leaving variable. This gives two more cutsets through
which mass flow in edge S6 can be estimated.
Step 8. Because the intersection of all sets of 8(1) (S6 ) is just S6 , the set of enter-
ing variable candidates obtained by using Equation (11.15) is Le = {S1 , S4 , S5 }.
The elements of 8(2) (S6 , S1 ) are the two following cutsets: K 1(2) (S8 , S6 , S1 ) =
{S1 , S6 , S8 } and K 2(2) (S7 , S6 , S1 ) = {S1 , S4 , S6 , S7 }. Therefore, 0{8(2) (S6 , S1 )} =
2. In addition, the total number of variables in these two cutsets is 7. Thus,
Ä(S6 , S1 ) = 3.5. In turn, the set {S6 , S4 } is present in two cutsets, which have
a total of eight variables. Similarly, the set {S6 , S5 } is present in two cutsets
totaling seven variables. Thus, the measures corresponding to the other two
entering candidates are: Ä(S6 , S4 ) = 4, Ä(S6 , S5 ) = 3.5. If we select edge S1
as the entering variable, the new set of measured edges is x M = {S4 , S5 , S6 , S7 }
and the network reliability actually reduces from 0.900 to 0.882, despite the
low value of its entering measure.
The selection of edge S4 as the entering variable does not improve the network
reliability and when edge S5 is chosen as the entering variable, the network
reliability improves from 0.900 to 0.964. Therefore, the choice of edge S5 as
the entering variable holds. On further iterations, the network reliability does
FIGURE 11.8. Simplified HDA process network.
not improve. This shows that a local optimal solution has been achieved by the
algorithm. Comparing this solution to the one obtained by explicit enumeration
Ali and Narasimhan (1995) found that this is also the global solution.
Ali and Narasimhan (1995) showed two more examples on the results of
the application of this algorithm for the simplified toluene hydrodealkylation
process (HDA) (Douglas, 1988). This process is shown in Figure 11.8. It has 8
nodes and 14 edges. The minimum number of sensors is 7.
Results of an exhaustive search are shown in Table 11.4, with results of the
algorithm included in the last two columns.
Less than 1% of the feasible solutions are optimal. GSENNET is capa-
ble of finding these solutions efficiently. In cases where the global optimum
TABLE 11.4. Optimal Solutions for the HDA Process (All Sensors with Sensor
Failure Probability = 0.1).
No. of No. of Optimal Optimal

No. of Feasible Optimal Global Solution Reliability
Sensors Solutions Solutions Reliability (GSENNET) (GSENNET)
7 992 55 0.729 S1 S5 S7 S9 0.729

S12 S13 S14
8 1,779 1 0.889 S2 S5 S6 S8 0.881
S10 S14
9 1,614 1 0.955 S2 S5 S7 S8 0.953
S10
10 929 1 0.975 S1 S2 S7 S8 0.975
11 358 1 0.986 S1 S6 S10 0.986
12 91 1 0.989 S1 S6 0.989
13 13 1 0.997 S4 0.997
14 1 1 0.998 — 0.998
Redundant Networks Featuring Maximum Reliability and Hardware Redundancy 245
TABLE 11.5. Failure Probabilities for the HDA Process.
Failure Failure Failure

Sensor Probability Sensor Probability Sensor Probability
S1 0.32 S6 0.23 S11 0.34

S2 0.16 S7 0.17 S12 0.24
S3 0.25 S8 0.10 S13 0.28
S4 0.37 S9 0.14 S14 0.12
S5 0.09 S10 0.27
is not obtained by GSENNET, the results differ from the optimum by less
than 0.5%.
An example with unequal sensor failures has also been solved. Table 11.5
shows the values of these failure probabilities, and Table 11.6 shows the results.
REDUNDANT NETWORKS FEATURING MAXIMUM RELIABILITY

AND HARDWARE REDUNDANCY
The GSENNET algorithm was extended by Ali and Narasimhan (1995) to

include hardware redundancy. For simplicity, it is assumed that at most two
sensors are used to measure a variable, although the method can be easily
generalized for a higher degree of hardware redundancy.
The leaving variable candidate set given by Equation (11.14) includes only
unmeasured variables. To allow for hardware redundancy, all measured vari-
ables that occur in cutsets 8(1) (Si ) that are currently measured by only one
TABLE 11.6. Reliabilities of the Initial Solution.
No. of No. of Optimal Optimal

No. of Feasible Optimal Global Solution Reliability
Sensors Solutions Solutions Reliability (GSENNET) (GSENNET)
7 992 2 0.443 S3 S4 S5 S6 0.433

S12 S13 S14
8 1,779 1 0.693 S1 S3 S4 S6 0.693
S10 S11
9 1,614 1 0.837 S1 S4 S6 S11 0.833
S13
10 929 1 0.897 S1 S4 S6 S11 0.881
11 358 1 0.945 S1 S4 S6 0.945
12 91 1 0.960 S4 S6 0.960
13 13 1 0.967 S4 0.967
14 1 1 0.967 — 0.967
sensor can also be leaving variable candidates, because these can also increase
the reliability of variable Si . In the case of entering variables, the candidate
set given by Equation (11.15) also includes all measured variables that are
being currently measured by more than one sensor. Both these modifications
increase the search space, leading to a corresponding increase in computational
requirements. Because it has been observed that hardware redundancy does
not increase the reliability as much as spatial redundancy, the first choice of
a leaving variable is an unmeasured variable. Similarly, the first choice of an
entering variable is one that is measured by more than one sensor.
In addition to the preceding modifications of leaving and entering variable
candidate sets, the computation of reliability has to be modified slightly to
account for variables that are measured by two sensors. If p1 and p2 are the
failure probabilities of two sensors measuring a variable, then the probability
that both of them fail is p1 p2 , assuming that the sensors fail independently.
Thus, for computing the reliabilities, the two sensors on a particular vari-
able can be considered as equivalent to one sensor with failure probability
p = p 1 p2 .
REDUNDANT AND RESTRICTED NETWORKS
As pointed out in Chapter 5, some variables may not be important and,

therefore, may not need to be observable. Similarly, some variables cannot be
measured because of technical and/or other economic reasons. In other words,
M R , the set of variables of interest, and M1 , the set of streams where sensors
can be placed, are not equal to the whole set of streams M. These networks are
called restricted. Vaclavek and Loucka (1976) and later Madron and Veverka
(1992) treated such specifications in their measurement placement strategies.
In this section, modifications to the GSENNET algorithm introduced by Ali
and Narasimhan (1995) are described so that such specifications on variables
can be performed. In addition to the sets of measured (x M ) and unmeasured
(xU ) variables, the following set of variables based on the specifications are
defined:
r x K ⊆ xU : set of key but unmeasurable variables

U
r x NK ⊆ xU : set of nonkey and unmeasurable variables
U
r x K ⊆ x M : set of key and measurable variables
M
r x NK ⊆ x M : set of nonkey, but measurable variables
M
In the design of a sensor network, only the reliabilities of key variables

(whether measurable or not) are of concern. As before, we attempt to maximize
the minimum reliability among all key variables.
Redundant and Restricted Networks 247
Well-Posedness
We first establish the conditions that these sets have to fulfill so that the
problem is not ill-posed. These conditions are:
NK
(1) The sets xUK , x M , xUNK and x M
K
are disjoint.
(2) All variables should be classified, that is, x = x M ∪ xU , xU = xUK ∪ xUNK
and x M = x MK
∪ xM NK
.
(3) The subgraph formed by the variables (streams) in xU should not contain
any cycle containing one or more variables from xUK . This is because in
such case these will be unobservable variables.
(4) The number of sensors selected should be such that observability of required
or important variables is feasible.
Number of Sensors for Restricted Networks
We now focus on a procedure that allows the determination of the minimum

number of sensors that can ensure observability. If one wants to estimate all
variables in a process, the minimum number of sensors n min is equal to n S −
n U + 1. If one is able to determine that certain streams can be unobservable,
then they can be eliminated by the merging procedure that was presented in
Chapter 5 to determine connectivity of systems and to obtain the redundant
network. In this section, the conditions on which different edges (streams) can
be eliminated are determined, and a new expression for the minimum number
of sensors is presented.
—All unmeasurable and nonkey variables xUNK will not be measured and are
not of interest. They can be eliminated.
—All measurable and key variables x M K
can belong to two types of cutsets.
r Cutsets containing some variables from set x NK . These cutsets are use-
U
less, because they contain a variable that cannot be measured. Measuring
variables from x M K
will at most contribute to the reliability of an unmea-
sured variable that is of no interest, and they can contribute to their own
reliability. Because they will not contribute to the reliability of any other
variable, they can be eliminated from the graph. We denote the number of
these variables n s,1 .
r Cutsets containing only key variables. Measuring such variables may con-
tribute to the reliability of other variables of interest.
NK
—All measurable and nonkey variables x M can belong to two types of cutsets.
r Cutsets containing some variables from set x NK . This cutsets are useless,
U
because they contain a variable that cannot be measured. Measuring vari-
NK
ables from x M will at most contribute to the reliability of unmeasured
variable that are of interest. Thus, these variables can be eliminated from
consideration, because they cannot possibly contribute to the reliability
of any key variable. To identify such variables, a spanning tree of the re-
duced network (obtained after eliminating nonkey unmeasurable and key
measurable variables in the preceding steps) is constructed choosing all
edges from set xUK first. This is possible because edges of set xUK do not
form cycles. Then, one edge at a time from set xUNK is tested. If it forms a
NK
cycle containing only edges of set x M that are in the tree, then this edge
is marked. It will be a chord of the spanning tree, but it does not need to
be measured. This can be explained by the fact that cutsets containing the
variable under analysis are formed by picking at least one other element
from each cycle to which the variable belongs. If one such cycle contains
only variables of no interest, then all cutsets will contain at least one vari-
able of no interest, and therefore measuring the variable under analysis is
NK
futile. After all edges from set x M have been considered, the spanning
tree can be completed by using edges from set x M K
. We denote by n s,2 the
number edges marked.
r Cutsets where all the rest of the variables are key variables. Measuring such
variables may contribute to the reliability of other variables of interest.
Then, the number of sensors required to observe all key variables is given by
n min = nrS − n rU + 1 + n S,1 − n S,2 (11.17)
where n rS and n rU are the number of edges and nodes of the reduced graph,
respectively.
This reduction method also provides the opportunity to reduce the compu-
tational requirements for a design of networks with a number of sensors larger
than the minimum above established. The algorithm for redundant networks
(GSENNET) can be applied to the reduced process graph using n min − n S,1
sensors to obtain their optimal locations. The remaining n S,2 sensors are placed
on edges belonging to the variables of the set x M
K
that were eliminated from the
original graph.
Modifications to GSENNET
The changes needed in GSENNET to address the existence of unmeasurable

and nonkey variables are:
r Only the reliabilities of important variables are computed at each iteration.
r Variables that belong to the set x K are not included in the leaving variables
U
candidate set Lc. This guarantees that at the final solution, no sensors are
placed on such variables.
TABLE 11.7. Data for the Constrained Steam Metering
System.
xM xU
xMK xMN K xUK xUN K
S1 S4 S5 S6 S7 S8 S9 S10 S14 S15 S2 S12 S18 S3 S13 S17 S11 S16 S19 S23
S20 S21 S22 S24 S25 S26 S27 S28
FIGURE 11.9. Steam metering network.
FIGURE 11.10. Reduced steam-metering network.
249
TABLE 11.8. Solution to the Steam-Metering Network.
Variable
No. of Optimal Solution (xU ) Global with Least
Sensors (GSENNET) Reliability Reliability
15 S3 S5 S10 S11 S12 S13 S14 S16 S17 S18 S19 S20 S23 0.430 S17
16 S3 S10 S11 S12 S13 S14 S16 S17 S18 S19 S20 S23 0.504 S17
17 S3 S11 S12 S13 S14 S16 S17 S18 S19 S20 S23 0.511 S17
Illustrative Example
To illustrate the performance of GSENNET in the case of where constraints

are imposed, the steam-metering network (Serth and Heenan, 1986) is used
for this purpose. This network is shown in Figure 11.9 and the constraints on
variables are shown in Table 11.7.
Nodal aggregation to eliminate all variables that are unmeasurable and non-
key results in the reduced graph shown in Figure 11.10. Basically, node U4 and
node U9 are merged with node U5 . In addition, nodes U8 and U10 are merged
with node U7 .
NK
As a result, streams S12 and S18 from x M become part of a loop that cannot
contribute to any other variable through balances. Thus, they can be eliminated.
None of the edges of x M K
are eliminated (n s,2 = 0). The reduced graph contains
NK
8 nodes and 22 edges. The only edges from x M left are S2 , and, therefore,
there is no loop. Thus, S2 is not eliminated (n s,1 = 0). The minimum number
of sensors is then 22 − 8 + 1 + 0 − 0 = 15.
Results of the design using GSENNET are shown in Table 11.8.
NOMENCLATURE

bx : branch x of a spanning tree
cp: chord p of a spanning tree
K s(1) (St , Si ): almost redundant cutset containing Si and unmeasured
variable St
K s(2) (St , Si , S j ): almost redundant cutset containing Si , mesured variable Si
and unmeasured variable St
K max : set of fundamental cutsets that have the maximum
cardinality
Lc: list of candidate leaving variables
Le: list of candidate entering variables
References 251

MR : set of key variables for reliability
n min : minimum number of measured streams to obtain
observability of all variables of interest
nU : number of units or nodes in a grap
n rS : number of streams or edges of a reduced graph
n rU : number of units or nodes of a reduced graph
pi : probability of variable i
R v (t): system reliability
Rmin : set of streams having the minimum reliability
Si : stream i
T: spanning tree
Ui : unit i
xU : set of unmeasured variables
xUK : set of key and unmeasurable variables
xUNK : set of nonkey and unmeasurable variables
xM K
: set of key and measurable variables
NK
xM : set of nonkey but measurable variables
Greek Letters
0{ }: cardinality function
8(1) (Si ): set of all cutsets containing Si that have only one
unmeasured variable St 6= Si
8(2) (Si , S j ): set of cutsets containing edges Si and S j containing some
unmeasured variable St 6= Si
Ä(Si , S j ): potential of measured variable S j to increase the reliability
of the network
REFERENCES
AIChE J., 39, 5, pp. 2237–2249 (1993).
Ali Y. and S. Narasimhan. Redundant Sensor Network Design for Linear Processes. AIChE J., 41,
10 (1995).
Douglas J. Conceptual Design of Chemical Process. McGraw Hill, New York (1988).
Process Networks. I. Theorems. Comp. & Chem. Eng., 12, 7, pp. 671–687 (1988a).
Models. AIChE J., 38, 2, pp. 227 (1992).
Nijenhuis A. and H. S. Wilf. Combinatorial Algorithms for Computers and Calculators. Academic
Press, New York (1978).
AIChE J., 32, pp. 733 (1986).
Vaclaveck V., M. Loucka M. Selection of Measurements Necessary to Achieve Multicomponent
Mass Balances in Chemical Plants. CES, 31, pp. 1199–1205 (1976).
CHAPTER 12
Design of Reliable Bilinear

INTRODUCTION
This chapter is devoted to extending the concepts presented in Chapter 11 to

the design of reliable bilinear sensor networks. Requirements of accuracy are
not considered.
BILINEAR MULTICOMPONENT SYSTEMS
We now present the method developed by Ali and Narasimhan (1996). The
following assumptions are made:
(1) Only minimal (nonredundant) sensor networks are considered. That is, net-
works in which only one way of estimating each variable exists. This is
equivalent to requesting estimability of order one for nonlinear systems.
(2) All concentrations/temperature of all components in all streams are nonzero.
This implies, for example, that no component is completely separated from
any stream by a separator. The assumption is trivial for bilinear systems
corresponding to energy balances.
(3) Either all or none of the compositions of a stream are measured, that is,
partial stream composition measurement is not considered. We call these
measure-all composition sensors.
(4) A single sensor is used to measure all compositions of a stream.
(5) Sufficient flow rates are measured so that it is possible to estimate all un-
measured flow rates by using flowmeters only (linearly estimable
networks).
253
254 DESIGN OF RELIABLE BILINEAR NONREPAIRABLE SENSOR NETWORKS
Assumption 1 limits the scope of the method to the development of nonredun-

dant sensor network design algorithms. However, these can serve as a starting
point for the optimal placement of redundant sensors.
Assumptions 2, 3 and 4 are relevant only for multicomponent processes,
and they make possible the quantification of the minimum number of sensors
required to estimate all variables.
The last assumption, which decomposes the bilinear problem into two cou-
pled linear problems, disallows certain feasible sensor network designs where
flow rates are estimated by using both flow rate and composition (or tempera-
ture) sensors. Ali and Narasimhan (1996) claim that this assumption is not so
restrictive because, in general, the estimation of flow rates using both flow rate
and composition sensors requires information from several sensors and leads
to lower accuracy and reliability.
Although the basic design algorithms are developed under these assumptions,
extensions to treat the general case are described later.
Processes without Splitters
We observed in Chapter 5 that when a multicomponent process consists of

mixers and separators only, and no splitters, then a complete set of nonredun-
dant equations that relate the variables is obtained by writing all the component
balances for every unit and normalization equation for every stream. It was also
assumed that compositions of a stream are not partially measured. Moreover,
if the composition sensor of a stream fails, then all compositions of that stream
become unmeasured, and none of them can be indirectly estimated by using the
normalization equation. Thus, only the overall flow and component balances
around each unit are useful in indirectly estimating variables and, hence, are
useful in computing reliabilities. A similar situation arises in multicomponent
processes, where one of the components in all streams is unmeasurable. For ex-
ample, in mineral beneficiation networks, the gangue component in all streams
is unmeasured. In these networks, normalization equations are not useful in
indirect estimation of mineral concentrations of streams.
It was also shown in Lemma 6.1, that under the above conditions the minimum
number of sensors needed is 2(n S − n U + 1). Moreover, it was shown that
flowmeters have to be placed on the chords of a spanning tree, and concentration
sensors should be located in the chords of another, not necessarily different
spanning tree.
Evaluation of Reliability
Because we are considering minimal sensor networks, measured variables

are nonredundant, and, therefore, the reliabilities of measured compositions
in a stream are all equal to the nonfailure probability of the set of sensors
used.
In the case of unmeasured compositions, they are observable by construction.
Their reliability should be computed by taking into account in how many cutsets
they are included. Moreover, all unmeasured compositions of a stream are
estimated indirectly by using the same set of cutsets. Therefore, all unmeasured
compositions in a stream have the same reliabilities. We now show how to
compute these reliabilities.
Let T x be the spanning tree formed by streams with unmeasured compo-
sitions. Let K sx (S j ) be the fundamental cutset with respect to this spanning
tree containing branch S j . To estimate the unmeasured mass fraction of com-
ponent t in stream j (x j,t ), the compositions of all the rest of the streams in
K sx (S j ) should be measured, and flow rates of all streams in K sx (S j ) should
be observable. Moreover, the cutset K sx (S j ) contains one or more streams with
unmeasured flow rates. This follows from the fact that if flow rates of all streams
in K sx (S j ) are measured, then it implies that all streams of this cutset are chords
of the spanning tree. This is impossible because a cutset cannot solely con-
sist of chords of some spanning tree. To illustrate this, consider the system of
Figure 12.1. This is similar to the system in Figure 5.9 with the difference that
a flash unit substitutes the splitter. Assume that we are trying to evaluate the
reliability of the unmeasured composition of stream S4 . The cutsets containing
stream S4 are also shown in the figure.
The measured flow rates are in S6 and S2 and the measured concen-
trations are in streams S1 and S5 . The corresponding spanning trees are:
FIGURE 12.1. Cutsets of the system in Figure 5.9.

T x = {S2 , S3 , S4 , S6 }, T m = {S1 , S3 , S4 , S5 }. It is easy to verify that some cut-

sets contain more than one unmeasured flow.
Let T m be the spanning tree consisting of unmeasured flows. Consider the set
of streams with unmeasured flow rates that belong to the union of all cutsets con-
taining S j . We denote this by {b1 , b2 , . . . , bn }. For S4 this set is {S1 , S3 , S5 }. Con-
sider now all the estimable cutsets containing these streams: K 1 , K 2 , . . . , K n .
All these cutsets contain measured flow rates that contribute to determine the
flow rates of S1 , S3 and S5 . Because by assumption flow rates have degree
of estimability one, there will be one such cutset per unmeasured flow rate.
Then the set Z j of flowmeters required to estimate flow rates of all streams of
{b1 , b2 , . . . , bn } is given by:
[
Zj = K sx ∪ K 1 ∪ K 2 ∪ . . . . . . . . ∪ K n − {b1 , b2 , . . . . , bk } (12.1)
s
In our example the estimable cutsets containing S1 , S3 , and S5 are K 1 =

{S1 , S6 }, K 2 = {S2 , S5 , S6 } and K 3 = {S2 , S3 , S6 }. Therefore Z j = {S2 , S6 }.
Thus, to estimate composition x j,t , composition sensors of all other streams
in Z j should be active. The reliability of x j,t is given by:
Y¡ ¢ Y
Rxv j,t (t) = 1 − pix × (1 − pi ) (12.2)
i∈K xj i∈Z j
where pi and pix are the failure probability of flow rate and composition sensor
of stream Si , respectively.
Properties of Optimal Bilinear Sensor Networks
As in Chapter 11, a sensor network for multicomponent processes is designed

to maximize the minimum reliability for all flow rate and composition variables.
In view of the assumptions made, the problem is then to determine which choice
of spanning trees T m and T x renders the maximum network reliability.
First, we prove that the least reliability is always attained by an unmeasured
composition.
Lemma 12.1: The minimum reliability is attained by an unmeasured com-
position in some stream and not by an unmeasured flow rate.
Proof: Let the least reliability among flow rates correspond to an unmea-
sured flow rate in stream Sk . This stream may be a chord or a branch of the
spanning tree T x , depending on whether the compositions of stream Sk are
measured or not. In either case, stream Sk must be present in some fundamen-
tal cutset of T x , because every branch and every chord of T x must be present
in some fundamental cutset. Let stream Sk be present in a fundamental cutset
K sx (S j ). Then, to estimate compositions of stream S j , all flowmeters necessary

for estimating the flow rate of stream Sk must be active together with other
flowmeters and composition sensors. Therefore, the reliability of x j,t is smaller
than the reliability of flow rates in stream Sk . Either x j,t has the least reliability
or some other unmeasured composition has lower reliability. In either case, the
least reliability is attained by an unmeasured composition. Q.E.D.
From this result, it follows that while evaluating the network reliability, we
need to evaluate the reliabilities of unmeasured compositions only.
In the case of equal failure probability for all sensors, consider the following
lemma:
Lemma 12.2: In bilinear sensor network with equal failure probabilities of
all flow sensors p, the optimal set of flow rates and composition spanning trees
(T m and T x ) are the same.
Proof: Consider first any given (not necessarily optimal) composition span-
ning tree T x . Consider a fundamental cutset K sx (S j ) of the process graph for
the given T x . The reliability of any composition x j,t , in stream j, is computed
by using Equation (12.2). From these equations, we can observe that for a given
T x , the only way of increasing
Q the reliability of x j,t is to choose the spanning
tree T m so that the product i∈Z j (1 − pi ) increases. For equal flowmeter failure
probabilities, this is achieved by minimizing the cardinality of set Z j . Because
every indirect estimation of a flow rate (using flow rate measurements only)
requires at least one other flowmeter, the set Z j has least cardinality when the
flow rates of all streams in K sx (S j ) (except stream S j ) are measured. In other
words, K sx (S j ) should also be a fundamental cutset corresponding to T m . Be-
cause this is valid for all fundamental cutsets of T x , therefore T m and T x should
be identical for the optimal choice of T x . Naturally, the largest reliability will
be obtained for the best spanning tree solution corresponding to the case when
only flow measurements are considered (T ∗ ). Q.E.D.
Corollary: In bilinear sensor network with equal failure probabilities of all
flow sensors p, and equal composition sensor failure probability p x, the optimal
set of flowmeters and composition spanning trees (T x and T m ) are the same as
the one corresponding to the optimal solution corresponding to pure flows only
T ∗.
Proof: Because T x = T m , then the reliability of the sensor network corre-
sponds to the spanning tree that has a cutset with largest cardinality. Q.E.D.
Consider now the case where the sensor failure probabilities for composition
sensors are not equal, but they are still equal for the flow sensors. To visualize
the effect, we present the example in Figure 12.2, where the simplified ammonia
process is presented. In this process the failure probabilities of flowmeters are
all 0.1 and the composition sensor failure probabilities are shown in Figure 12.2.
FIGURE 12.2. Ammonia plant network and composition sensor failure probabilities.
If all composition sensors have the same failure probability, then the best
solution is to locate flowmeters and composition sensors on streams S3 , S5 and
S6 . The corresponding spanning tree for maximum reliability of flows only, T ∗ ,
is shown in Figure 12.3a. However, if we adopt this pattern of flow measurement
for the present problem, the minimum reliability is 0.486, obtained for the
compositions of stream S4 . The optimal spanning tree solution for this problem
is shown in Figure 12.3b. By locating flowmeters and composition sensors on
streams S6 , S7 and S8 , a network with reliability of 0.605, corresponding to
the reliability of compositions in streams S1 , S2 and S3 , is obtained. This shows
that the solution T x = T m = T ∗ is not optimal when composition sensor failure
probabilities are unequal. However, as pointed out in Lemma 12.2, T x = T m
still holds.
The following trivial example shows that when the flow rate sensors have
unequal failure probabilities, the optimal solution does not feature the same
spanning trees, that is T x = T m is not a necessary condition of optimum. Con-
sider a simple process unit with one input stream and two output streams shown
in Figure 12.4. The numbers within parentheses for each stream correspond to
failure probabilities of flowmeters and composition sensors.
FIGURE 12.3. Optimal ammonia plant solutions.

FIGURE 12.4. System with unequal sensor failure probability.
The optimal solution for this process is to measure flow rates of streams S1
and S2 and compositions of streams S2 and S3 , giving T x = {S1 } and T m =
{S3 }, which are not identical. Such a possibility arises whenever the variable
reliability of a flow rate is higher if it is indirectly estimated than if it is measured.
For example, if we measure the flow rates of streams S1 and S2 , then the
reliability of estimating the flowrate of stream S3 is 0.56, which is higher than
directly measuring it with a reliability of 0.50.
Design Procedure of Bilinear Multicomponent Systems
In the previous section, three different cases of increasing complexity have

been discussed. We now present algorithms for the three cases.
r All sensor failure probabilities are equal: SENNET provides the optimal
solution.
r Equal sensor failure probabilities for flows only: Given the above finding
that both spanning trees coincide, then the solution is obtained with a slight
modification of SENNET. The entering and leaving variable candidate sets
are constructed in the same way as in the linear nonredundant case, that is,
exchanging a chord by a branch, so that a spanning tree structure is always
maintained. The modification consists in a different way of choosing the
entering and the corresponding leaving variable, which cannot be decided
by comparing the cardinalities of the cutsets involved because these are not
correlated with reliabilities. The choice is now arbitrary.
r Unequal sensor failure probabilities: For this case, the algorithm proposed
by Ali and Narasimhan (1996) is presented in the next section:
Algorithm for Unequal Sensor Failure Probabilities
Start with an initial spanning tree for T x , and keeping this fixed, attempt
to obtain a spanning tree T m that gives maximum network reliability. Phase 1
provides the spanning tree T m and Phase 2 improves on T x :

Phase I:
—Step 1. Set T x = T m .
—Step 2. Find the unmeasured composition, x j,t with the least reliability.
—Step 3. Obtain the set of streams, Z j whose flow rate measurements are
required to estimate x j,t .
—Step 4. For each stream c p of set Z j , identify all the fundamental cutsets of
T m that contain c p as a member. Let {K 1 , K 2 , . . . . , K r } be the set of such
fundamental cutsets, corresponding to the branches b1 , b2 , . . . , br of T m .
—Step 5. If for any of the fundamental cutsets, K q , obtained in step 4, the
following is satisfied:
¡ ¢ Y
1 − pc p ≤ (1 − pi ) (12.3)
i∈K q
i6=c p
then c p should enter and bq should leave T m .

—Step 6. If no such cutset exists, stop; otherwise, go to step 2.
Step 3 identifies all the flowmeters (chords of T m ) that are required to estimate
the least reliable composition variable. In step 4, an attempt is made to improve
the network reliability by interchanging one of these chords with a branch of a
fundamental cutset of which it is a member (elementary tree transformation of
T m ). The network reliability improves if the indirect estimation of the flow rate
of the chord is better than its direct measurement, which is checked in step 5.
Phase 1 of the algorithm thus gives the spanning tree T m corresponding to the
given T x that results in the maximum network reliability.
Phase 2: In this phase, an elementary tree transformation of T x is performed
in an attempt to find a better solution (T x , T m ), where T m is optimal for the
chosen T x . The entering and leaving variables for T x are obtained in the same
way as described in the previous section. For each choice of entering and leaving
variable candidates, a spanning tree T x is obtained, and its corresponding best
possible flow rate spanning tree can be obtained by using phase 1. If this solution
renders an improved network reliability, we accept it and proceed with the next
iteration of the algorithm; otherwise, we try the next choice of entering and leav-
ing variables for T x . If none of these choices leads to an improvement in network
reliability, the iterative process is terminated and the final solution is accepted.
Example 12.1
The simplified ammonia network shown in Figure 12.2 is used to illustrate

the proposed algorithm. Flowmeters are assumed to have a failure probability
equal to 0.1 and the failure probabilities of composition sensors are as shown
in the figure. We now apply SENNET.
—Construction of the spanning tree: Start with the spanning tree shown in
Figure 12.3a. This tree is T m = {S1 , S2 , S4 , S7 , S8 }.
—Identification of fundamental cutsets: The fundamental cutsets with respect
to this tree are:
r K 1 (S1 ) = {S1 , S3 }
r K 1 (S2 ) = {S2 , S3 }
r K 1 (S4 ) = {S3 , S4 , S5 }
r K 1 (S7 ) = {S5 , S6 , S7 }
r K 1 (S8 ) = {S3 , S6 , S8 }
—Calculation of minimum network reliability: The minimum network relia-
bility is 0.486, corresponding to the concentration in stream S4 . Therefore
bx = S4 .
—Determination of leaving variables: The leaving variable candidates are
those branches whose fundamental cutsets have at least one chord in common
with the fundamental cutset of stream S4 . Moreover, stream S4 may itself be
a leaving variable candidate. The fundamental cutset containing stream S4 is
K 1 (S4 ) = {S3 , S4 , S5 }. One chord of this cutset is S3 , and the fundamental cut-
sets containing this chord are K 1 (S1 ), K 1 (S2 ) and K 1 (S8 ). Similarly, the other
chord is S5 , and the fundamental cutsets containing this chord are K 1 (S7 ) and
K 1 (S8 ). Thus, the leaving variable candidate set is Lc = {S1 , S2 , S4 , S7 , S8 }.
We arbitrarily select S7 as the candidate for leaving variable.
—Determination of entering variables: The corresponding set of entering vari-
ables is Le = {S5 }, so we choose S5 as the entering variable. The new tree
becomes {S1 , S2 , S4 , S5 , S8 }, giving a network reliability of 0.498. We ob-
serve that the network reliability improves; therefore, the choice of leav-
ing and entering variables holds. In the next iteration, the leaving and en-
tering variable candidates are S8 and S3 , respectively, giving a new tree
T x = T m = {S1 , S2 , S3 , S4 , S5 } with a network reliability of 0.605. Because
the network reliability does not improve any further for any choice of en-
tering and leaving variables, we accept this as the final solution. Explicit
enumeration of all possible spanning tree solutions shows that this is also the
globally optimal solution.
Example 12.2
Consider the ammonia process shown in Figure 12.5. The failure probabilities
of flowmeters and composition sensors for each stream are, respectively, shown
within parentheses in the figure.
FIGURE 12.5. Ammonia plant network with unequal composition sensor failure proba-
bilities.
Choose an initial spanning tree T x = {S1 , S2 , S4 , S7 , S8 } as shown in

Figure 12.3a.
Phase 1: The optimal T m corresponding to this T x is obtained by using the
algorithm as follows:
—Steps 1–2. Initialize T x = T m = {S1 , S2 , S4 , S7 , S8 }. The network reliability

for this choice is 0.274, obtained for compositions in stream S8 .
—Step 3. The set Z 8 is equal to {S3 , S6 }.
—Step 4. The fundamental cutsets of T m containing chord S3 are:
(1) K 1 (S1 ) = {S1 , S3 }
(2) K 1 (S2 ) = {S2 , S3 }
(3) K 1 (S4 ) = {S4 , S3 , S5 }
(4) K 1 (S8 ) = {S8 , S3 , S6 }
The fundamental cutsets containing chord 6 are:
(5) K 1 (S7 ) = {S7 , S5 , S6 }
(6) K 1 (S8 ) = {S8 , S3 , S6 }
—Step 5. The fundamental cutsets containing S3 do not give an indirect way
of estimating the flow rate of chord S3 with better reliability than its direct
measurements. However, cutset K 1 (S7 ) provides an indirect way of estimat-
ing flowrates of S6 that is better than its direct measurement. So we exchange
branch S7 with chord S6 and get a new T m = {S1 , S2 , S4 , S6 , S8 } giving an im-
proved network reliability of 0.323, corresponding to compositions in stream
S8 . Further iteration does not lead to any improvement in the network reli-
ability. Thus, the best T m obtained for the initial T x is T m = {S1 , S2 , S4 ,
S6 , S8 }.
Phase 2: We now attempt to find an improved solution by elementary tree

transformations of T x . The leaving variable candidates can either be stream
S8 or branches of T x whose fundamental cutset has a common chord with

K 1x (S8 ),. Thus, the leaving variable candidate set is {S8 , S1 , S2 , S4 , S7 }. We
arbitrarily select edge S1 as the leaving variable and obtain the corresponding
set of entering variables:
Le = {S3 , S6 , S8 } − {{S3 , S6 , S8 } ⊕ {S1 , S3 }} = {S3 }
Because edge S3 is the only element of Le, we choose it as an entering

variable. The new T x is {S2 , S3 , S4 , S7 , S8 }. Corresponding to this T x , we find
the optimal T m , which happens to be identical. The network reliability for this
trial solution is equal to 0.303. We observe that the network reliability does
not improve; therefore, the choices of entering and leaving variable candidates
are rejected. We continue until all candidates from set Lc have been verified
and find that the network reliability does not improve. Thus, we accept T x =
{S1 , S2 , S4 , S7 , S8 } and T m = {S1 , S2 , S3 , S6 , S8 } as the solution.
Treatment of Splitter Units
In Chapter 6, it was discussed the way splitter units account for the role
of branches and chords of spanning trees. The net result of that analysis is
that one cannot consider spanning trees for T x containing one or two splitter
branches. If T x contains only one splitter branch, one of the splitter stream
compositions must be measured. In this case, among all the streams associated
to a splitter, the one whose composition sensor has the least failure probability
is measured. If T x contains two splitter streams as branches, then none of the
splitter streams have measured compositions. It is now possible to use the sensor
network design algorithm for the three different cases described earlier with the
following modifications.
(1) When splitters are present, it can be shown that in the optimal sensor net-
work design, T x and T m are identical for equal failure probabilities of
flowmeters, only if T x contains one splitter stream. Because, in general, it
cannot be determined whether T x and T m should be identical in the optimal
solutions, the algorithm for unequal sensor failure probabilities are always
used for any given failure probabilities of flowmeters and composition sen-
sors. Moreover, the initial spanning tree and entering variable set for T x is
chosen so that among streams incident on any splitter, not more than two
are branches of T x . This can be easily incorporated in the spanning tree
generation algorithm. At each iteration, we also chose the entering variable
so that not more than two streams incident on the same splitter became
branches of T x .
(2) To compute the reliability of a composition variable, one needs to identify
all flowmeters and all composition sensors that are useful for its estimation.
These are obtained as follows:
r If T x contains only one stream of a splitter, then this implies that the com-
positions of one of the splitter streams that has the least failure probability,
say psx , is measured. Thus, the composition reliability of all streams inci-
dent on that splitter are equal to (1 − psx ). If the fundamental cutset of any
other stream K xj (Si ) contains one or more streams of a splitter, then in the
reliability computation of composition x j , the composition of all these
splitter streams is assumed to be measured by using one composition
sensor with failure probability psx .
r If T x contains two streams of a splitter, say S1 and S2 , then none of the
splitter stream compositions of a splitter stream are indirectly estimated
by using the fundamental cutset, K xj (S1 ) or K xj (S2 ). Let K xj (S1 ) be the
cutset through which compositions of a splitter stream are estimated. Let
c1 , c2 , . . . , cr be the chords of K xj (S1 ) that do not correspond to a splitter.
The set of flowmeters required to estimate the compositions of streams
associated with a splitter is obtained as before by using Equation (12.2).
However, only the composition sensors of chords c1 , c2 , . . . , cr are re-
quired for estimating splitter stream compositions. Thus, the reliabilities
of all splitter stream compositions are given by:
Y¡ ¢ Y
R(x j,i ) = 1 − pix × (1 − pi ) (12.4)
i∈K xj i∈Z j
If the fundamental cutset of any other stream, K xj (S1 ) contains one or

more streams of a splitter, then the reliability of x1 also depends on the
flowmeters and composition sensors required to estimate compositions
of splitter streams.
It was proved in Lemma 12.1 that the least reliability is attained by an un-
measured composition. However, if splitters are present, there are rare instances
when an unmeasured flow rate of some stream associated with a splitter may
have lower reliability than all unmeasured compositions. This may occur, for
instance, if a splitter stream S j is a branch both in T x and T m , and if one of the
splitter streams has measured composition. In this case, the unmeasured flow
rate of stream S j will not be useful for estimating any of the unmeasured com-
position and Lemma 12.1 cannot be used. For the flow rate of stream S j to have
the least reliability, the flowmeter failure probabilities of splitter streams must
be substantially higher than other flowmeter and composition sensor failure
probabilities. Therefore, this case was ignored in the development.
Example 12.3
In this example, we illustrate the design of a sensor network when splitters

are present in the process. Consider the ammonia network given in Figure 12.2.
For reasons of simplicity, we assume that all flowmeters have equal failure prob-
ability of 0.10 and all composition sensors also have equal failure probability
of 0.20. We also assume that only a single splitter is present in the process that
is represented by node U5 .
To observe flowrates and compositions of all streams, the minimum number
of flowmeters is three, whereas that of composition sensors is two, which is
given by Equation (6.20). In this equation n S = 8, n U = 6, n sp = 1 (one splitter)
and m s,1 = 3 (three nodes incident to the splitter). We start with an initial tree
T x = T m = {S2 , S3 , S4 , S5 , S8 }. Note that we cannot start with the spanning
tree given in Figure 12.3b, because it contains three splitter branches. The
fundamental cutsets of this spanning three are:
r K 1 (S2 ) = {S1 , S2 }
r K 1 (S3 ) = {S1 , S3 }
r K 1 (S5 ) = {S5 , S6 , S7 }
r K 1 (S4 ) = {S1 , S4 , S6 , S7 }
r K 1 (S8 ) = {S1 , S6 , S8 }
Since T x contains two splitter branches, all splitter stream compositions
are unmeasured. Thus, the compositions of stream S7 (a splitter edge) is un-
measured. However, the flow rate of this stream is measured. Reliabilities of
compositions for all streams, except for S4 , S5 and S7 , can be computed by
using Equation (12.2), because the fundamental cutsets of these streams in T x
do not contain any splitter stream.
The compositions of splitter stream S8 are observable by using the flow rate
and composition measurements of streams S1 and S6 , that is, by using funda-
mental cutset K 1 (S8 ) with a reliability of 0.518. Therefore, the compositions
of splitter streams S4 and S7 are also observable with the same reliability. The
compositions of stream S5 can be estimated by using the flow rate and compo-
sition measurements of streams S6 and S7 (fundamental cutset 3). However, the
compositions of stream S7 are indirectly estimated by using flow rate and com-
position sensors of streams S1 and S6 . Therefore, the compositions of stream
S5 are observable through flow rate measurements of streams S1 , S6 and S7
and the composition measurements of streams S1 and S6 , with a reliability of
0.467.
Corresponding to this T x we obtain the optimal T m using phase 1 of the
algorithm for unequal sensor failure probabilities. We observe that T x and T m
are identical. The network reliability is 0.467, which is attained by compositions
of stream S5 . As in SENNET (Ali and Narasimhan, 1993), we find the set of
leaving variables is Lc = {S2 , S3 , S4 , S8 }. If edge S4 is chosen as the leaving
variable, the corresponding set of entering variables is Le = {S1 , S6 , S7 }. If we
choose either S6 or S7 as the entering variable, the network reliability does
not improve. However, the network reliability improves if edges S1 or S8 are
chosen as leaving and entering variables. This gives a new spanning tree solution
{S1 , S2 , S3 , S4 , S5 }. Further iterations do not improve the network reliability.
Explicit enumeration of all solutions shows that this solution is also the global
optimum.
ENERGY NETWORKS
For this case, the same approach as the one shown in Chapter 6 is used. All
the algorithms developed for multicomponent networks can now be applied to
obtain the optimal placement of flow and temperature sensors. The streams with
unmeasured flows will form a spanning tree of G m , whereas the streams with
unmeasured temperatures will form a spanning tree of G e . As an example, Ali
and Narasimhan (1996) solved a flowsheet of a crude preheat train of a refinery
by using the same technique. The example is omitted because it does not add
new conceptual material.
NOMENCLATURE

K xj : jth cutset of the concentration graph
Lc: list of candidate leaving variables
Le: list of candidate entering variables
nU : number of units or nodes in a grap
pi : failure probability of flow rate sensor of stream Si
pix : failure probability of composition sensor of stream Si
Si : stream i
T: spanning tree
T x: spanning tree corresponding to concentration measurements
in its chords
T m: spanning tree consisting of unmeasured flows
T ∗: optimal solution corresponding to pure flows only
Ui : unit i
xU : set of unmeasured variable
Z j: set of flowmeters required to estimate all flowrates
of K xj
References 267
REFERENCES
AIChE J., 39, 5, pp. 2237–2249 (1993).
Ali Y. and S. Narasimhan. Sensor Network Design for Maximizing Reliability of Bilinear Processes.
AIChE J., 42, 9, pp. 2563–2575 (1996).
CHAPTER 13
Design of Reliable and Cost-Efficient

INTRODUCTION
Chapters 11 and 12 presented methodologies to design reliable systems based

on the concept of maximizing the system reliability where cost is not taken into
consideration. Most of the material for this chapter, which adds cost consider-
ations to the problem, is taken from Bagajewicz and Sánchez (2000).
MINIMUM COST MODEL
The design of the sensor network subject to reliability constraints can be

written as follows:
X 
Min ci qi 



∀i∈M1 
s.t. (13.1)
Rkv (q) ≥ Rk∗ ∀k ∈ M R 




qi ∈ {0, 1} ∀i ∈ M1
where M1 is the set of streams where sensors can be placed, M R is the set
of variables whose reliability is to be constrained, qi are the binary variables
determining whether a sensor is located in stream Si (qi = 1), or not (qi = 0)
and ci are the corresponding costs.
The reliability of each variable is calculated using the failure probabilities of
all the sensors participating in the cutsets as discussed in Chapter 10.
269
270 DESIGN OF RELIABLE AND COST-EFFICIENT
MODEL FOR MINIMAL NETWORKS
Assume first that all sensors have the same cost c and that N is the number
of sensors. Then,
X
ci qi = cN (13.2)
∀i
Therefore, one can rewrite model (13.1) as a minimum number of sensors
model:
X 
Min qi 

∀i


s.t. (13.3)
Rkv (q) ≥ Rk∗ ∀k ∈ M R 



qi ∈ {0, 1} ∀i ∈ M1
Solution Procedure
Models (13.1) and (13.3) are mixed integer programming problems. More-
over, the constraints regarding reliability are in principle constructive con-
straints, in the sense that they involve an algorithm for the evaluation of re-
liability. However, an explicit mathematical expression to calculate Rkv (q) can
also be used. One can also use the tree search methodology developed
by Bagajewicz (1997) for this type of problems, which was presented in
Chapter 6. This methodology is based on an enumeration of every branch of the
tree of alternatives aided by a stopping criterion. In essence, the enumeration
proceeds until a feasible node of the tree is found. This node is evaluated and
compared with the existing best solution, and the developing of that branch is
stopped because any addition of sensors leads to an increase in cost. Although
the computational effort for this technique is reasonable for small problems,
the technique is not very efficient for large problems. There is, therefore, a need
to develop algorithms that are more efficient to solve these problems. Such
methods have not yet been developed.
Relation to Other Models
The model presented by Ali and Narasimhan (1993), which was discussed
in detail in Chapter 11, can be put in the following form:

Max Min Rkv (q) 

∀k∈M R 

s.t. 


∗
E i (q) ≥ E i ∀i ∈ M R
X (13.4)
qi = N ∗ 



∀i 


qi ∈ {0, 1} ∀i ∈ M1
The first constraint puts a lower bound on the estimability of the variables
of interest. The second constraint fixes the number of sensors. In the case of
minimal networks, N ∗ should be chosen to be the number of chords of the
spanning tree and the threshold estimability is set to one. In this case, these two
can be substituted by a unique constraint requiring estimability of order one in
all variables. We saw in Chapter 6 that this is a property of a minimal network.
For the case of redundant networks, the number of sensors can be increased, and
the estimability requirements can be still set to one. Such use of model (13.4)
is equivalent to the use of the methodology presented in Chapters 11 and 12.
Limitations of Previous Models
The model (13.4) is successful in identifying reliable sensor networks, as

described in Chapters 11 and 12. However, it has the following limitations.
(1) Cost is not explicitly considered so that the solutions may not be cost-
effective.
(2) It does not allow limiting the reliability of specific variables, because it
relies on maximizing only the smallest, which does not guarantee desired
reliability levels in others.
(3) It cannot be used in conjunction with other goals such as accuracy, error
detectability etc.
(4) In the case of the design of redundant sensor networks, there is no control
over which variable is redundant. Moreover, there is no control on what is
the degree of redundancy of the system or the different ways a particular
variable can be estimated.
(5) The set of variables for which reliability is requested M R is equal to the
whole set of variables of the system.
(6) The set of variables where sensors can be placed (M1 ) is also the whole
set of variables. Although some modifications can be made to address this
problem in the context of the methodology based on the choice of leaving
and entering variables, the effect of such restrictions on the ability to find
optimal solutions has not been investigated.
To address these limitations a more general maximum reliability model is
needed. In the next sections, such model is presented, and its connections with
the minimum cost model given by model (13.1) are established.
GENERALIZED MAXIMUM RELIABILITY MODEL
Assume now that a new constraint is added to model (13.1) in which the
minimum reliability of all the variables of interest is used, as follows:
X 
Min ci qi 

s.t. ∀i 



v ∗
Min Rk (q) ≥ Min Rk (13.5)
∀k3M R ∀k3M R


Rkv (q) ≥ Rk∗ ∀k ∈ M R  


qi ∈ {0,1} ∀i ∈ M1
Note that the addition of the constraint is trivial, because it is guaranteed by

the other constraints of reliability of individual streams. We now use the concept
of duality according to Tuy (introduced in Chapter 8) to obtain the following
result:
 X 

 cT = α ≤ Min f (x) = Min ci qi 


 


 i∈M1 


 

s.t. 
v ∗
 g(x) = Min Rk ≥ Min Rk = β 

 ∀k∈M R ∀k∈M R 


 Rv ≥ R∗ 


 ∀k ∈ M R 



k k 

qi ∈ {0,1} ∀i ∈ M1
 
 β = Min Rk∗ ≥ Max g(x) = Max Min Rkv 

 ∀k∈M ∀k∈M 



R R



 s.t. 


 X 

⇔ f (x) = c q
i i ≤ c T = α (13.6)

 



i∈M1



 Rk ≥ Rk ∗
∀k ∈ M R 


 

 
qi ∈ {0,1} ∀i ∈ M1
Thus, the maximum reliability model constrained by cost follows:

Max Min Rkv 

∀k∈M R 



s.t. 

X 
ci qi ≤ cT (13.7)


i∈M1


Rkv ≤ Rk∗ ∀k ∈ M R 




qi ∈ {0,1} ∀i ∈ M1
The objective function of this problem is the same as the one proposed by
Ali and Narasimhan (1993). A few simplifying assumptions will produce a
problem constrained by a fixed and minimum number of sensors, which is the

model they presented.
Consider all the sensors having the same cost, so that the cost constraint
becomes a constraint on the total number of sensors.
X X
ci qi ≤ cT qi ≤ N ∗ (13.8)
i∈M1 i∈M1
Thus, when the total number of sensors is set to a minimum and the cons-
traints on the reliability of individual variables is dropped, the problem
becomes:

Max Min Rkv 
∀k∈M R 



s.t. 

X
qi ≤ N ∗ (13.9)




i∈M1 


qi ∈ {0,1} ∀i ∈ M1
Consider now the following problem, where the constraint is made an

equality.

Max Min Rkv 

∀k∈M R 



s.t. 
X (13.10)
qi ≤ N ∗ 



i∈M1 


qi ∈ {0,1} ∀i ∈ M1
Lemma 13.1: The solution to model (13.9) has the same objective function
value as the solution of model (13.10).
Proof: If the constraint is not binding, then, if e
q is the vector corresponding
to the optimal solution
X
qi < N ∗
e (13.11)
i∈M1
Therefore, it is possible to add a sensor to the system. Such an addition can

leave the reliability of the system unaltered or increase it. Assume, without loss
of generality, that the optimum of the problem corresponds to the reliability of
estimation of the flowrate of stream S1 . That is

v v
Min e
Rk = e
R1 (13.12)
∀k∈M R
A new sensor can be located in a stream where a new balance equation

involving the flowrate of S1 cannot be written. In such case, the reliability of S1
is not altered. On the contrary, if a new balance equation can be written, then
the reliability of S1 will increase, because reliability is a monotone function of
the number of sensors involved. Q.E.D.
Example 13.1
We now illustrate the above lemma in the example of Figure 13.1. Assume
that in this example the set M R = {S1 , S2 , S3 }.
Assume that a feasible solution of model (13.9) is given by only one mea-
surement in streams S3 , that is, q = (0,0,1,0,0), as indicated in Figure 13.1.
It is easy to see that an addition of a measurement in stream S2 will increase
the reliability of S1 , whereas the addition of a measurement in S4 or S5 will
not. Indeed, the addition of a measurement in S4 cannot be used in conjunction
with S3 to provide a new estimate of S1 . Therefore, if the original constraint
sets a maximum of two sensors (N ∗ = 2), then the constraint on the number
of sensors will be binding if sensors in S1 and S2 can be placed. Otherwise,
the one sensor solution q = (0,0,1,0,0) will have the same reliability as the two
sensors solutions q = (0,0,1,1,0), or q = (0,0,1,0,1).
The above model addresses most of the concerns raised when analyzing the
model presented by Ali and Narasimhan (1993). It explicitly considers the cost,
it can impose lower bounds on the reliability of specific variables and it can
restrict the set where sensors can be located. Other goals, such as accuracy
goals, can be easily added to the minimum cost model as constraints. These
constraints restrict the feasible region and remain as such in the Tuy-type dual
problem. Finally, this model cannot explicitly control the degree of redundancy.
This is considered in the next section.
Reliable Sensor Networks with Estimability Constraints
We now proceed to generalize the maximum reliability model to include

estimability constraints. Consider the addition of estimability constraints to the
minimum cost model (13.1).
X 
Min ci qi 



∀i∈M1 

s.t. 

Rkv (q) ≥ Rk∗ ∀k ∈ M R  (13.13)


E k (q) ≥ E k∗ ∀k ∈ M E 




qi ∈ {0,1} ∀i ∈ M1
Using the duality according to Tuy, one can convert this model into maximum
reliability models.
Reliable Sensor Networks with Accuracy Constraints
We now proceed to generalize the maximum reliability model to include

accuracy constraints. The resulting model is:
X 
Min ci qi 



∀M1 

s.t. 

Rkv (q) ≥ Rk∗ ∀k ∈ M R  (13.14)


σk (q) ≥ σk∗ ∀k ∈ M P 




qi ∈ {0,1}∀i
Using the duality according to Tuy, one can convert this model into maximum
reliability models or into a maximum precision one. Note that estimability
constraints can still be added.
Other Illustrative Examples
Consider the same ammonia network example presented in Chapter 11

(Figure 11.1). The cutsets corresponding to this network are given in
Table 11.2. We now present different cases illustrating the above concepts. Data
for these cases are given in Table 13.1 and solutions are given in
Table 13.2
TABLE 13.1. Data for Illustrative Examples.
Estimability
Reliability Lower Bounds Lower Bounds
Case Instrumentation Cost R∗ E∗
1 2000 * [1 1 1 1 1 1 1 1] 0.81 * [1 1 1 1 1 1 1 1] —
2 1000 * [1.5 2.0 2.3 2.8 1.7 2.0 1.5 2.8] 0.81 * [1 1 1 1 1 1 1 1] —
3 2000 * [1 1 1 1 1 1 1 1] 0.9 * [1 1 1 1 1 1 1 1] —
4 1000 * [1.5 2.0 2.3 2.8 1.7 2.0 1.5 2.8] [0.9 0 0.81 0 0 0 0.729 0] —
5 1000 * [1.5 2.0 2.3 2.8 1.7 2.0 1.5 2.8] [0.95 0 0.95 0 0 0 0.95 0] —
6 2000 * [1 1 1 1 1 1 1 1] 0.81 * [1 1 1 1 1 1 1 1] [ 2 0 2 0 0 0 2 0]
7 1000 * [1.5 2.0 2.3 2.8 1.7 2.0 1.5 2.8] 0.81 * [1 0 1 0 0 0 1 0] [ 2 0 2 0 0 0 2 0]
8 1000 * [1.5 2.0 2.3 2.8 1.7 2.0 1.5 2.8] 0.81 * [1 0 1 0 0 0 1 0] [ 2 0 3 0 0 0 2 0]
9 1000 * [1.5 2.0 2.3 2.8 1.7 2.0 1.5 2.8] 0.9 * [1 0 0 1 0 0 1 0] [ 0 3 0 0 0 3 0 3]
TABLE 13.2. Solutions of Illustrative Examples.

Case
(Cost) Solutions Reliabilities Estimabilities
S1 S4 S8 0.90 0.90 0.90 0.90 0.81 0.81 0.81 0.90
S1 S5 S6 0.90 0.90 0.90 0.81 0.90 0.90 0.81 0.81
S2 S4 S8 0.90 0.90 0.90 0.90 0.81 0.81 0.81 0.90
S2 S5 S6 0.90 0.90 0.90 0.81 0.90 0.90 0.81 0.81
1 S3 S4 S8 0.90 0.90 0.90 0.90 0.81 0.81 0.81 0.90 —
(6000) S3 S5 S6 0.90 0.90 0.90 0.81 0.90 0.90 0.81 0.81
S4 S5 S7 0.81 0.81 0.81 0.90 0.90 0.81 0.90 0.81
S6 S7 S8 0.81 0.81 0.81 0.81 0.81 0.90 0.90 0.90
2 S1 S5 S6 0.90 0.90 0.90 0.81 0.90 0.90 0.81 0.81 —
(5200)
S1 S4 S6 S7 0.973 0.973 0.973 0.973 0.964 0.973 0.973 0.964
S1 S5 S7 S8 0.973 0.973 0.973 0.964 0.973 0.964 0.973 0.973
S2 S4 S6 S7 0.973 0.973 0.973 0.973 0.964 0.973 0.973 0.964
3 S2 S5 S7 S8 0.973 0.973 0.973 0.964 0.973 0.964 0.973 0.973 —
(8000) S3 S4 S6 S7 0.973 0.973 0.973 0.973 0.964 0.973 0.973 0.964
S3 S5 S7 S8 0.973 0.973 0.973 0.964 0.973 0.964 0.973 0.973
S4 S5 S7 S8 0.964 0.964 0.964 0.973 0.973 0.973 0.964 0.973
4 S1 S7 R1 = 0.9 R3 = 0.9 R7 = 0.9 —
(3000)
5 S1 S5 S7 S8 R1 = 0.973 R3 = 0.973 R7 = 0.973 —
(7500)
6 Same as in case 3 All binding
(8000)
7 Same as in case 5 All binding
(7500)
8 S1 S2 S5 S7 S8 R1 = 0.9973 R3 = 0.9973 R7 = 0.9802 E1 = 3 E3 = 3 E7 = 2
(9500)
9 S1 S2 S5 S6 R1 = 0.9988 R3 = 0.988 R7 = 0.989 E2 = 3 E6 = 3 E8 = 3
(11500) S7 S8
276
TABLE 13.3.
Instrumentation Cost Reliability Lower Bounds Precision Upper Bounds
1000 * [1.5 2.0 2.3 2.8 1.7 2.0 1.5 2.8] R ∗ = 0.9 * [1 0 0 1 0 0 1 0] σ ∗ = [0 3 0 0 0 2 0 2]
Case 1: This first case corresponds to the same problem presented by Ali
and Narasimhan (1993) in this opportunity solved by using the minimum cost
model. The constraints in reliability were chosen equal to the solution obtained
by Ali and Narasimhan (1993). Different alternative solutions having the same
cost of 6000 are obtained.
Case 2: This case explores variants of the above solution by altering the cost
of the different sensors. The optimal solution is now only one.
Case 3: In this case uniform costs are used again, but higher reliability
thresholds are introduced.
Case 4: Different costs and different reliability constraints are used in this
case. However, only three streams are restricted. These types of solutions cannot
be obtained using the models presented in Chapter 11.
Case 5: This case is the same as case 4, with an increase in the lower bound
on reliability.
Case 6: This case explores the effect of adding estimability constraints. All
estimability constraints are binding, with no variable having an estimability
larger than two. Observing that the solution is one, as in Case 3, this example
illustrates how estimability constraints can play a role similar to reliability
constraints. However, as discussed before, they are not completely equivalent.
Case 7: Cost is changed in this case with only some reliability and estima-
bility constraints. All the estimability constraints are binding.
Case 8: This case explores the increase of the estimability of variable S3 in
case 7 to three (E 3∗ = 3), while the rest of the data remains the same. Estima-
bilities are not all binding.
Case 9: Tight reliability constraints in variables and different estimability
constraints, are explored in this case.
Case 10: To show how reliability and accuracy constraints interact, the ex-
ample of Table 13.3 was prepared. Results of model (13.14), with a resulting
cost of C = 8000, are shown in Table 13.4.
All the examples presented above show the richness of the sensor location
problem, when several constraints of different nature are imposed. The chal-
lenge is now to be able to solve large systems. Techniques capable of addressing
this issue will be presented in Chapter 16.
TABLE 13.4. Results for Case 10.

Solution Reliabilities Precision
S2 S5 S6 S7 S8 R1 = 0.9729 R3 = 0.9639 R7 = 0.9729 σ2 = 1.3297 σ6 = 1.9911 σ8 = 1.9249
NOMENCLATURE
N: number of sensors
N ∗: number of sensors fixed for design
q: binary vector indicating whether sensors are located (1) or not
(0) in variables
Riv (q): variable estimation reliability as a function of q
Ri∗ : variable estimation reliability threshold q
Si : stream i
Ui : unit i
Greek Letters
σ: precision
REFERENCES
AIChE J., 39, 5, pp. 2237–2249 (1993).
2306 (1997).
Bagajewicz M. and M. Sánchez. Cost-Optimal Design of Reliable Sensor Networks. Comp. &
Chem. Eng., 23, 11/12, pp. 1757–1762 (2000).
CHAPTER 14
Design of Repairable Sensor Networks
INTRODUCTION
In this chapter the service and estimation availability and reliability of sen-
sor networks that are subject to corrective maintenance are studied. Finally, a
minimum cost model is presented.
As it was discussed in Chapter 10, when repairs are not present, the service
availability of a sensor is equal to its service reliability. In addition, the failure
rate has been considered constant. However, in the presence of repairs, failure
is no longer an event that depends on how many hours the sensor survived
from the time it has been put in service. It is also conditioned by the fact that
maintenance is made and/or the sensor has been repaired at a certain time after
being put in service. These events condition the failure rate. Thus, we distinguish
unconditional from conditional events in failure and repair. These concepts are
first formally introduced to be later used to determine sensor maintenance cost,
which accounts for nearly 20% of all maintenance cost (Masterson, 1999). Its
reduction or containment is essential.
FAILURE INTENSITY
Definition: The unconditional failure intensity wi (t0 , t) at time t is the prob-

ability that the sensor fails per unit time at time t, given they are normal at time
t0 .
Definition: The conditional failure intensity of sensor i, λi (t) at time t is the

probability that the sensor fails per unit time at time t, given that it was at the
normal state at time t.
279
280 DESIGN OF REPAIRABLE SENSOR NETWORKS
To help understand the difference between these two concepts, assume a sys-
tem of many alike sensors with a constant failure rate. As time goes by, sensors
fail at a constant rate, that is, the number of sensors failing is proportional to the
number of normally functioning sensors. This is no other than the conditional
failure intensity. Thus, for a system with constant failure rate the conditional
failure intensity is equal to the failure rate (λi (t) = r ). However, the uncondi-
tional failure intensity is the proportion of the original sensors that fail. For the
system in question, at a constant failure rate, the number of sensors failing is
lower as time goes by; therefore, the unconditional failure rate drops. In the
general case, when the rate of failure r is not constant, then the conditional
failure intensity is not equal to the failure rate.
REPAIR INTENSITY
Once one is able to calculate what the probability is of a given number of

sensors failing, we assume that they will be repaired. Thus, much in a similar
way as in the case of failures the following is defined:
Definition: The unconditional repair intensity vi (t0 , t) at time t is the proba-
bility that the component is repaired per unit time at time t, given that it jumped
into the normal state at time t0 .
Definition: The conditional repair intensity µi (t) is the probability that a
component is repaired per unit time at time t, given that is failed at time t.
Assume again a system of many alike sensors. The unconditional repair
intensity counts the number of sensors being repaired, as a fraction of the
original number of sensors. On the other hand, the conditional repair intensity
is the number of repairs as a fraction of the number of failed sensors at time t.
Thus, if the repair rate m is constant, the conditional repair intensity is equal to
the repair rate (µi (t) = m) (Henley and Kumamoto, 1981).
EXPECTED NUMBER OF REPAIRS
The real number of repairs in the interval (t, t + dt) is directly related to the
unconditional repair intensity by: vi (t0 , t)dt. The conditional repair intensity
is not used because the unconditional repair intensity counts the fraction of
sensors being repaired with respect to the total original number of sensors, not
with respect to those that had failed at time t. Thus, the expected number of
repairs (3i ) in the interval [t0 , t), is
Z t
3i (t0 , t) = vi (t0 , t) dt (14.1)
t0
To calculate the unconditional repair intensity vi , we will assume the follow-

ing:
—The conditional failure intensity (λi ) is known and is constant.
—The sensor does not deteriorate in time, that is, that the probability of failing
before and after the repair does not change. These are known as AGAN
(as-good-as-new) type of maintenance.
—If the sensor has been repaired, even if constant, its failure intensity does not
change.
—The conditional repair intensity (µi ) is known and is constant. This is given
by the rate at which the instruments can be repaired, or in more practical
terms the number of these sensors per unit time that can be repaired.
Systems with an increasing failure rate with time are intentionally left out
from consideration, because no design procedures for sensor networks featur-
ing nonconstant failure rates have been developed yet. Similar consideration
applies to repair rates. Finally, sensors can have different failure modes, because
different components can have different failure rates. However, for the same
reasons outlined above, only one failure mode is considered.
Number of Repairs under Corrective Maintenance
We now aim at obtaining an expression for the number of repairs when the
instruments are repaired immediately after they fail. This last assumption is
crucial for a development of an expression of the unconditional repair intensity.
Assume the following events:
r Bi = the component is failed at time t.
r Ci = the component is repaired during (t, t + dt).
r Wi = the component was in the normal (operating) state at time t0 .
Then we can write an expression for the unconditional repair intensity:
vi (t0 , t) = Pr(Ci | Wi ) (14.2)
However, events Bi and Ci are by assumption simultaneous. In other words,

Ci is included in Bi or implied by Bi . In other words, Ci = Ci ∩ Bi . Thus, using
laws of probability:
vi (t0 , t) = Pr(Ci | Wi ) = Pr(Ci ∩ Bi | Wi ) = Pr(Ci | Bi ∩ Wi ) Pr(Bi | Wi )

(14.3)
The first term is the probability that the system is being repaired at time t
given it has failed at time t and it was in normal state at time t0 . This is no other
than the unconditional repair intensity, that is
µi (t) = Pr(Ci | Bi ∩ Wi ) (14.4)
The second term in Equation (14.3) is the probability that the system has
failed at time t given it was in normal state at time t0 . This is no other than the
service unavailability, that is
Uis (t) = Pr(Bi | Wi ) (14.5)
therefore,
vi (t0 , t) = µi (t)Uis (t) (14.6)
and thus for a constant unconditional failure intensity (µi = m i ) the expected
number of repairs is:
Z t
3i (t0 , t) = µi Uis (t) dt (14.7)
t0
We now proceed to develop a Markov model that will help us determine the
unavailability of sensor i. We assume that transitions from the failed state to
the normal state (and vice versa) occur one at a time, that is, they cannot take
place simultaneously. In other words, the possibility of two or more transitions
is negligible. This is depicted in Figure 14.1.
The basis for these markovian models has been introduced in Chapter 10.
Using Equations (10.6) and (10.7) for our case we obtain:
d Ais (t0 , t)
= µi Uis (t0 , t) − λi Ais (t0 , t) (14.8)
dt
Equation (14.8) is explained as follows: The rate at which the sensor changes
its availability (left side) is equal to the rate at which unavailable sensors are
FIGURE 14.1. Markov representation of corrective maintenance.

repaired minus the rate at which they fail. Using Equation (10.1) and integrating:
µ ¶
λi λi
Uis (t0 , t) = + Uis (t0 , t0 ) − e−(λi +µi )(t−t0 ) (14.9)
λi + µi λi + µi
Ais (t0 , t) = 1 − Uis (t0 , t) (14.10)
Thus, from Equations (14.6) and (14.9) we conclude that after a long period
of time the unconditional repair intensity is given by:
µi λi
lim vi (t0 , t) = (14.11)
t→∞ λi + µi
In other words, when the repair rate is much larger than the failure rate (µi À
λi ), the unconditional repair intensity is given by the failure rate (vi (t0 , t) → λi ).
Conversely, if the failure rate is much larger than the repair rate the unconditional
repair intensity approaches the repair rate (vi (t0 , t) → µi ).
The expected number of repairs is obtained by using Equation (14.7), that is:
·
λi
3i (t0 , t) = µi (t − t0 )
λi + µi
µ ¶ ¸
Uis (t0 , t0 ) λi ¡ −(λi +µi )t ¢
− − e − e(λi +µi )t0 (14.12)
λi + µi [λi + µi ]2
Figure 14.2 shows the evolution of service unavailability, service availability,

and the expected number of repairs as a function of time. As expected, the
FIGURE 14.2. Availability, unavailability and expected number of repairs.

availability of a repairable system does not go to zero, but rather levels off at a
certain time and stays constant.
MAINTENANCE AND TOTAL COST
The maintenance cost is for simplicity related to the expected number of

repairs, that is
X
c M (t1 , t2 ) = 3i (t1 , t2 )c R,i (14.13)
∀i
where c R,i is the cost of repairing sensor i, which can be broken down into
spare parts cost and manpower costs.
If the life cycle is n years, the present value of the life cycle maintenance
cost can be calculated as follows:
X
n
c M ( p − 1, p)
c M,0 = (14.14)
p=1
(1 + I ) p
where I is the interest rate and p is the operating time of the instrument in years.
Therefore, the total cost of instrumentation can be stated in terms of its
purchase cost and the present value of the maintenance cost.
cT = c P + c M,0 (14.15)
Other formulations, such as perpetuity formulas or average yearly cost that

include depreciation, can be developed. We refer the reader to textbooks such
as Peters and Timmerhaus (1991).
RESIDUAL PRECISION
In attempting the design of sensor networks, one may ask each key variable
to have a certain degree of estimability (Chapter 6) or a certain estimation
availability to avoid solutions consisting of nonredundant networks. This can
also be achieved easily by requesting the variance of estimates of measured
variables to be lower than the variance of the measurement.
In many cases, measurements contain biases, and, therefore, they can no
longer be used. Thus, their corresponding measurement are eliminated from
further consideration. Once a redundant measurement is found to have a bias
and is eliminated, the overall degree of redundancy is lower, and the precision
Residual Precision 285
of all variables decreases. It is desired that the precision of certain key vari-
ables remain above the specified threshold on any deletion of gross errors. This
motivates the following definition (Bagajewicz, 1997).
Definition: Residual precision of a sensor is defined as the ability of the
network to guarantee a certain level of precision in key selected variables when
gross errors are detected and the measurements are eliminated.
Because many gross errors can be found and, therefore, the corresponding
variables are eliminated, the residual precision order is defined next.
Definition: A variable has residual precision of order k, ψi (q, k), when a
certain level of precision ψi∗ (k) is guaranteed after k gross errors are detected
and the measurements are eliminated, regardless of their position in the network.
Note that these concepts were introduced originally by Bagajewicz (1997)
under the name “availability.” They are being renamed in this book so that the
term “availability” is reserved for the concept introduced in Chapter 10.
The residual precision can be obtained as follows: Consider a fixed set of
measurements given by the binary vector q. Let tk be another binary vector of
the same dimension as q defined as follows:
½
1 If variable i is eliminated
tk (i) = (14.16)
0 otherwise
The vector tk has the following properties:

X
tk (i) = k (14.17)
∀i
q(i) − tk (i) ≥ 0 (14.18)
The first condition indicates that only k elements of tk are nonzero. The second
indicates that only measured variables can be signaled in tk to be nonzero. We
now proceed to provide a way of calculating ψi (k). As k measurements are
deleted, the new measurement vector is q − tk , and, therefore, a new precision
for each variable σi (q − tk ) is obtained. Because this precision varies with the
choice of the measurements deleted, then the maximum precision should be
picked. Thus,
ψi (q, k) = Max σi (q − tk ) (14.19)

∀tk
Remark 1: If the residual precision requested is equal to the bound requested

on deletion, that is, ψi∗ (k) = σi∗ , the constraint on precision can be dropped,
as the former contains the latter.
TABLE 14.1. Solutions of the Residual

Precision Constrained Problem
x1 x2 x3 x4
1% 2% 2% —
1% 2% — 2%
Example 14.1
Consider Example 8.1, with the same costs. Two nonredundant solutions of
cost cT = 3000 are shown in Table 8.1. Consider now that residual precision
of order k = 1 is added to variables S1 and S4 as follows: ψ1∗ (1) = 1.5% and
ψ4∗ (1) = 3%. The solution is: x M = (2%, 3%, 3%, 3%) with cost cT = 3900.
Assume now that residual precision is requested to the same level as precision.
Then two alternative solutions with cost cT = 5500 are obtained (Table 14.1).
Not only the cost is higher but also there is one more degree of redundancy.
For larger problems, the number of alternatives will increase, requiring new
criteria to further screen alternatives.
MINIMUM COST MODEL
We now generalize the minimum cost model introduced in Chapter 6. The

model minimizes total overall cost, including maintenance cost, and includes
all the desired constraints on estimability and precision. In addition, it also
features constraints on availability.
X 
Min (c P,i + c M,i )qi 



i∈M1 

s.t. 



E j (q) ≤ E ∗j ∀ j ∈ ME 
σ j (q) ≤ σ j∗ ∀ j ∈ Mp  (14.20)

ψ j (q, k j ) ≤ ψ ∗j (k j ) ∀ j ∈ Mp 




Avj (q, t) ≥ A∗j ∀ j ∈ MA


qi = {0, 1} ∀i ∈ M1
where M A is the set of variables for which availability constraints are im-
posed.
A simplified version of this model has been considered by Sánchez and
Bagajewicz (1998). In previous chapters instruments were considered nonre-
pairable systems, and, therefore, instrument reliability was used to evaluate
variable reliability and build constraints. In the above minimum cost model,
reliability constraints are purposely omitted. When instruments are considered

repairable systems, the concept of availability of a variable is more adequate.
Availability constraints present a problem because availability is a function
of time. Thus, the constraint in model (14.20) should be considered as applied
to the limiting value for large time. Finally, a model for multiple instruments,
similar to model (8.5), is presented next:
m 
XX
ni


Min (c P,i,k + c M,i,k ) pi,k 



∀i 

k=1


s.t. 



E j ( p) ≤ E ∗j ∀j ∈ Is 



σ j ( p) ≤ σ j∗ ∀j ∈ Mp
ψ j ( p, k j ) ≤ ψ ∗j (k j ) ∀j ∈ Mp (14.21)


Avj ( p, t) ≥ A∗j ∀j ∈ 

MA 



X
nim




pi,k ≤ 1 ∀i ∈ M1 




m
k=1
pi,k = {0, 1} ∀i ∈ M1 , ∀k = 1, n i
Example 14.2
We now show a design of a sensor network for the simplified ammonia

network (Figure 6.5). All cutsets are shown in Table 6.1. In this example,
sensors for each stream may be selected from a set of three instruments with
different precision, purchase cost and failure rate. These data are included in
Table 14.2.
Maintenance corrective costs are evaluated considering spare part cost, Spc =
10, manpower cost, M pc = 40 for all instruments, life cycle of 5 years, annual
interest I = 6%. Constraints of precision, residual precision and availability
are included for only two flow rates. The limits on these variables are presented
in Table 14.3.
The repair rate of instruments, a parameter that is a characteristic of the
plant in consideration, has been varied between 1 and 20. The results of the
TABLE 14.2. Instrumentation Data.
No. 1 No. 2 No. 3
Purchase cost 350 250 200

Precision 1.5% 2.5% 3%
Failure rate
(failures /yr) 0.3 0.6 0.7
TABLE 14.3. Constraints of the Optimization Problem.
Precision Residual Precision Availability

Stream Requirements Requirements Requirements
S1 — — 0.9
S2 1.5% 2%
S5 2.5% 3%
S7 — — 0.9
optimization problem are presented for each case in Table 14.4.

In the first case, the repair rate is comparatively low. Consequently, the avail-
ability of instruments in the life cycle is also relatively low. To satisfy the avail-
ability of key variables, the optimal solution includes a set of six instruments.
Three of these instruments are of type 1, which are sensors of low failure rate,
high precision and high cost. For this reason, precision and residual precision
constraints are not binding.
When the repair rate is assumed to be 2, an optimal solution exists that consists
of five instruments. Two of these instruments are of type 1 and the rest are of
type 3. Consequently, the total instrumentation cost decreases. Furthermore, the
residual precision for variable S2 and the availability for variable S7 are binding
constraints.
A lower instrumentation cost is obtained for a repair rate equal to 4. Even
though sensors are located on the same streams as in the previous case, one
sensor of higher failure rate is installed on stream S8 . This occurs because the
repair rate is now higher, making the constraint on availability of variable S7
not binding.
The results of the last case show that the influence of availability constraints
TABLE 14.4. Optimization Results for the Simplified Ammonia Process

Flowsheet.
Residual
Instrument Precision(%) Precision (%) Availability
Repair Measured Precision (S 2) (S 2) (S 1)
Rate Variables (%) Cost (S 5) (S 5) (S 7)
1 S1 S4 S5 S6 S7 S8 3 1 1 1 3 2 2040.2 0.8067 0.9841 0.9021

1.2893 1.2937 0.9021
2 S4 S5 S6 S7 S8 33131 1699.8 0.9283 1.9712 0.9222
1.9928 2.0086 0.9062
4 S4 S5 S6 S7 S8 33133 1683.7 1.2313 1.9712 0.9636
1.9963 2.0086 0.9511
20 S4 S5 S6 S7 S8 33133 1775.2 1.2313 1.9712 0.9983
1.9963 2.0086 0.9969
Nomenclature 289
decreases for high repair rates. The cost increases because of the effect of
increasing the repair rate µ (from 4 to 20) in the maintenance cost model. In
this case only the residual precision of variable S2 is binding.
It can be concluded that the repair rate has a direct influence on the availability
of a variable. If the repair rate is high, the design follows the requirements of
precision and residual precision constraints. Thus, the availability of a variable
may be a binding constraint for lower repair rates. In this situation, cost may
increase because it is necessary to incorporate more instruments to calculate
the variable by alternative ways.
NOMENCLATURE

Ai∗ (t): variable estimation threshold
C M (t1 , t2 ): maintenance cost for the interval (t1 , t2 )
C M,0 : present value of maintenance cost
cP : purchase cost
C R,i : cost of repairing sensor i
cT : total cost
I: interest rate
K j: constants
m i (t): instrument repair rate
MA: set of key variables for avaliability
M pc: manpower cost
p: operating time
Pr {.}: probability
q: binary vector indicating whether sensors are located (1) or not (0)
in variables
ri (t): rate of repair of instrument
rmax : maximum rate of repairs
Si : stream i
Spc: spare part cost
tk (i): auxiliary binary vector being nonzero when variable i is
eliminated
Ui : unit i
Uis (t): sensor service unavailability
Uis,c (t): corrective maintenance sensor service unavailability

s, p
Ui (t): preventive maintenance sensor service unavailability
vi (t0 , t): instrument unconditional repair intensity
wi (t0 , t): instrument unconditional failure intensity
Greek Letters
λi (t): instrument conditional failure intensity

3i (t0 , t): instrument expected number of repairs
µi (t): instrument conditional repair intensity
σ: precision
ψi (q, k): residual precision of order k of variable i as a function of q
ψi∗ (k): residual precision threshold
REFERENCES
2306 (1997).
Henley E. J. and H. Kumamoto. Reliability Engineering and Risk Assessment. Prentice Hall, New
Jersey (1981).
Masterson J. S. Reduce Maintenance Costs with Smart Field Devices. Hydrocarbon Processing,
Jan. (1999).
Peters M. S. and K. D. Timmerhaus. Plant Design and Economics for Chemical Engineers. Fourth
Edition. McGraw-Hill, New York (1991).
Sánchez M. C. and M. Bagajewicz. On The Design of Robust and Reliable Sensor Networks. AIChE
National Meeting, Paper 243f., Miami (1998).
CHAPTER 15
Design of Robust Sensor Networks
INTRODUCTION
Instruments fail in many ways. One important noncatastrophic failure is bias,

also referred to as systematic error. When data reconciliation is performed,
systematic errors produce a smearing effect in all the data so that the whole
data set has to be discarded. Without data reconciliation, no smearing exists,
but there is no other way than direct inspection to determine this situation.
In previous chapters, different properties have been introduced so that redun-
dancy is guaranteed. Estimability constraints can be used as a means to directly
guarantee more than one way of calculating a variable. Similarly, precision
and residual precision constraints can be used to indirectly guarantee a certain
level of redundancy. However, redundancy requirements are not a direct way to
request that a sensor network behaves properly in the presence of gross errors.
In this chapter, two new sensor network properties dealing directly with the
presence of gross errors are presented, and a general model for the design of
robust networks is presented.
ORIGIN OF GROSS ERRORS
Steady-state data reconciliation is implemented in practice by using averages

of single instrument measurements. As outlined in Chapter 4, gross errors from
different sources are in this way lumped in a single value. These sources of
gross errors are:
—Instrument biases: These are consistent fixed value departures from the
averaged values of the signals. Drifting is also a bias.
291
292 DESIGN OF ROBUST SENSOR NETWORKS
FIGURE 15.1. Ammonia oxidation reactor feed. Reprinted from Process Plant Perfor-
mance: Measurement and Data Processing for Optimization and Retrofits, Frantisek
Madron, D. Sharp. 1992, Ellis Horwood Series in Chemical Engineering.
—Leaks: Leaks can be classified as predictable (tank evaporation, for exam-

ple) or unpredictable. Both are typically not considered in the plant models,
although the predictable ones can be incorporated.
—True outliers: These are occasional measurements that depart significantly
from all other measurements. If the number and size of these outliers is large
and one-sided, the averaged values may get distorted and confused with
biases (Figure 4.4b).
—Departure from steady state: When plants present small drifts and oscilla-
tions, averaged values treated as steady-state values do not reflect real plant
behavior. This phenomenon has multiple effects, not only on gross error de-
tection but also on reconciliation itself as well as on variance (tolerance)
estimation. Consider for example, the signal in Figure 15.1:
This signal corresponds to the flow rate of ammonia into a reactor of am-
monia oxidation by air. The fluctuations are a result of the control scheme.
Again, the average value of a certain number of measurements will approach
the real average only as the number of measurements is impracticably large.
When a small number of values are averaged, the large peak at t = 20 can
introduce a departure from the real average. In addition, when signals like
these are averaged, the variance used cannot be the instrument variance any
longer, and process variations should be allowed if the true average flow rate
is to be recovered. This question goes to the heart of the problem, which is
the assumption of steady state. Although methods to perform dynamic data
reconciliation exist, procedures to develop robust dynamic sensor networks
have not been yet developed. Therefore, we omit discussing this issue fur-
ther and concentrate on methods driven by the steady-state data reconciliation
needs.
Test for Gross Error Presence 293
GROSS ERROR HANDLING
The challenging task in data reconciliation is to

r Identify the existence of gross errors.
r Identify the gross errors location.
r Identify the gross error type.
r Determine the size of the gross error.
At least one method that allows the detection of the existence of gross errors
exists. Some of the methods for gross error identification are to a certain extent
capable of discerning the location and type. Very little work has been performed
to address departures from steady state.
After the gross errors are identified, two responses are possible and/or desired:
r Eliminate the measurement containing the gross error.
r Or correct the measurements or the model and run the reconciliation again.
The first alternative is the one implemented in commercial software, which

in general only considers biases. This leaves the system with a smaller de-
gree of redundancy and, as we saw, the precision of the reconciled variables
deteriorates.
In the next few sections, some elements of gross error detection as well as
size estimation are introduced. A more thorough treatment of the subject can be
found in the following books: Madron (1992), Narasimhan and Jordache (2000)
and Sánchez and Romagnoli (2000).
TEST FOR GROSS ERROR PRESENCE
Testing for gross errors is based on the statistical principles of hypothesis

testing. We present the three most popular tests.
Global Test
The null hypothesis H0 is that there is no gross error. Let r be the vector or
residuals of the material balances, that is, r = C R z, where z are the flow rate
measurements. Then, the expected value of r is:
E(r ) = 0 (15.1)
and the covariance matrix of r
Cov (r ) = C R Q R C RT (15.2)
where Q R is the covariance matrix of random measurement errors. In the ab-

sence of gross errors, the following variable:
¡ ¢−1
χm2 = r C R Q R C RT r (15.3)
2
follows a chi-square distribution with m degrees of freedom (χm,α ), where m is
the number of rows of C R . This number can be obtained before even performing
data reconciliation. If it falls within the interval of confidence, that is, if it is
lower than a certain critical value, then the null hypothesis is accepted. On the
other hand, if it is larger than the critical value, it is said that a gross error has
been detected.
The global test is performed as follows:
a. Compute χm2 .
2
b. Determine the threshold value χm,α of the chi-square distribution corre-
sponding to a desired level of confidence α (usually 90–95%)
c. If χm2 > χm,α
2
, then there is a gross error in the data set.
Remark: The global test cannot determine where the gross errors are or
how many are there.
Measurement Test
The measurement test is based on the vector of measurement adjustments (or

corrections)
eR
a = FR+ − F (15.4)
where a is the vector of measurement adjustments.

The test is based on the assumption that the random errors for measurements
are independently and normally distributed with zero mean. Under the null
hypothesis, H0 , the expected value of a:
E(a) = 0 (15.5)
and the covariance matrix of a:
eR
Cov (a) = Q (15.6)
Gross Error Detection in Dynamic Data Reconciliation 295
Thus, the following variable is expected to follow a normal distribution

N (0,1).
ai
ηi = q (15.7)
eR )ii
(Q
If no gross error is present, the above value should be lower than a critical
value. If α is the confidence level, then this critical value xα/2 , is obtained
directly from the normal distribution tables xα/2 .
The measurement test can then be summarized as follows: The ith measure-
ment contains a gross error with an error probability of α, if ηi > xα/2
Several studies, modifications and improvements have been proposed for this
test (Mah and Tamhane, 1982, 1987; Crowe et al., 1983)
Nodal Test
In the absence of gross errors the constraint residuals r follow an m-variate

normal distribution (m is the rank of C R ). Therefore,
ri
ξi = q¡ ¢ (15.8)
C R Q R C RT ii
follows a standard normal distribution, N (0,1), under H0 .

If ξi is larger than the critical value based on a confidence level α, then one
concludes that there is at least one gross error in the set of measurement that
participates in the corresponding node balance.
Other tests used specifically for gross error exists. Among the most popular
are the generalized likelihood ratio (Narasimhan and Mah, 1987, 1988), princi-
pal component tests (Tong and Crowe, 1995) and the Bonferroni tests (Rollins
and Davis, 1992).
GROSS ERROR DETECTION IN DYNAMIC

DATA RECONCILIATION
The detection of gross errors in the context of dynamic data reconciliation

shares some of the difficulties of tests based on steady state models. Narasimhan
and Mah (1988) proposed to apply the generalized likelihood ratio (GLR) to dy-
namic situations with small departures from steady state values. Kao et al. (1992)
studied the effect of serially correlated data on gross error detection. They
proposed composite test procedures based on window averages, pre-whitening
procedures and the generalized likelihood ratio. Rollins and Devanathan (1993)
suggested the use of unbiased estimation techniques, and finally Albuquerque
and Biegler (1996) also proposed gross error detection techniques as an ex-
tension of their discretization approach. There are many other contributions
that would be worth mentioning. However, because the design of robust sen-
sors capable of using techniques for dynamic data reconciliation has not been
addressed, we will not explore this field in more detail here.
INACCURACY IN GROSS ERROR DETECTION
The following are the types of errors that one can incur while doing the
hypothesis testing above proposed.
r Type I Error: A gross error is identified, while in reality there is none.

r Type II Error: A gross error exists but is not identified.
Of these two, the first one is relatively harmless, as an elimination of a variable

that does not contain a gross error results only in loss of precision. However,
when a gross error is not detected, smearing persists and all the results are
corrupted.
MULTIPLE GROSS ERROR IDENTIFICATION
The tests described above are suitable for the detection of one gross error.
However, when more gross errors exist, strategies are needed to identify them.
This area of multiple gross error detection, identification and estimation has
been the object of several articles in the last 20 years.
One of the first strategies proposed is serial elimination (Ripps, 1965), which
consists of coupling a certain test with an elimination strategy. If the test failed,
then a strategy is proposed to identify one or more variables, which are the
“most suspected ones.” The measurements of these variables are eliminated
and the test is run again. Commercial versions of this procedure (DATACON,
SIGMAFINE) eliminate one measurement at a time and use the measurement
test or similar. The procedure has been the object of several variations of the
original scheme (Nogita, 1972; Romagnoli and Stephanopoulos, 1980; Iordache
et al., 1985; Rosenberg et al., 1987; Bagajewicz et al., 2000).
Several other procedures exist. For a review of these the reader is referred to
the books by Madron (1992), Narasimhan and Jordache (2000) and Sánchez and
Romagnoli (2000). In addition, more recent work not covered in these books
includes articles by Bagajewicz et al. (2000), Bagajewicz and Jiang (1998,
2000), Sánchez et al. (1999), Jiang and Bagajewicz (1999), Jiang et al. (1999)
and Renganathan and Narasimhan (1999).
Sensor Network Error Detectability 297
GROSS ERROR SIZE ESTIMATION
Once the gross errors have been identified, it is desired to determine their
size. There are several methods that have been developed in recent years to
perform this. When one gross error is present, Madron proposed an expression
based on the statistical properties of r (C R Q R C RT )−1r. Basically, when the mea-
surements follow a normal distribution with no gross errors, r (C R Q R C RT )−1r
follows a central chi-square distribution with m degrees of freedom. However,
in the presence of gross errors, r (C R Q R C RT )−1r follows a noncentral chi-square
distribution χm2 (ρ). When one gross error is present, the noncentrality parameter
ρ is related to the gross error δi as follows:
¡ ¢1/2
σi2 − e
σi2
ρ(m, α, β) = δi (15.9)
σi2
where α is the level of confidence, σi2 is the variance of the measurement and
σi2 the corresponding variance of the estimate. The parameter β is the power of
e
the test and is related to the probability of making a type II error (γ ) through
β = 1 − γ.
We are interested in certain power characteristics that are directly related to
the probability of Type II error. Thus, one can look for threshold values of the
noncentrality parameter of the chi-square distribution, so that a certain power
of the test β is obtained. Thus, the threshold value of the gross error can be
obtained inverting Equation (15.9).
σi2
δi∗ (m, α, β) = ρ(m, α, β) ¡ ¢1/2 (15.10)
σi2 − e
σi2
Thus, δi∗ (m, α, β) is the smallest size of gross error that can be detected with
probability β. Typical values of β are 50% and 90%. Madron (1992) presented
tables for ρ(m, α, β) and an empirical expression for large m.
In the presence of multiple gross errors, the above formulas do not apply. Most
of the research in the last 20 years has been devoted to the task of identifying
and assessing the size of many gross errors. Several methods, some of which
are efficient, have been developed. Because there are no methods developed to
design sensor networks featuring capabilities of detecting more than one gross
error, such techniques are not covered here.
SENSOR NETWORK ERROR DETECTABILITY
We now introduce a property that is related to the response of the sensor

network under the presence of gross errors (Bagajewicz, 1997).
Definition: The ability of the network to detect k gross errors of a certain

adimensional size κ D or larger is called error detectability of order k.
The adimensional gross error is given by κ D = δi /σi . We now concen-
trate on the error detectability of order k = 1. As it was explained above, in
the presence of a gross error of size δi in variable xi , the objective function
Ä = (x − z)T Q −1 2
R (x − z) has a noncentral χm (ρ) distribution with noncentral-
ity parameter ρi , that is, the mean of Q is ρi + m. Having discussed this effect,
we are in a position to work out this concept backward. One can choose a
threshold size of gross error for each variable δi∗ and request those errors larger
than this threshold to be detected by the network with a certain confidence. If
one chooses one common adimensional value for all variables κ D , the error
detectability criterion for variable i becomes
σi
ρ(m, α, β) ≤ κD (15.21)
(σi − σ̃i )1/2
Mathematical expressions capable of assess error detectability of larger order

have not been developed yet.
SENSOR NETWORK GROSS ERROR RESILIENCE
A sensor network should be capable of preventing undetected gross errors

from corrupting the reconciled values. This motivates the following definition:
Definition: The ability of the network to limit the smearing effect of k unde-
tected gross errors of a certain adimensional size κ R or lower is called gross
error resiliency of order k.
In the event that a gross error of a certain magnitude in any variable occurs
and is not detected, a certain corruption of data will take place after the recon-
ciliation is performed. Let 1xi = xi − z i be the change of the reconciled value
of variable xi when a set of gross errors of adimensional size κ R are present in
the network. Let Mδ be a set of k gross error locations. Then δ R , the vector of
gross errors, is defined as follows:
X
δR = κR σi ei (15.22)
i∈Mδ
Thus, replacing δ R in Equations (7.21) and (7.23), one obtains:

( ¡ ¢−1
−Q R C RT C R Q R C RT C R δR if xi is measured
1xi = (15.23)
C R δR otherwise
Robust Sensor Networks 299
Then, a desired level of resiliency is fixed for variable xi , requesting that 1xi
is lower than a certain threshold ri∗ . Sometimes this threshold can be quantified
in cost terms (e.g., loss of product). Thus, the resilience criterion becomes:
τi (k, κ R ) ≤ ri∗ (15.24)
where
¯ Ã !¯
¯ ¡ ¢ X ¯
¯ T T −1 ¯
τi (k, κ R ) = Max ¯ei Q R C R C R Q R C R C R
T
σi ei ¯κ R (15.25)
∀Mδ ¯ ¯
i∈M δ
if variable xi is measured, and

¯ Ã !¯
¯ X ¯
¯ T ¯
τi (k, κ) = Max ¯ei C R σi ei ¯κ R (15.26)
∀Mδ ¯ ¯
i∈M δ
otherwise.
ROBUST SENSOR NETWORKS
One might be tempted here to give a definition of what is a robust sensor

network. In principle, such a network is one that features meaningful values
of precision, precision availability, variable availability, error detectability and
resilience. These five properties encompass all the most desired features of a
network. Indeed, precision and precision availability guarantee that data are
always of the desired quality. Variable availability controls redundancy and
reliability and allows the data to be most of the time at hand. Finally, error
detectability and resilience make sure that the probability of data being gross
error free is large. Although the two first properties are deterministic and depend
directly on the quality of the instrument selected, the other three are statistical
in nature.
A formal definition of sensor network robustness is intentionally not given,
because many other features that a sensor network must have could be added.
For example, by the time this book is being edited, neural networks, wavelet
analysis, principal component analysis (PCA), partial least squares (PLS) and
other techniques for process monitoring have been developed for process mon-
itoring. For all these methods, very little analysis about goals that a sensor
network design or upgrade procedure can undertake was done. It is expected
that in the next few years, such methods will emerge. Thus, to maintain the
term alive, and just because it is ambitious, we simply define a robust sensor
network as one in which all the existing network performance measures are
below desired threshold values.
MINIMUM COST MODEL FOR ROBUST NETWORKS
We now generalize the minimum cost model given by Equation (14.37)

adding resilience and error detectability.


m
XX
ni

Min (C P,i,k + C M,i,k )Pi,k 



∀i 

k=1 

s.t. 



E j ( p) ≤ E ∗j ∀ j ∈ Is 



σ j ( p) ≤ σ j∗ ∀ j ∈ M p 





ψ j ( p, k j ) ≤ ψ ∗j (k j ) ∀ j ∈ M p 
Avj ( p, t) ≥ A∗j ∀ j ∈ M A (15.27)


σi 

ρ(m, α, β) 1/2 ≤ κ D ∀ j ∈ M D


(σi − σ̃i ) 

∗
τ j (k, κ R ) ≤ r j ∀ j ∈ M R 





Xn m



i
pi,k ≤ 1∀i 



k=1 


pi,k = 0,1 ∀i, ∀k = 1, n i m
Example 15.1
We now concentrate on adding error detectability to Example 14.1. We will

show that error detectability increases the precision of the sensor network and
can even make the design infeasible. Consider adding an error detectability of
κ D = 3.9 (with γ = 50%) to the availability constrained problem of Example
14.1. Two solutions from a set of only four feasible solutions are found with
cost cT = 4800 (Table 15.1).
If an error detectability of κ D = 3.4 for all measurements is requested, the
problem has only one solution, namely x M = (1%, 3%, 1%, 1%), with cost
cT = 8300.
Consider now the addition of resilience. If error detectability is requested
at a level of 3.9 times the standard deviation of the measurement for all
TABLE 15.1. Effect of Error Detectability

Constraints (η D = 3.9).
x1 x2 x3 x4
1% 3% — 2%
1% 3% 2% —
Minimum Cost Model for Robust Networks 301
measurements and resilience is requested at a level of 3 times the standard devia-

tion for all flow rate measurement, then the solution is again x = (1%, 3%, 1%,
1%), with cost cT = 8300. Relaxing (increasing) the resilience levels maintain-
ing the error detectability at the same level may actually lead to solutions of
higher cost, even to infeasibility.
Example 15.2
Consider a network proposed by Madron and Veverka (1992) (Figure 15.2). It

is of larger size and contains several measured variables. Originally the problem
was posed as selecting what measurements should be added to make streams
S1 and S5 observable. Meyer et al. (1994) added the costs of Table 15.2. We
will consider the flow rates of Table 15.3.
Madron and Veverka report the solution x M = {S2 , S4 , S8 }, with a cost cT =
35. Meyer et al. (1994) reported the same solution using their own cost data.
If the requirement of precision is sufficiently relaxed, the same solution is
found, that is, x M = (S2 , S4 , S8 ). When flowmeters of 2.5% precision are used,
this network has a precision of 27.8% in stream S5 . Because this value is
too high, resulting in poor monitoring performance, consider requiring a 2.5
precision on stream S1 and S5 . The solution is x M = {S2 , S4 , S5 , S8 }, with a
cost cT = 60. However, the network is nonredundant so it does not have error
detectability. If error detectability is requested at a level of κ D = 3.9 (with
γ = 50%) for flows S1 and S3 , which are the larger flows, the solution obtained
is x M = {S1 , S2 , S4 , S5 , S6 , S7 } with a cost of cT = 90.
TABLE 15.2. Cost of Flowmeters.
Stream 1 2 3 4 5 6 7 8 9
Cost 19 17 13 12 25 10 7 6 5
Source: Meyer et al. (1994).
TABLE 15.3. Flow Rates.
Stream Flow Stream Flow
1 140 13 10
2 20 14 10
3 130 15 90
4 40 16 100
5 10 17 5
6 45 18 135
7 15 19 45
8 10 20 30
9 10 21 80
10 100 22 10
11 80 23 5
12 40 24 45
NOMENCLATURE
a: vector of adjustments (difference between measured and

reconciled values)
Aiv ( p, t): variable estimation availability as a function of p
Ai∗ (t): variable estimation threshold
c M,i,k : cost of sensor k for variable i
cP : purchase cost
cT : total cost
Cov(.): covariance
ei : unit vector
E(.): expected value
FR+ : measured flow rate of stream i
e
FR : estimated flow rate of stream i
MA: set of key variables for avaliability
Mδ : set of gross error locations
N (0,1): normal distribution
pi,k : binary vector indicating whether sensor k is located (1) or not (0)
in variable i
References 303
q: binary vector indicating whether sensors are located (1) or not (0)
in variables
Q R: variance matriz of redundant measurements
eR :
Q variance matriz of estimates
r: residuals
Si : stream i
xα/2 : threshold value for measuremnt test
Greek Letters
α: confidence level for hypothesis testing

β: auxiliary variable defined by Equation (15.25)
γ: probability of committing Type II error in hypothesis testing
δ: gross error
δ∗: smallest size of detectable gross error
δD : threshold for error detectability
δR : adimensional gross error
ξ: nodal test statistics
φ: noncentrality parameter of the noncentral chi-squared distribution
χm2 : chi-squared distribution
2
χm,α : chi-squared distribution threshold
κD : adimensional gross errors for gross error detectability
κR : adimensional gross errors for gross error resilience
η: measurement test statistics
σ: precision
ψi (q, k): residual precision of order k of variable i as a function of q
ψi∗ (k): residual precision threshold
REFERENCES
Albuquerque J. S. and L. T. Biegler. Data Reconciliation and Gross-Error Detection for Dynamic
Systems. AIChE J., 42, 10, pp. 2841 (1996).
2306 (1997).
Bagajewicz M. and Q. Jiang. Gross Error Modeling and Detection in Plant Linear Dynamic Rec-
onciliation. Computers and Chemical Engineering, 22, 12, pp. 1789–1810 (1998).
Bagajewicz M. and Q. Jiang. A Mixed Integer Linear Programming-Based Technique for the Estima-
tion of Multiple Gross Error in Process Measurements. Chemical Engineering Communications,
177, pp. 139–155 (2000).
Bagajewicz M., Q. Jiang and M. Sánchez. Removing Singularities and Assessing Uncertainties in
Two Efficient Gross Error Collective Compensation Methods. Chemical Engineering Commu-
nications, to appear (2000).
Matrix Projection. I. The Linear Case. AIChE J., 29, pp. 818 (1983).
Iordache C., R. Mah and A. Tamhane. Performance Studies of the Measurement Test for Detection
of Gross Errors in Process Data. AIChE J., 31, pp. 1187 (1985).
Jiang, Q. and M. Bagajewicz. On a Strategy of Serial Identification with Collective Compensation
for Multiple Gross Error Estimation in Linear Data Reconciliation. Ind. & Eng. Chem. Research,
38, 5, pp. 2119–2128 (1999).
Jiang, Q., M. Sánchez and M. Bagajewicz. On the Performance of Principal Component Analysis in
Multiple Gross Error Identification. Ind. & Eng. Chem. Research, 38, 5, pp, 2005–2012 (1999).
Kao C. S., A. C. Tamhane and R. S. H. Mah. Gross Error Detection in Serially Correlated Process
Data 2. Dynamic Systems. Ind. & Eng. Chem. Research, 31, pp. 254 (1992).
Madron F. Process Plant Performance, Measurement. Data Processing for Optimization and
Models. AIChE J., 38, 2, pp. 227 (1992).
Mah R. S. H. and A. C. Tamhane. Detection of Gross Errors in Process Data. AIChE J., 28, 828
(1982).
Mah R. S. H. and A. Tamhane. Detection of Gross Errors in Process Data. AIChE J., 33, pp. 1514
(1987).
Meyer M. J., M. Le Lann, B. Koehret and M. Enjalbert. Optimal Selection of Sensor Location on a
Complex Plant Using a Graph Oriented Approach. Computers and Chemical Engineering, 18,
Suppl., pp. S535–S540 (1994).
Narasimhan S. and C. Jordache. Data Reconciliation & Gross Error Detection. Gulf Publishing
Company, Houston, Texas (2000).
Narasimhan S., and R. S. H. Mah. Generalized Likelihood Ratio Method for Gross Error Identifi-
cation. AIChE J., 33, 9, pp. 1514–1521 (1987).
Narasimhan S. and R. S. H. Mah. Generalized Likelihood Ratios for Gross Error Identification in
Dynamic Processes. AIChE J., 34, pp. 1321 (1988).
Nogita S. Statistical Test and Adjustment of Process Data. Ind. Eng. Chem. Process Des. Develop.
2, pp. 197 (1972).
Renganathan T. and S. Narasimhan. A Strategy for Detection of Gross Errors in Nonlinear Processes.
Ind. & Eng. Chem Research, 38, 6, pp. 2391 (1999).
Ripps D. L. Adjustment of Experimental Data. Chem. Eng. Progr. Symp. Ser. No. 55, 61, pp. 8–13
(1965).
Rollins D. K. and J. F. Davis. Unbiased Estimation of Gross Errors in Process Measurements.
AIChE J., 38, 4, pp. 563–572 (1992).
Rollins D. K. and S. Devanathan. Unbiased Estimation in Dynamic Data Reconciliation. AIChE
J., 39, 8, pp. 1330 (1993).
Chemical Plants. Chem. Eng. Sci., 35, 5, pp. 1067–1081 (1980).
Rosenberg J., R. S. H. Mah and C. Iordache. Evaluation of Schemes for Detecting and Identifying
Gross Errors in Process Data. Ind. & Eng. Chem. Research, 26, pp. 555 (1987).
References 305
Sánchez M. and J. Romagnoli. Data Processing and Reconciliation for Chemical Process Opera-
tions. Academic Press, San Diego, California (2000).
Sánchez M., J. Romagnoli, Q. Jiang and M. Bagajewicz. Simultaneous Estimation of Biases and
Leaks in Process Plants. Computers and Chemical Engineering, 23, 7, pp. 841–858 (1999).
Tong H. and C. M. Crowe. Detection of Gross Errors in Data Reconciliation by Principal Compo-
nent Analysis. AIChE J., 41, 7, pp. 1712 (1995).
CHAPTER 16
Genetic Algorithms
INTRODUCTION
The tree enumeration procedure presented in Chapter 6 can be used for the so-
lution of all the sensor network design models. However, this procedure works
reasonably well for relatively small problems, but it cannot handle large systems.
Recently, Chmielewski et al. (1999) proposed an alternative method based on
traditional MILP techniques. However, this chapter will cover recent advances
in the use of genetic algorithms to solve the problem. Other approaches, such
as special tailored MINLP solving techniques or the investigation of neces-
sary conditions of optimum to reduce the integer search space, have not been
investigated.
GENETIC ALGORITHMS
If the design problems are presented using binary variables, as in model (6.3),
then the standard genetic algorithm (GA), as described in Goldberg (1989) can
be applied. Sen et al. (1998) proposed a graph-based genetic algorithm for the
design of sensor networks.
The graph-theoretic operation used for the design of genetic operators is
an elementary tree transformation or cyclic interchange (Deo, 1974), which
generates a new spanning tree from an initial spanning tree by adding a chord
and deleting a branch belonging to the fundamental circuit of the chord that is
added. The genetic operators of crossover and mutation are implemented using
this operation.
307
308 GENETIC ALGORITHMS
Selection
As in the case of a standard genetic algorithm, one starts with an initial pop-
ulation of randomly chosen spanning tree solutions of the process graph. The
operator for selecting the mating pool is the same as in a standard GA imple-
mentation where the selection is based on the value of the objective function
(fitness) for each spanning tree. The tournament selection strategy was used by
Sen et al. (1998).
Crossover
Consider two spanning trees T1 and T2 that have one or more chords not in
common. Let cl be a chord of T1 and a branch of T2 . The following property
can be proved (Deo, 1974):
r Let bl be one of the branches of T1 that is also an element of the fundamental
circuit formed by cl . Then bl must be a chord of T2 . (A circuit is a sequence
of connected edges such that the first and the last edge are the same.)
The crossover operation is defined as the cyclic interchange of bi and c j
(satisfying above properties) in both trees T1 and T2 . The crossover operation
results in two new spanning trees (offspring) that have one more chord in
common between them compared with T1 and T2 . Because a chord set uniquely
defines a spanning tree, the “transfer” of a chord from one spanning tree to
another essentially transfers the “characteristics” of the spanning tree, which is
the objective of the crossover operator in GA. The exact implementation of the
crossover operator between two spanning trees T1 and T2 is as follows:
—Step 1. Find the chords not common between T1 and T2 .
—Step 2. Select, in sequence, an uncommon chord of T1 , say c j , identified in
step (1) and find a branch bi satisfying property (2) described above.
—Step 3. Generate a uniform random number between 0 and 1. If this random
number exceeds pc , the crossover probability, then interchange bi and c j for
both spanning trees. Update the chord sets of the spanning trees and their
fundamental circuits.
—Step 4. If all chords identified in step 1 have been tried, then stop. Otherwise,
return to step 2.
The two spanning trees obtained at the end of the above procedure are the
offspring produced by crossover. The above operation ensures that both the
offspring are feasible spanning trees.
As an example, consider the two spanning trees of the graph of Figure 11.1,
shown in Figure 16.1. These spanning trees are represented by the chord sets
T1 = (S2 , S5 , S6 ) and T2 = (S6 , S7 , S8 ).
FIGURE 16.1. Crossover operation: parent solutions.
The number of chords in each set not common for the two trees is equal
to two. Indeed, the set of chords in T1 and not in T2 is (S2 , S5 ), whereas the
set of chords in T2 but not in T1 is (S7 , S8 ). Let chord S2 of T2 be selected and
approved as a candidate for crossover. The fundamental circuit formed by chord
S2 in T1 is (S2 , S3 , S4 , S8 , S1 ). It can be deduced from the two spanning trees
that the chord-branch pair (S2 , S8 ) of tree T1 occurs as a branch-chord pair in
T2 . Thus, by interchanging S2 and S8 in both trees, one obtains two new trees
T1 = (S5 , S6 , S8 ) and T2 = (S2 , S6 , S7 ) as shown in Figure 16.2.
It can be observed that although the parents have two uncommon chords, each
of the offspring has only one chord uncommon with either of their parents. Thus,
both the offspring are a cross between their parents, having some chords drawn
from each parent.
Mutation
A cyclic interchange is also used to implement the mutation operation. For

each fundamental circuit of each spanning tree, depending on the mutation
FIGURE 16.2. Crossover operation: children solutions.

FIGURE 16.3. Mutant spanning tree.
probability, a cyclic interchange with a randomly selected branch of the funda-

mental circuit is carried out.
For example, if for chord S6 of the spanning tree shown in Figure 16.1(a),
the mutation probability is exceeded, then a randomly selected branch from
its fundamental circuit (S6 , S1 , S2 , S3 , S4 , S7 ), say branch S2 , is selected for
interchange. The chords of the resulting tree after mutation are (S2 , S5 , S8 ) as
Both crossover and mutation operators allow adequate propagation of build-
ing blocks and help genetic algorithms to quickly find the optimal solution
without digressing to finding many infeasible solutions (Goldberg, 1989).
Extensions to the Algorithm
The algorithm proposed in the previous section is useful for designing a sensor
network that optimizes any one selected criterion. In practice, it is necessary
to optimize several performance measures simultaneously. Sen et al. (1998)
suggested the use of the modified genetic algorithms for multiobjective opti-
mization proposed by Srinivas and Deb (1995). In addition, the above crossover
and mutation operators are useful for the design of nonredundant networks.
Objective Functions
Different objective functions can be used for the design of these minimal
networks. The four shown in Table 16.1 will be tried in the example below:
For the case of maximum precision, ki is the number of occurrences of chord
i in all the fundamental cutsets of the spanning tree.
Example 16.1
The above algorithm was tested on the example of a steam-metering net-

work shown in Figure 11.9. The data for this network were obtained from
Sen et al. (1998), who in turn quote Bansal et al. (1994) and is reproduced in
TABLE 16.1. Objectives Used in

Conjunction with GA.
P
Cost ci
i∈chords
Q ¡ ¢
System reliability 1 − Ris
i∈chord
Maximum branch reliability Min v

i∈branches Ri
P
Maximum precision (ki + 1)σi2
i∈chords
Table 16.2. The parameters of the GA used in all simulation runs are given
in Table 16.3. Although these parameters were not optimized, they were suit-
ably selected after a few trial runs. These parameter values maintained the
diversity of the population and prevented premature convergence to a subopti-
mal solution.
The GA was applied to optimize the sensor network for each of the above
four objectives, individually. A summary of the simulation results is given in
Table 16.4.
The first column gives the objective function used in the simulation. The
second column shows the chords of the spanning tree corresponding to the best
solution obtained in the final population. The last four columns give the values
of the four criteria for this solution.
For all the four objective functions, the best solution obtained corresponds to
the global optimum as verified through explicit generation of all 106 spanning
trees of the process graph. In the implementation of the GA, about 50 × 10 =
500 spanning trees are evaluated, which correspond to less than 0.05% of the
total number of feasible solutions. Furthermore, for the objective functions of
cost, system reliability and accuracy, there is a unique global optimum solution,
whereas for the minimum branch reliability objective five different spanning
tree solutions correspond to the global optimum. Despite the fact that there is a
unique (or few in the case of minimum branch reliability) global optimum solu-
tion, the GA has been able to identify it. It can also be observed that there exists
a trade-off between different criteria. The minimum cost solution (first row) has
a low reliability, whereas the one with maximum reliability has high cost and
low accuracy. This points to the need for multiobjective optimization to obtain
adequate trade-offs. In Chapter 8, a special type of duality embedded in all these
problems was pointed out. For example, a minimum cost network is a dual of a
maximum precision network. Likewise, the same type of duality exists between
cost and reliability. It is no coincidence that many of the solutions obtained have
the same value of these parameters, as seen in the last column in Table 16.4.
Sen et al. (1998) also compared the efficiency of the GA with graph-theoretic
algorithms for maximizing the minimum branch reliability or accuracy. To
TABLE 16.2. Data for Steam Metering

Network.
Standard
Failure Deviation of
Variable Probability Error Cost
S1 0.141 0.0215 3.7

S2 0.174 0.025 4.5
S3 0.104 2.8 132.2
S4 0.07 2.749 129.2
S5 0.174 1.332 65.3
S6 0.096 2.807 132.4
S7 0.164 0.058 5.0
S8 0.165 4.101 193.9
S9 0.055 0.0215 2.06
S10 0.099 1.31 62.8
S11 0.153 0.3715 20.2
S12 0.092 1.682 80.0
S13 0.052 2.782 130.4
S14 0.154 2.296 109.8
S15 0.095 1.5 71.6
S16 0.116 0.591 29.7
S17 0.077 0.8182 39.5
S18 0.08 0.4057 20.4
S19 0.099 0.1987 11.1
S20 0.074 0.2625 13.6
S21 0.071 2.182 102.9
S22 0.094 0.1362 8.1
S23 0.170 0.0648 6.3
S24 0.066 1.166 55.5
S25 0.088 2.136 101.0
S26 0.143 2.033 93.7
S27 0.128 1.769 84.7
S28 0.075 1.806 85.4
asses this efficiency, they used the number of sensor network design solu-
tions evaluated (function evaluations). The algorithm for maximizing accuracy
developed by Bansal et al. (1994) requires 30–70 function evaluations, de-
pending on the initial spanning tree used, whereas the algorithm by Ali and
Narasimhan (1993) for maximizing the minimum branch reliability requires
TABLE 16.3. Parameter Values of GA.
Population size 40–60

Number of generations 10
Crossover probability 0.45
Mutation probability 0.03
Tournament size 10% of population size
References 313
TABLE 16.4. Simulation Results of GA.
Performance
Objective Optimal System Minimum Branch

Criteria Spanning Tree Cost Reliability Inaccuracy Reliability
Cost S1 S2 S7 S9 S10 S11 555.5 0.141 85 0.265

S16 –S24 S27 S28
System S3 S4 S6 S9 S10 1178.2 0.235 267.86 0.265
reliability S12 S13 S15
S17 –S22 S24 S25 S28
Inaccuracy S1 S2 S7 S9 S10 S11 555.5 0.141 85 0.265
S16 –S24 S27 S28
Minimum S5 S8 –S11 S13 S15 1281.9 0.144 242.26 0.528
branch S17 S22 –S28
reliability
50 function evaluations. On the other hand, GA requires about 500 function

evaluations. However, GA also provides near optimal solutions, which may be
useful in practice, in addition to the fact that it is an algorithm applicable for
other objectives than reliability.
NOMENCLATURE
bx : branch x of a spanning tree

cp: chord p of a spanning tree
Si : stream i
T: spanning tree
Ui : unit i
Greek Letters
σ: precision
REFERENCES
Ali Y. Sensor Network Design for Maximizing Reliability of Processes. Ph.D. Thesis, Indian Institute
of Technology. Kanpur, India (1993).
Bansal P., Y. Ali and S. Narasimhan. Sensor Network Design for Linear Processes. Proceedings of
IFAC. Workshop on Integration of Process Design and Control. Baltimore, Maryland (1994).
Deo N. Graph Theory with Applications to Engineering and Computer Science. Prentice Hall,
Englewood Cliffs, NJ (1974).
Goldberg D. E. Genetic Algorithms in Search, Optimization and Machine Learning. Addison Wesley
Reading, (1989).
Sen S., S. Narasimhan and K. Deb. Sensor Network Design of Linear Processes Using Genetic
Algorithms. Comp. & Chem. Eng., 22, 3, pp. 385–390 (1998).
Srinivas N. and K. Deb. Multi-Objective Function Optimization Using Non-Dominated Sorting
Genetic Algorithms. Evolutionary Computation, 2 (3), pp. 221–248 (1995).
CHAPTER 17
Design of Sensors Networks for

Process Fault Diagnosis
INTRODUCTION
Alarm systems differ from normal monitoring systems. Although the basic
goal of monitoring systems is to provide a good estimate of the state of the sys-
tem, alarm systems are designed to detect process malfunction. In turn, process
faults, which typically have only one source in some unit, propagate throughout
the process, altering the readings of instruments (pressures, temperatures, flow
rates etc.). Thus, these sensors should be able to determine departures from
normal operation. In this sense, this task is different from that of gross error
detection, which concentrates on instrument malfunction instead. As a conse-
quence, the discrimination between instrument malfunction is an additional task
of the alarm system. Thus, the problem of designing an alarm system consists
of determining the cost-optimal position of sensors, such that all process faults,
single or multiple and simultaneous, can be detected and distinguished from
instrument malfunction (biases).
FAULT DETECTION, DIAGNOSIS AND ALARMS
A process fault is a departure from an acceptable range of operation or “degra-

dation from normal operation conditions, and includes symptoms of a physical
change (such as deviations in measured temperature or pressure) as well as the
physical changes themselves (scaling, tube plugging, etc.) and deviations in pa-
rameters (such as a heat transfer coefficient)” (Wilcox and Himmelblau, 1994a).
Faults originate in a process and propagate to a set of sensors (Figure 17.1).
These sensors are also subject to faults themselves. These faults on the sensors
are either biases or catastrophic failures. In the latter case, the fault detection
315
316 DESIGN OF SENSORS NETWORKS FOR PROCESS FAULT DIAGNOSIS
FIGURE 17.1. Fault diagnosis and alarm system.
is compromised, whereas in the former case the process fault can go either
undetected, or false alarms may be induced. Therefore, a good alarm system
should be able to filter the disturbances affecting the sensors and the gross
errors induced by their faults. The next step is the process fault detection itself,
although these two steps can be performed simultaneously. This procedure
needs also some capabilities of distinguishing process disturbances from real
faults. Once a process fault has been identified, the final step consists of taking
corrective actions or determining a shutdown of the process. This is performed
by implementing an alarm logic.
Fault detection and diagnosis have been addressed in several books (Him-
melblau, 1978; Pau, 1981; Gertler, 1998). We briefly list some of different
approaches, putting emphasis in those that have been later used for sensor
location methodologies.
Fault Trees
Fault trees have been used to build a library of diagnostic symptoms. A set
of undesired events (top events) is selected. For each top event, a digraph is
developed, modeling the interaction between variables, and later the fault tree
itself is built. Lapp and Powers (1977) presented an algorithm to build the fault
tree automatically. These trees are composed of AND and OR logical gates,
connecting the top events to their causes. We briefly illustrate the fault tree
analysis with a centrifugal pump (Figure 17.2). This system consists of a pump,
pressure and flow indicators, a check valve and a control valve.
Consider the event consisting of a backflow to the pump feed. The corre-
sponding fault tree is shown in Figure 17.3.
By assigning probabilities to events, failure rates, mean time to repair and
frequencies of events that are certain, the failure rate of the top event can be
FIGURE 17.2. Centrifugal pump system.
calculated by combining probabilities across logical gates, that is, P(A1 ∪ A2 )

for an OR gate (A1 OR A2 ) and P(A1 ∩ A2 ) for an AND gate (A1 AND A2 ).
The probabilities are calculated using the known formulas for probabilities of
union and intersection of sets.
In addition to the failure rate of the top event, the determination of minimal
cutset events, which are the events that alone can cause the top event, can
be performed. Qualitatively, the fewer the minimal cutsets and the larger the
number of members in each minimal cutset, the safer the process.
Fault trees have been used to perform risk analysis and suggest flow sheet
modifications. The reader can consult a variety of such tress for the pump
system of Figure 17.2 in Aelion and Powers (1993). Hazard identification
can also be performed by using a model-based approach (Catino and Ungar,
1995).
Signed Digraph Approach
Signed digraphs (SDG) are used to make cause effect (CE) analysis (Iri
et al., 1979; Umeda et al., 1980; Shiozaki et al., 1985; Wilcox and
Himmelblau, 1994a,b). Consider the system representing two tanks in series
FIGURE 17.3. Fault tree for backflow to feed event in Figure 17.2.
FIGURE 17.4. Two tanks in series.
presented in Figure 17.4, which is adapted from Wilcox and Himmelblau

(1994a). The digraph shown in Figure 17.4(b) is similar to the one presented
in Chapter 5 for a cyclohexane plant (Figure 5.1). The nodes of the digraph
are the units and the edges are the streams. To represent the level changes
(increasing in the first tank and decreasing in the second), pseudostreams are
added.
Consider now another type of digraph where the nodes represent the streams
of the process and the edges represent the influence that one exerts on the other.
Such a digraph is shown in Figure 17.5. The influence can be either positive or
negative, indicating that a positive change in one node will result on a positive
or negative in the other.
In this example, a positive change in a flow rate has a positive effect on the
level of the tank in which it discharges, that is, it will increase it, but it will have
a negative effect in the level of the tank from which it comes. This is represented
by the pluses and minuses (+ and −) on the edges of the digraph. The same
can be said about the level: It will have a positive effect on the flow rate of the
stream that is leaving the tank and no effect on the one discharging in it.
More formally, a signed digraph is obtained by adding a function that maps
the branches of the graph to the set {+, −}. Thus, in the SDG, the nodes of the
graph represent the immediate influences between variables, and the function
indicates the sign of the influences between variables. This representation of
positive and negative influences between variables in a process is due to Iri
et al. (1979, 1980).
FIGURE 17.5. Signed directed digraph.

TABLE 17.1. A Possible Pattern

for Two Tanks in Series.
Node n i F0 L1 F1 L2 F2
Pattern p(n i ) + + + 0 0
The state of a system is described qualitatively by a pattern p, which is a

function from the nodes of the graph to the set {+, 0, −}. A node mapping to
+, 0, or − indicates that the corresponding process variable is high, normal or
low, respectively. Table 17.1 describes a possible pattern for the two tanks in a
series.
In this pattern, one can see that the flow rate F0 , the level L 1 and flow rate
F1 are higher than normal. However, the level of tank 2, L 2 and its flow rate F2
remain normal.
Two additional definitions are required to understand how the signed digraph
is used for diagnosis: valid nodes and consistent branches.
Definition: A valid node in the signed digraph is any node which is abnormal
(i.e., p maps the node to + or −).
Definition: A consistent branch bk is a branch whose sign multiplied by the

signs of the initial and terminal nodes is positive.
For example, in the case of the pattern of Table 17.1, the branch connecting
nodes F0 and L 1 is consistent because the branch has a positive sign and the
two nodes also have a positive state. However, the branch connecting nodes F1
and L 2 is not consistent because the state of node L 2 is zero. Thus, a consistent
branch can be thought of as a consistent path for the propagation of the influence
of its initial node to its terminal node. On the basis of the state of the system
and the signal digraph (SDG), one can construct the cause-effect graph (CEG).
For the case of the two tanks in series, given the pattern in Table 17.1 and
the signed digraph of Figure 17.5, the corresponding CE graph is shown in
Figure 17.6.
The arc from F1 to L 1 is inconsistent on the basis of the assumption that an
“effect cannot compensate for its own cause” (Kramer and Palowitch, 1987)
and has been cut in the CE digraph.
To aid in the fault identification algorithms, a special type of node is defined
next.
FIGURE 17.6. Cause-effect (CE) digraph.

Definition: A maximally strongly connected component (MSCC) in the SDG

consists of a node or a cycle with no input arcs (Iri et al., 1979).
For example, in the case of the SDG of Figure 17.6, the node F0 is a maximally
strongly connected component.
Thus, in a CE digraph, the faults are assumed to originate from an element
in one of the MSCCs (Iri et al., 1979). This assumption that only nodes in the
maximally strongly connected components can be the root causes of faults is
a trivial assumption that can be made without any loss of generality (Raghuraj
et al., 1999). Indeed, if there is a fault node that is not a maximally strongly con-
nected component, then a pseudonode with just one output arc can be attached
to this fault node. Now in our treatment, this pseudonode becomes the fault
node corresponding to the original fault node. If the single-fault assumption is
valid (one fault at a time) (Iri et al., 1979; Kramer and Palowitch, 1987), then
a single MSCC exists in the CE digraph.
Every fault in the process has its own rooted digraph and valid nodes in the
SDG, representing the process variables that are influenced by the particular
fault. Thus, the problem of sensor location for fault observability can be viewed
as a systematic procedure of identifying the root nodes corresponding to all the
faults and placing the sensors on the measurable valid nodes in the SD graph. The
partial pattern obtained from these sensors then helps in detecting symptoms
of every fault with a given resolution.
In a series of articles, Iri et al. (1979), Umeda et al. (1980) and Shiozaki
et al. (1985) proposed the use of implicit enumeration to assign qualitative
states to each unmeasured node in the SDG. A depth-first search algorithm
(Tarjan, 1972) is used to locate maximum strongly connected components
(MSCC). If a single MSCC is found, the enumeration continues until all nodes
have a qualitative state assigned. If more than one MSCC is found, then the
enumeration backtracks to change the last qualitative state assignment. This pro-
cedure is computationally very expensive. Finally, Chang and Hwang (1992,
1994) developed a digraph-based approach for the synthesis of fault trees.
Rule-Based Approach
Rule-based algorithms are constructed by using SDG representations. Kramer

and Palowitch (1987) proposed the determination of “if-then” rules involving
tests of the measured variables only. These rules are developed off-line by us-
ing the SDG. Two rules are produced for each node, one testing if the +1 state
of the node is a failure and one testing the −1 state. This method has been,
nevertheless, challenged in several fronts (Mohindra and Clark, 1993). First,
some inconsistencies appear when handling control loops, which generate a
path from a nonzero state at a disturbance variable, through a zero state at
a controlled variable, to a nonzero state at the controller input. Second, little
guidance is given on how to remove unmeasured nodes, and finally, no proof is

given that the results are identical to those produced by other methods. Chang
and Yu (1990) addressed the same task by simplifying the SDG according to
states. Finally, Mohindra and Clark (1993) introduced the concept of path-SDG,
which consists of measured nodes only. This work addresses the issue of the
computationally expensive task of finding all MSCCs and the disadvantages of
the method proposed by Kramer and Palowitch (1987). They propose to find
single-node and single-loop strongly connected components by partitioning the
PSDG into “fragments.” A distributed reasoning strategy is used to determine
the faults.
Failure Propagation Networks
Kokawa and Shingai (1982) and Kokawa et al. (1983) proposed to construct
a digraph where the nodes are the devices and the arcs are the direction of
propagation of failure. Failure propagation time and failure probabilities are
also used. A backward search from the abnormal nodes to the source nodes
is used. However, the state of all nodes is assumed to be known and cycles
cannot be present. Picking up on the deficiencies of the work by Kokawa et al.,
Qian (1990) proposed to use a failure propagation graph using OR nodes and
arrows. The OR nodes represent the devices of the plant or the failure modes
of theses devices. The arrows show the failure propagation direction between
adjacent nodes. Branches are then classified into several types according to the
interaction between the nodes indicating the positive or negative abnormality
condition of the pointed node. Certain rules are then outlined to indicate how a
consistent branch can be identified so that backtracking from an abnormal node
is properly performed. Failure propagation time is not required, and the failure
probability is replaced by a degree of confidence.
Neural Networks
Neural networks have very good abilities to “learn” and recognize patterns.
Thus, a neural network can recognize deviations from a pattern of values of pro-
cess variables and detect and possibly diagnose a fault. Venkatasubramanian
and Chan (1989) compared such approach with the performance of a rule-
based methodology. Watanabe et al. (1989) used a two-stage neural network
efficient only for single faults. Fan et al. (1993) proposed orthonormal trans-
formations of the inputs, thus allowing the successful identification of multiple
fault causes and their severity. Finally, Watanabe et al. (1994) provided a long
list of applications of artificial neural networks for fault detection and diagnosis.
They also proposed a hierarchical artificial neural network structure for detect-
ing simultaneous “slight” faults with the same degree of deviation in the pro-
cess variable. Kavuri and Venkatasubramanian (1993a) used a hidden layer of
ellipsoidal activation functions to obtain bounded regions of process faults and

later fuzzy clustering techniques to improve the fault classification (Kavuri and
Venkatasubramanian, 1993b). Chen et al. (1998) used a wavelet transform to
extract features of the transients, which are later processed by a neural network.
Finally, variable structure neural networks have been recently proposed (Luan
et al., 1998).
Knowledge-Based Approach
Knowledge-based systems have also been proposed for fault diagnosis.

Kramer (1987) introduced nonboolean reasoning in expert systems. Venkata-
subramanian and Rich (1988) proposed an expert system that integrates com-
piled with deep level knowledge, which is in turn based on constraints (material
and energy balances), confluence equations (qualitative differential equations
that represent the influence of one variable on another), fault models to explain
local causes of faults and causal models of process units to indicate the local
effect of unit failures on variable. Fathi et al. (1993) presented an integrated ap-
proach of analytical redundancy obtained by using extended Kalman filters with
compensation and a knowledge base organized on the basis of malfunction hier-
archy. Su and Fan (1993) proposed a rule-based expert system integrated with a
neural network, which also makes use of wavelet decomposition. Finally, Leung
and Romagnoli (1998) proposed an expert system based on possible cause and
effect graphs (PCEG) with Bayesian belief networks and applied it successfully
to a distillation column.
Model-Based Approach
Quantitative models used for fault detection and isolation are usually
linear models. Several review articles and books have been devoted to cover
all the advances. Among the most important review articles are the ones
by Gertler (1988), Frank (1990) and Patton (1995) and the books by
Himmelblau (1978), Patton et al. (1989) and Gertler (1998), among others.
Quantitative modeling uses a dynamic model of the system being monitored
to generate estimates of measured and unmeasured variables and parameters. In
turn, estimates of the measured variables are contrasted with the measurements
generating error signals known as residuals. Patton (1995) offers the explanatory
diagram of Figure 17.7. The residual generation, the decision making and the
operator are assisted by a knowledge base.
There are four approaches to residual generation:
—Kalman filter: The prediction error of the Kalman filter can be used as a
residual. Because these are white, diagnosis can be performed by using sta-
tistical tests. This is the most common approach.
FIGURE 17.7. Quantitative model-based fault diagnosis. Adapted from Patton, 1995.
—Diagnostic observers: Observer innovations are used as residuals, which

are decoupled from disturbances with “unknown input” design techniques.
Residuals are not white, so statistical tests are somehow complex.
—Parity relations: These are rearranged input-output equations. Residuals
generated by this technique are also colored. These relations generate identi-
cal results as the observer-based techniques when the residuals are generated
for the same specification. (Gertler, 1998)
—Parameter estimation: According to Gertler (1998), this is the natural ap-
proach to fault detection and isolation. A reference model is used, identifying
the plant in a fault-free situation. Parameters that are reidentified on-line are
used to generate residuals. A weak relation between parity relations and
parameter estimation also has been found (Gertler, 1998).
The literature on this fault detection and diagnosis is really very large,
because this is one of the approaches that has received most attention. Be-
cause sensor location techniques have not been proposed using these tools,
this area is not covered in any more detail, and the reader is referred to the
review by Patton (1995) and the book by Gertler (1998) for additional
material.
Multivariate Statistical Methods
In process monitoring it is customary to arrange the measurement into two

sets of data, process variables (X) and product variables (Y). It is further as-
sumed that there is a relation between them. Principal component analysis
(Jackson, 1991) relies on an eigenvector decomposition, which usually finds
that data can be described with far fewer factors than the original variables, sim-
plifying the monitoring task. PCA models have been combined with univariate
statistical testing or scoring.
Multivariate regression modeling includes inverse least squares, principal
component regression and partial least squares (PLS) (Wise and Gallagher,
1995). PLS is similar to PCA, except that it simultaneously reduces the dimen-
sions of process and product variables (X and Y), whereas PCA only reduces
the process variables only. Of the three PLS is the most popular.
In both cases, the monitoring task consists of performing tests on the loadings
and latent scores determining if they are out of range (Kresta et al., 1991). Both
methods and their application to process monitoring are excellently reviewed
by Kresta et al. (1991) Wise and Gallagher (1995). A good example of PLS
use for fault detection is provided by Vinson and Ungar (1994). Finally, Gertler
et al. (1998) recently showed the connections between PCA and parity relations.
Extensions of these two techniques have flourished in recent years.
McGregor et al. (1994) used multiblock PLS methods to analyze processes
with subsections. Dunia and Qin (1998) recognize that each multidimensional
fault is described in a subspace on which the displacement occurs. They propose
a reconstruction procedure that enables fault isolation. Extension to nonlinear
cases have been presented by Qin and McAvoy (1992), who embedded neu-
ral networks into a PLS framework. Dynamic systems have been addressed by
Lakshminarayanan et al. (1997). Dong and McAvoy (1996) used nonlinear PCA
based on neural networks, and Jia et al. (1998) improved on the same idea by us-
ing nonlinear PCA based on input-training neural networks. Shao et al. (1998)
used a combination of wavelet filtering followed by nonlinear PCA, control
limit and contribution plots. Vedam and Venkatasubramanian (1998) proposed
to use signed directed digraphs to identify root causes of the error in PCA model
predictions. Finally, Qin (1998) proposed a recursive PLS method.
Other Techniques
In a recent workshop of IFAC (Lyon, France, 1998) some new approaches

to the problem have been proposed. Semiquantitative model-based techniques
using fuzzy rules were proposed by Ozyurt et al. (1998). Parallel evolutionary
algorithms of the genetic type were proposed to perform parameter
estimation to generate residuals (Marcu and Frank, 1998). Aggregative learn-
ing for neural networks was also presented (Akhmentov and Dote, 1998).
Statistical techniques and relationships between fuzzy systems and radial ba-
sis function networks were applied to prune a trained neural network and to
extract qualitative rules (Gomm et al., 1998). Casual reasoning using fuzzy
cognitive maps was used in a forward inference engine (Kim and Lee, 1998).
Singular value decomposition was proposed to obtain least squares residuals
(Han and Frank, 1998). Szigeti and Tarantino (1998) proposed to augment the
state space with instrument and control dynamics and proposed a filter based
on this model. A hierarchical procedure based on a triangular representation of
process trends (Cheung and Stephanopoulos, 1990) in combination with prin-
cipal component analysis was proposed by Cho and Han (1998). Stanley and
Vaidhayanathan (1998) studied the use of generic fault propagation in casual
directed graphs by an event correlator. Finally, Bensaker and Ouchene (1998)
proposed to perform parameter estimation by using Poisson moment functional
and generating residuals.
Role of Sensor Failure
Until recently, most of the literature downplayed the sensor fault detection
task associated with fault detection as a simultaneous filtering problem.
Dunia et al. (1996a,b) presented a PCA-based method to capture the correla-
tion between the sensors, which can be later used for sensor fault identification.
Busson et al. (1998) present an analytical residual generation approach for a
steam condenser that allows them to obtain signatures for sensor faults. These
attempts, however, do not seem capable of distinguishing a sensor fault from a
process fault.
Zafiriou and coworkers (Naidu et al., 1989; Yao and Zafiriou, 1990) discussed
the use of neural networks to pick up sensor failures in control systems. Negiz
and Cinar (1994) developed a multipass PLS technique that allows the elimi-
nation of incipient faulty sensors. Luo et al. (1998) presented a new approach
to sensor validation in real time. In this approach, the signal is represented by
wavelets, which are decomposed into different frequency ranges. The features
at different frequencies are analyzed by using statistical testing. The authors
claim that the physical changes in the sensor can be separated from the noise and
the process changes. Vedam et al. (1998) used a B-spline-based algorithm that
achieves data compression by denoising the data from each sensor. Later, they
perform a trend analysis, detecting abnormal frequencies, changes in correla-
tion among sensors and perform a root cause analysis. A knowledge base is used
to distinguish between process changes and abnormalities in the sensors. These
are very encouraging results, even though the techniques have only been tested
on a few simulated examples. Quite recently, Dunia and Qin (1998) presented
a unified PCA geometric approach for a unidimensional fault, which is capable
of distinguishing a sensor from a process fault and seems to be amenable for
generalization to multiple faults.
Although it is obvious that the use of the methodologies proposed throughout

this book, sensors networks can be designed to exhibit the needed accuracy,
reliability and appropriate gross error detection robustness, these properties are
not enough for the design of a fault detection network. No work has been yet
performed to design sensor networks combining these robustness features with
fault detection.
Alarm Logic
Some incipient work is starting to be performed in the designing of the

appropriate alarm logic. Tsai and Chang (1997) proposed a method that is
based on data reconciliation first and that minimizes the economic losses due
to Type I and Type II errors.
It is abundantly clear that the sophistication of tools that are being applied
to the detection and diagnosis of faults has far exceeded the humble level of
the optimal sensor location problem for the same purpose. We now proceed
to review the few existing results and hope that this fertile research ground
flourishes in the next years.
FAULT OBSERVABILITY
Observability of a fault refers to the ability of a sensor network to detect

its presence. Thus, a fault is observable if the sensor network can measure its
effects in locations that are not necessarily the location of the fault.
As in the case of the estimation of variables in steady-state processes, ob-
servability of faults can be achieved by using many different sets of sensors.
The design task consists of determining which combination is the most cost-
effective. In addition, in the same way it was done for variable estimation,
one may add reliability constraints and other filtering abilities to the problem,
such as the ability to distinguish disturbances from process faults, the ability to
identify gross errors in the instruments etc. Such model has not been presented
yet.
FAULT RESOLUTION
A fault can be “observed,” when its symptoms are observed. However, the
same type of symptoms can correspond to different faults. Thus, the term res-
olution has been coined to refer to the ability of determining the exact pro-
cess fault occurred given the set of symptoms. This is, therefore, another
constraint that can be added to the problem of designing a robust sensor
network.
SENSOR NETWORK DESIGN
In previous paragraphs some aspects of the problem of designing a cost-

effective sensor network for fault handling have been suggested. As outlined in
Chapter 2, the problem can be defined as follows:

s.t.

 Desired Observability of Faults


Desired Level of Resolution of Faults

 Desired Level of Reliability of Fault Observation


Desired Level of Gross-Error Robustness in the Sensor Network
The first attempt to present a technique to locate sensors was done by

Lambert (1977), who used fault trees based on failure probabilities. Failure
probabilities are hard to assess, fault trees cannot handle cycles and the con-
struction of the tree is cumbersome for large-scale systems. Because of these
limitations, the technique has not been developed further. We now concentrate
on a recent sensor location technique for fault observability.
SENSOR LOCATION FOR FAULT OBSERVABILITY
We start with the definition of the problem: Given a process DG, the sen-
sor location problem for observability is one of finding a set of nodes that is
connected to all the nodes with only output arcs (root nodes).
For the purpose of defining a strategy to positioning sensors, Raghuraj
et al. (1999) used directed graphs (DG), that is, graphs without signs. In this
way, the arcs of a DG represent a “will cause” relationship, that is, an arc from
node A to node B implies that A is a sufficient condition for B. This in general
is not true for an SDG. As an example, the effect of a fault on a connected vari-
able might be offset by the opposing effect of some other fault on that variable.
Another example might be the case of a controlled variable that would assume
a normal value at steady state, though there would be nodes with abnormal
values connected to it. Therefore, the arcs in the SDG represent a “can cause”
relationship, that is, an arc from node A to node B only implies that A can
cause B, and not that A will cause B. If the signs of the arcs were to be consid-
ered, then suitable modifications can be made to the algorithms presented by
Raghuraj et al. (1999). Because the notion of signs is not used, the concepts of
consistent and inconsistent cycles are also not considered. An important point
to note is that the consistent and inconsistent cycles do not really change the
CE behavior analysis of the process. A methodology for the optimization of
different possible sensor locations in the DG representing the process is the

focus of this section. The strategy used to solve the problem is based on the
following claim.
Lemma 17.1: In a DG that is weakly connected (i.e., the corresponding
undirected graph is connected) with no cycles, there is at least one directed path
from a root node (node with only output arcs) to one node of the observability
set.
Proof: Consider a node with only output arcs. Consider the longest directed
path from the node to some other node in the DG. Now, the last node in the
directed path should be a node with only input arcs; otherwise, it would not be
the longest path. Hence, every root node is connected through a directed path
to one of the nodes of the observability set. It is clear that this claim is valid
only for DG with no cycles. Q.E.D.
Procedure: Given a process DG, the observability problem is solved through
the following sequence of steps:
(1) As a first step in the algorithm, a cycle is located. All the nodes in the cycle
are collapsed into a supernode.
(2) This procedure is repeated until there are no more cycles in the DG. This
should make the DG acyclic.
(3) Choose all the nodes with only input arcs. This results in a set of sensors
that can observe all the faults. If one of the chosen nodes happens to be a
supernode, then any sensor that is a part of the supernode can be used.
To illustrate the algorithm, consider the digraph of Figure 17.8(a) (Raghuraj
et al., 1999). The reduced digraph in which the cycles have been eliminated is
shown in Figure 17.8(b). There are two cycles consisting of nodes N9 , N10 and
nodes N11 , N12 . Figure 17.8(b) shows the DG after the two cycles have been
collapsed to supernodes S N2 and S N1 , respectively.
Definition: The observability set is the set of nodes in the reduced graph
that have only input arcs.
From Figure 17.8, one can conclude that the observability set is given by three
nodes, namely [C6 , C7 , C8 ]. This set of sensors would ensure the observability
condition, but the set might not be minimal.
The possibility of obtaining observability sets that are not minimal is illus-
trated through Figure 17.9. In this figure, the nodes with only input arcs are
[C6 , C7 ]. Clearly, these nodes would not give the observability set, as root node
R1 cannot be observed with these two sensors. Now, if nodes N13 and N12 are
collapsed into a single node, then this supernode would also become a node
with only input arcs. Now, the observability set could be either [C6 , C7 , N12 ]
or [C6 , C7 , N13 ].
FIGURE 17.8. Collapsing of nodes in the digraph.
Minimum Set of Sensors Ensuring Observability
To generate a minimal set of sensors, a minimal subset of sensors from the

observability set that have at least one directed path from every root node needs
to be picked.
Clearly, enumeration is one way of exactly solving the problem, but with
an increasing number of nodes in the observability set and root nodes, the
problem will become combinatorially complex. Raghuraj et al. (1999) proposed
FIGURE 17.9. Illustration of nonminimal observability sets.

FIGURE 17.10. Bipartite graph.
a greedy search, which is later modified to remove redundant members of the

observability set. They do not guarantee optimality, that is, a minimum set may
not be found by this algorithm.
The greedy search consists of first drawing a bipartite graph between all the
nodes of the observability set (top row) and the root nodes. A bipartite graph is
one whose vertex set can be partitioned into two sets in such a way that each
edge joins a vertex of the first set to a vertex of the second set. Such bipartite
graph corresponding to Figure 17.8(b) is shown in Figure 17.10.
Now the problem is one of choosing the minimum number of sensors (nodes
in the top row) that would cover all the root nodes. This is the well-known
“minimum set covering problem” (Parker and Rardin, 1988). All the root nodes
are said to be “covered” if a directed path exists from every root node to at least
one of the nodes of the observability set.
Greedy algorithm: In this algorithm, the node of the observability set that
has the maximum number of arcs incident on it is first chosen. A check is made
to determine if all the root nodes are covered by the chosen component. If they
are covered, then this gives the minimum number of sensors for observability.
If some of the root nodes are not covered with this node, then all the arcs from
already covered root nodes to other nodes of the observability set are deleted.
After deletion of the arcs, the node with the maximum number of arcs incident
on it is chosen again. At this point, a check is made to see if all the root nodes
are covered.
The result of the application of the greedy algorithm to Figure 17.10 is shown
in Figure 17.11 where nodes [C6 , C7 , C8 ] form the observability set, because
these are the nodes in the DG with only input arcs. It is also clear that by
placing sensors on these nodes, one could detect all the faults. However, the
set [C6 , C7 , C8 ] is not the minimal set. The minimal set is, in fact, [C7 , C8 ],
which results from the application of the greedy algorithm. As a first step in
the algorithm, node C7 is chosen, because it has the maximum number of arcs
incident on it. In the next step, all the arcs other than the arcs from R5 to
C8 are cut because all the other root nodes are already covered by C7 . These
arcs are shown as dotted lines in the figure. Node C7 is marked, because this
node is already chosen. Now C8 is chosen as the next node, and with this it is
FIGURE 17.11. Result of the application of the greedy algorithm.
clearly seen that all the root nodes are now covered. This, then, is the minimal
observability set, as given by Algorithm 1.
This algorithm may not give the actual minimum number of sensors. To
explain this, let us consider Figure 17.12. By using the greedy algorithm, C1
would be chosen as the first key node, because it has the maximum number
of arcs incident on it (C1 covers the root nodes [ f 1 , f 2 , f 3 , f 4 ]). As the second
step, the arcs from [ f 1 , f 2 , f 3 , f 4 ] to [C2 , C3 , C4 , C5 ] are removed. One still
has to choose [C2 , C3 , C4 , C5 ] to cover root nodes [ f 5 , f 6 , f 7 , f 8 ]. Thus, the
greedy algorithm identifies [C1 , C2 , C3 , C4 , C5 ] as the minimal set. But clearly,
the minimal set is actually [C2 , C3 , C4 , C5 ]. The presence of a redundant com-
ponent, C1 , makes the observability set non-minimum. Although the node C1
was chosen first, the nodes chosen afterward cover all the roots nodes covered
by C1 , by themselves. Hence node C1 becomes redundant in the observability
set. To solve this problem, a backtracking procedure is used in the algorithm to
facilitate identification and removal of the key node.
Backtracking algorithm: In this algorithm, the key node with the maximum
number of arcs incident on it is chosen first and marked. All the arcs from the
root nodes covered by the selected node to all the previously marked nodes are
deleted. All the other arcs from the root nodes covered by the chosen node to
FIGURE 17.12. Illustration of the failure of the greedy algorithm.

unmarked nodes are stored in a buffer. Now, a node is selected again based on
the maximum number of incident arcs. The number of arcs that are incident on
a node is taken to be the difference between the actual number of arcs incident
on the node and the arcs incident on the node that are stored in the buffer.
This procedure is continued until all the root nodes are covered. At the end
of the algorithm, all the marked nodes that have no arcs incident on them are
removed from the minimal set. This ensures the removal of the redundant key
nodes. The algorithm is given in Figure 17.13 in terms of a flow chart (Raghuraj
et al., 1999).
To illustrate the use of this algorithm, consider again the example of
Figure 17.12. Denote an arc from Ri to C j as ai j . As a first step in the algorithm,
component C1 is chosen and marked. No arcs are deleted, because there is only
one marked node. Arcs a12 , a23 , a34 and a45 are stored in a buffer, because they
connect covered root nodes (R1 , R2 , R3 and R4 ). Now, all the other nodes have
one arc incident on them (two actual arcs minus one arc in the buffer). Choose
C2 . Root nodes R1 and R5 are covered by this sensor. At this stage arc a11 is
deleted, because it is connected to an already marked node from R1 , which is
a root node covered by the currently chosen node C2 . Similarly, arcs a21 , a31
a41 are deleted when components C3 , C4 and C5 are chosen, respectively. At
the end of the algorithm, nodes C1 , C2 , C3 , C4 and C5 would be chosen as the
minimal set. Now, because node C1 has no arcs incident on it, this node would
be deleted. Hence set [C2 , C3 , C4 , C5 ] would be identified as the minimal set.
It is obvious that although nodes are being removed at the end of the algorithm,
observability is always ensured.
SENSOR LOCATION FOR FAULT RESOLUTION
The minimum requirement for a fault-monitoring system is that it ensures

observability. In the previous section, this problem was formulated and solved.
For a fault-monitoring system to be useful, it should not only be able to observe
all the faults but also resolve them to the maximum extent possible. The reso-
lution is of course restricted by the topology of the digraph and the position of
the fault or root nodes in the digraph. Also, the assumption of single-fault or
multiple-fault would lead to different resolutions. In this section, the problem
of sensor locations for maximum resolution under single-fault and multiple-
fault assumptions is formulated and solved. The problem of sensor location for
maximum resolution under single-fault assumption could be formally stated as
follows:
Definition: A sensor network for maximum fault resolution is such that each
fault has one and only one set of nodes from which it is observable.
Assume Ai is the subset of nodes of the observability set such that its elements
are connected to the root node Ri . Assume now that for each Ri , a set Ci ⊆ Ai
FIGURE 17.13. Flowchart of backtracking algorithm.
is picked. By construction, the set C = ∪ Ci allows full observability of the

faults. The problem then consists of determining a minimal set C such that
Ci 6= C j whenever Ai 6= A j and at least one element from every Ai is a member
of C. It follows that two faults are indistinguishable if and only if Ci = C j .
Whenever they are not equal, there is at least one sensor that can be used to
distinguish the faults. We illustrate this as follows: consider three fault nodes
FIGURE 17.14. Maximum fault resolution.
R1 , R2 , and R3 (Figure (17.14). In this figure A1 = [C1 , C5 ], A2 = [C2 , C3 ],

and A3 = [C1 , C3 , C4 ].
Clearly, if only one fault is expected to occur at a time, then the set [C1 , C3 ]
would be adequate to distinguish between the three faults. Indeed, a fault in R1
is reflected in C1 , but not in C3 . Similarly, a fault in R2 is reflected in C3 , but
not in C1 . Finally, a fault in R3 is reflected in bot C1 , and C3 simultaneously.
To obtain the minimum set for maximum fault resolution, Raghuraj et al.
(1999) proposed the following procedure:
(1) Define Bi j = B ji = Ai ⊕ A j = Ai ∪ A j − Ai ∩ A j . There are n × (n −
1)/2 such sets.
(2) Denote each Bi j = B ji by a node. Draw a bipartite graph between this node
and the nodes in the set Bi j = B ji . Repeat this procedure for all the new
nodes generated.
(3) Without the observability condition, the minimum set of nodes that would
cover all the root nodes in the bipartite graph is the solution to the problem.
The sets Bi j represent the ring sum of the observability subsets of Ri . They
contain nodes that can observe either one of the roots Ri and R j , but not both
simultaneously. If the root node Bi j = B ji is covered, then faults Ai and A j can
be differentiated. If this procedure is repeated for all the faults pairwise, then
all the faults could be distinguished except for the ones that affect the identical
set of sensors. Our aim here is to choose a minimal set of sensors that would
perform single-fault identification.
Consider the example of Figure 17.15. Three sets are generated: B12 = B21 =
[C1 , C2 , C3 , C5 ], B23 = B32 = [C1 , C2 , C4 ], B13 = B31 = [C3 , C4 , C5 ].
The application of the backtracking algorithm renders the following: Node
C1 is chosen and marked. Thus, the connections from B12 to C2 , C3 and C5 as
well as the connections from B23 to C2 and C4 are stored in the buffer. These
are shown in dotted line in Figure 17.16.
Because node B13 remains uncovered, the next step consists of picking the
node with the largest number of connections. Nodes C3 , C4 and C5 have one
connection each. Assume one picks C3 . Then, node B13 is also covered and
the solution C = {C1 , C3 } has been found. If C4 or C5 is picked, the sets
FIGURE 17.15. Example to illustrate the backtracking algorithm.
C 6= {C1 , C4 } or C = {C1 , C5 } are also be solutions. The reader can verify

that if for example C2 is marked first, then the backtracking algorithm will end
up eliminating it.
The only aspect of the formulation that has not been considered in this solution
procedure is the requirement of the presence of at least one element from every
Ai in C. For example, Raghuraj et al. (1999) showed that the observability of
fault Ri is not ensured by the above algorithm if Ai ⊆ A j , ∀ j 6= i. Indeed, no
element from Ai would be present in the sets Bi j . However, the observability
of fault Ri could still be ensured only if some element of Ai is present in some
other set Bk,l , where k, l 6= i. This is not possible either because if Ai ⊆ Ak , then
Bk,i ∩ Ai = φ ∀k 6= i. Hence, the observability of fault i cannot be ensured.
To overcome this difficulty, Raghuraj et al. (1999) suggested that this can be
overcome by also adding Ai as a root node in the bipartite graph.
Multiple Faults
The sensor-location problem for multiple faults and maximum resolution

can be solved as an extension of the single-fault assumption problem. Let us
consider a specific case where single faults and two simultaneous faults are
important. Raghuraj et al. (1999) suggested the use of the following steps.
(1) Define Ai j = A ji = Ai ∪ A j and add them to the observability set in the

bipartite graph.
FIGURE 17.16. Application of the backtracking algorithm.

FIGURE 17.17. Illustration of multiple faults.
(2) Solve the single-fault-assumption problem for System 1. This will give the
solution to the problem just posed.
In the solution to the double-fault case, some redundant sets might be gen-
erated, and one could considerably reduce the number of root nodes for which
the observability problem has to be solved. Instead of attempting this here, we
simply point out how the single-fault-assumption solution can be used to handle
the double-fault case also. In fact, this approach provides us with a framework
for posing and solving various kinds of sensor-location problems. This is a more
important aspect of the proposed solution strategy. In a typical plant scenario,
one might not be concerned with all multiple-or double-fault possibilities. There
might be some faults that might have a high probability of occurring together.
One could simply add such sets of faults to the original single-fault sets and get
a solution to the sensor-location problem.
Illustrative Example
Consider the system of Figure 17.17. The backtracking algorithm is applied

to get the minimal set of sensors for observability. This set is {C7 , C8 }.
We now obtain the maximum resolution set of sensors. Table 17.2 shows
the observability sets for each root. Notice that the roots R1 and R2 should be
merged in one, because they are indistinguishable.
The algorithm for maximum fault resolution is applied rendering C = {C6 ,
S N2 , C8 }. As anticipated, faults R1 and R2 are indistinguishable. Under
TABLE 17.2. Observability Subsets

for Figure 17.7a.
Root Nodes Observability Sets
{R1 , R2 } A1 = {C6 , S N2 , C7 }
R3 A3 = {C7 , S N1 }
R4 A4 = {C7 , C8 }
R5 A5 = {C8 }
Nomenclature 337
TABLE 17.3. Auxiliary Sets Ai j and Bi j .
A12 = {C6 , S N2 , C7 , S N1 } B12 = {C6 , S N2 , S N1 }

A13 = {C6 , S N2 , C7 , C8 } B13 = {C6 , S N2 , C8 }
A14 = {C6 , S N2 , C7 , C8 } B14 = {C6 , S N2 , C7 , C8 }
A23 = {C7 , S N1 , C8 } B23 = {S N1 , C8 }
A24 = {C7 , S N1 , C8 } B24 = {C7 , S N1 , C8 }
A34 = {C7 , C8 } B34 = {C7 }
the double-fault assumption, the solution is C = {S N2 , C7 , S N1 , C8 }. Clearly,

the number of sensors required for the double-fault diagnosis is more than
for the single-fault case. Indistinguishable faults can also be identified, just as
in the single-fault assumption case. For example, from Table 17.3, A23 = A24 .
This means that the occurrence of A2 and A3 together cannot be distinguished
from the occurrence of A2 and A4 together. From Table 17.2, we also see that
A3 , A4 and A5 correspond to faults R3 , R4 , R5 . Hence, if they occur together,
faults R3 and R4 cannot be distinguished from the simultaneous occurrence of
faults R3 and R5 .
The application of the algorithms to a CSTR and an FCC unit are presented
by Raghuraj et al. (1999).
Future Scope
According to Raghuraj et al. (1999), there are a number of enhancements that

can be incorporated into the proposed algorithm. First, signs could be placed on
the DG, and sensor-location problems could be solved using the SDG. Instead
of a purely CE analysis, which is used in both DG and SDG, one could pose the
sensor-location problem based on semi-qualitative, order-of-magnitude models.
Other aspects that have not been considered yet are failure probabilities, sen-
sor failure probability, severity of particular faults and the cost of the
sensors.
NOMENCLATURE
Ai : subset of nodes connected to node Ri

Ci : node
Ri : node
S Ni : supernode
Li : level in unit i
Si : stream i
REFERENCES
Aelion V. and G. J. Powers. Risk Reduction of Operating Procedures and Process Flowsheets. Ind.
Eng. Chem. Res., 32, pp. 82–90 (1993).
Akhmentov D. F. and Y. Dote. Aggregative Learning Method and Its Applications to Fault Diagno-
sis. Proc. IFAC. Workshop on On-Line Fault Detection and Supervision in the Chemical Process
Industries, Lyon, France (1998).
Bensaker B. and F. Ouchene. Fault Detection for Continuous-Time Systems. The PMF Identification
Approach. Proc. IFAC. Workshop on On-Line Fault Detection and Supervision in the Chemical
Process Industries, Lyon, France (1998).
Busson F., A. Aitouche, B. O. Boumama and M. Staroswiecki. Sensor Failure Detection in Steam
Condensers. Proc. IFAC. Workshop on On-Line Fault Detection and Supervision in the Chemical
Catino C. A. and L. H. Ungar. Model-Based Approach to Automated Hazard Identification of
Chemical Plants, AIChE J., 41, 1, pp. 97–109 (1995).
Chang C. C. and C. C. Yu. On-Line Fault Diagnosis Using Signed Directed Graphs. Ind. Eng.
Chem. Res. 29, pp. 1290 (1990).
Chang C. T. and H. C. Hwang. New Developments of the Digraph-Based Technique for Fault Tree
Synthesis. Ind. Eng. Chem. Res. 31, pp. 1490 (1992).
Chang C. T. and H. C. Hwang. Studies on the Digraph-Based Approach for Fault Tree Synthesis.
Ind. Eng. Chem. Res. 33, pp. 1520 (1994).
Chen B. H., X. Z. Wang and C. McGreavy. On-Line Operational Support System for Fault Diagnosis
in Process Plants. Comp. & Chem. Eng., 22, Suppl., pp. S973-S976 (1998).
Cheung J. T.-Y. and G. Stephanopoulos. Representation of Process Trends-part I. A Formal Rep-
resentation Framework. Comp. & Chem. Eng., 14, pp. 495–510 (1990).
Cho H.-W. and C. Han. Hierarchical Plant-Wide Monitoring and Triangular Representation-Based
Diagnosis. Proc. IFAC. Workshop on On-Line Fault Detection and Supervision in the Chemical
Dong D. and T. J. McAvoy. Nonlinear Principal Component Analysis Based on Principal Curves
and Neural Networks. Comp. & Chem. Eng., 20, 1, pp. 65–78 (1996).
Dunia R. and S. J. Qin. Subspace Approach to Multidimensional Fault Identification and Recon-
struction. AIChE J., 44, 8, pp. 1813–1829 (1998).
Dunia R. and S. J. Qin. A Unified Geometric Approach to Process and Sensor Fault Identification
and Reconstruction: the Unidimensional Fault Case. Comp. & Chem. Eng., 22, 7–8, pp. 927–943
(1998).
Dunia R., S. J. Qin, T. F. Edgar and T. J. McAvoy. Use of Principal Component Analysis for Sensor
Fault Identification. Comp. & Chem. Eng., 20, Suppl., pp. S713–S718 (1996a).
Dunia R., S. J. Qin, T. F. Edgar and T. J. McAvoy. Identification of Faulty Sensors Using Principal
Component Analysis. AIChE J., 42, 10, pp. 2797–2812 (1996b).
Fan J. Y., M. Nikolau and R. E. White. An Approach to Fault Diagnosis of Processes via Neural
Networks. AIChE J., 39, 1, pp. 82–88 (1993).
Fathi Z., W. F. Ramirez and J. Korbicz. Analytical and Knowledge-Based Redundancy for Fault
Diagnosis in Process Plants. AIChE J., 39, 1, pp. 42–56 (1993).
Frank P. M. Fault Diagnosis in Dynamic Systems Using Analytical and Knowledge-Based
Redundancy—A Survey of Some Results. Automatica, 26, 3, pp. 459–474 (1990).
References 339
Gertler J. J. A Survey of Model-Based Failure Detection and Isolation in Complex Plants. IEEE
Control Systems Magazine, 8, 6, pp. 3–11 (1988).
Gertler J. J. Fault Detection and Diagnosis in Engineering Systems. Marcel Dekker, New York
(1998).
Gertler J. J., W. Li, Y. Huang and T. McAvoy. Isolation Enhanced Principal Component Analysis.
Proc. IFAC. Workshop on On-Line Fault Detection and Supervision in the Chemical Process
Gomm J. B., M. Weerasinghe and D. Williams. Pruning and Extraction of Qualitative Fault
Diagnosis Rules from a Neural Network. Proc. IFAC. Workshop on On-Line Fault Detection
and Supervision in the Chemical Process Industries, Lyon, France (1998).
Han Z. and P. M. Frank. A New Scheme of Identification-Based Fault Detection Proc. IFAC.
Workshop on On-Line Fault Detection and Supervision in the Chemical Process Industries,
Lyon, France (1998).
Himmelblau D. Fault Detection and Diagnosis in Chemical and Petrochemical Processes. Elsevier,
Amsterdam (1978).
Iri M., K. Aoki, E. O’Shima and H. Matsuyama. An Algorithm for Diagnosis of System Failures in
the Chemical Process. Comp. & Chem. Eng., 3, pp. 489–493 (1979).
Iri M., K. Aoki, E. O’Shima and H. Matsuyama. A Graphical Approach to the Problem of Locating
the Origin of the System Failure. J. Opers Res. Soc. Jap. 23, pp. 295–311 (1980).
Jackson J. E. A User’s Guide to Principal Components. Wiley, New York (1991).
Jia F., E. B. Martin and A. J. Morris. Nonlinear Principal Component Analysis for Process Fault
Detection. Comp. & Chem. Eng., 22, Suppl., pp. S851–S854 (1998).
Kavuri S. N. and V. Venkatasubramanian. Representing Bounded Fault Classes Using Neural
Networks with Ellipsoidal Activation Functions. Comp. & Chem. Eng., 17, 2, pp. 139–163
(1993).
Kavuri S. N. and V. Venkatasubramanian. Using Fuzzy Clustering and Ellipsoidal Units in Neural
Network to Improve Fault Classifications. Comp. & Chem. Eng., 17, 8, pp. 765–784 (1993b).
Kim S.-H. and K.-S. Lee. A Study on the Development of Robust Fault Diagnostic System Based
on Neuro-Fuzzy Scheme. Proc. IFAC. Workshop on On-Line Fault Detection and Supervision in
the Chemical Process Industries, Lyon, France (1998).
Kokawa M. and S. Shingai. Failure Propagation Simulation and Non-Failure Path Search in Net-
work Systems. Automatica, 18, pp. 335 (1982).
Kokawa M., S. Miyasaki and S. Shingai. Fault Location Using Digraph and Inverse Direction
Search with Application. Automatica, 19, pp. 729 (1983).
Kramer M. A. Malfunction Diagnosis Using Quantitative Models with Non-Boolean Reasoning in
Expert Systems. AIChE J., 33, 1, pp. 130–140 (1987).
Kramer M. A. and B. L. Palowitch, Jr. A Rule-Based Approach to Fault Diagnosis Using the Signed
Directed Graph. AIChE J., 33, 7, pp. 1067–1078 (1987).
Kresta J. V., J. F. McGregor and T. Marlin. Multivariate Statistical Monitoring of Process Operating
Performance. Can. J. Chem. Eng., 69, pp. 35–47 (1991).
Lambert H. E. Fault Trees for Locating Sensors in Process Systems. CEP, August, pp. 81–85 (1977).
Lakshminarayanan S., S. L. Shah and K. Nandakumar. Modeling and Control of Multivariable
Processes: Dynamic PLS Approach. AIChE J., 43, 9, pp. 2307–2322 (1997).
Lapp S. A. and G. J. Powers. Computer-Aided Synthesis of Fault Trees. IEEE Trans. Reliab., 2,
pp. 13 (1977).
Leung D. and J. Romagnoli. Fault Diagnosis Expert System with Probability Calculations. Proc.
IFAC. Workshop on On-Line Fault Detection and Supervision in the Chemical Process Industries,
Lyon, France (1998).
Luan D.-N., D.-S. Chang, M.-P. Chang and C.-C. Yu. Diagnosis of Abrupt Faults Using Variable-
Structure Neural Network. Proc. IFAC. Workshop on On-Line Fault Detection and Supervision
in the Chemical Process Industries, Lyon, France (1998).
Luo R., M. Misra, S. J. Qin, R. Barton and D. Himmelblau. Sensor Fault Detection via Multiscale
Analysis and Nonparametric Statistical Inference. Ind. Eng. Chem. Res. 37, 3, pp. 1024–1032
(1998).
McGregor J. F., C. Jaeckle, C. Kiparissides and M. Koutoudi. Process Monitoring and Diagnosis
by Multiblock PLS Methods. AIChE J., 40, 5, pp. 826–838 (1994).
Marcu T. and P. M. Frank. Parallel Evolutionary Approach to System Identification for Process
Fault Diagnosis. Proc. IFAC. Workshop on On-Line Fault Detection and Supervision in the
Chemical Process Industries, Lyon, France (1998).
Mohindra S. and P. A. Clark. A Distributed Fault Diagnosis Method Based on Digraph Models:
Steady-State Analysis. Comp. & Chem. Eng., 17, 2, pp. 193–209 (1993).
Naidu S., E. Zafiriou and T. J. McAvoy. Application of Neural Networks on the Detection of Sensor
Failure during Operation of a Control System. Proc. ACC, Pittsburgh, pp. 1136 (1989).
Negiz A. and A. Cinar. Automated Detection of Incipient Sensor Faults by Multivariable Statistical
Methods. Proc. PSE’94, pp. 919–924 (1994).
Ozyurt I. B., A. K. Sunol and L. O. Hall. Semi-Quantitative Model-Based Chemical Process Fault
Diagnosis via Episodic Fuzzy Rules. Proc. IFAC. Workshop on On-Line Fault Detection and
Supervision in the Chemical Process Industries, Lyon, France (1998).
Parker R. G. and R. L. Rardin. Discrete Optimization. Academic Press, San Diego (1988).
Patton R. J. Robustness in Model-Based Fault Diagnosis: The 1995 Situation. Proceedings of the
IFAC On-Line Fault Detection and Supervision in the Chemical Process Industries, Newcastle
upon Tyne, UK (1995).
Patton R. J., P. M. Frank and R. N. Clark. Fault Diagnosis in Dynamic Systems: Theory and
Application. Prentice Hall, Englewood Cliffs, NJ (1989).
Pau L. F. Failure Diagnosis and Performance Monitoring. Marcel Dekker, New York (1981).
Qian D.-Q. An Improved Method for Fault Location of Chemical Plants. Comp. & Chem. Eng., 14,
1, pp. 41–48 (1990).
Qin S. J. Recursive PLS Algorithms for Adaptive Data Modeling. Comp. & Chem. Eng., 22, 4/5,
pp. 503–514 (1998).
Qin S. J. and T. J. McAvoy. Nonlinear PLS Modeling Using Neural Networks. Comp. & Chem.
Eng., 16, 4, pp. 379–391 (1992).
Raghuraj R., M. Bhushan and R. Rengaswamy. Locating Sensors in Complex Chemical Plants
Based on Fault Diagnostic Observability Criteria. AIChE J., 45, 2, pp. 310–322 (1999)
Shao R., F. Jia, B. Martin and A. J. Morris. Fault Detection Using Wavelet Filtering and Non-
Linear Principal Component Analysis. Proc. IFAC. Workshop on On-Line Fault Detection and
Shiozaki J., H. Matsuyama, E. O’Shima and M. Iri. An Improved Algorithm for Diagnosis of System
Failures in the Chemical Process. Comp. & Chem. Eng., 9, pp. 285 (1985).
Stanley G. M. and R. Vaidhayanathan. A Generic Fault Propagation Modeling Approach to On-
Line Diagnosis and Event Correlation. Proc. IFAC. Workshop on On-Line Fault Detection and
References 341
Su H. B. and L. T. Fan. A Rule-Based Expert System Integrated with Neural Networks for Process
Fault Diagnosis. Proceedings of the AIChE Annual Meeting, St. Louis (1993).
Szigeti F. and R. Tarantino. Augmented Space Method for Fault Detection and Isolation Filter.
Proc. IFAC. Workshop on On-Line Fault Detection and Supervision in the Chemical Process
Tarjan R. Depth-First Search and Linear Graph Algorithms. SIAM J. Comput., 1, pp. 146 (1972).
Tsai C.-S. and C.-T. Chang. Optimal Alarm Logic Design for Mass Flow Networks. AIChE J., 43,
11, pp. 3021–3030 (1997).
Umeda T., T. Kuriyama, E. O’Shima and H. Matsuyama. A Graphical Approach to Cause and
Effect Analysis of Chemical Processing Systems. Chem. Eng. Sci., 35, pp. 2379 (1980).
Vedam H. and V. Venkatasubramanian. Automated Interpretation of PCA-Based Process Moni-
toring and Fault Diagnosis Using Signed Digraphs. Proc. IFAC. Workshop on On-Line Fault
Detection and Supervision in the Chemical Process Industries, Lyon, France (1998).
Vedam H., V. Venkatasubramanian and Mohan Bhalodia. A B-Spline-Based Method for Data Com-
pression, Process Monitoring and Diagnosis. Comp. & Chem. Eng., 22, Suppl., pp. S827–S830,
(1998).
Venkatasubramanian V. and K. Chan. A Neural Network Methodology for Process Fault Diagnosis.
AIChE J., 35, pp. 1993 (1989)
Venkatasubramanian V. and S. H. Rich. An Object Oriented Two-Tier Architecture for Integrating
Compiled and Deep-Level Knowledge for Process Diagnosis. Comp. & Chem. Eng., 12, 9/10,
pp. 903–921 (1988)
Vinson J. M. and L. H. Ungar. Using PLS for Fault Analysis: A Case Study. AIChE Annual Meeting
(1994).
Watanabe K., I. Matsura, M. Abe, M. Kubota and D. M. Himmelblau. Incipient Fault Diagnosis of
Chemical Processes via Artificial Neural Networks. AIChE J., 35, pp. 1803 (1989).
Watanabe K., S. Hirota, L. Hou and D. Himmelblau. Diagnosis of Multiple Simultaneous Fault via
Hierarchical Artificial Neural Networks. AIChE J., 40, 5, pp. 839–848 (1994).
Wilcox N. A. and D. M. Himmelblau. The Possible Cause and Effect Graphs (PCEG) Model for
Fault Diagnosis-I. Methodology. Comp. & Chem. Eng., 18, 2, pp. 103–116 (1994a).
Wilcox N. A. and D. M. Himmelblau. The Possible Cause and Effect Graphs (PCEG) Model for
Fault Diagnosis. II. Applications. Comp. & Chem. Eng., 18, 2, pp. 117–127 (1994b).
Wise B. M. and N. B. Gallagher. The Process Chemometrics Approach to Process Monitoring
and Fault Detection. Proceedings of the IFAC On-Line Fault Detection and Supervision in the
Chemical Process Industries, Newcastle upon Tyne, UK (1995).
Yao S. C. and E. Zafiriou. Control System Sensor Failure Detection via Networks of Localized
Receptive Fields. Proc. ACC, pp. 2472 (1990).
INDEX
Accuracy, 5, 74, 76 Fundamental cutsets, 106

Alarm Logic, 326 Ring sum, 227
Alarms, 26, 315 Cycle, 100
Availability, 211
Sensor service availability, 211 Data reconciliation, 7, 159, 170
System availability, 222 Dynamic data reconciliation, 170
Linear data reconciliation, 162
Bias, 69 Nonlinear steady state data reconciliation,
169
calibration Precision of estimates, 165, 185
Base conditions, 35 Steady state linear data reconciliation,
Compensation, 35, 75 163
Curves, 74 Variance Estimation, 168
Testing conditions, 35 Variance of observable quantities, 165
Canonical matrices, 100 Dead band, 72
Canonical matrix, 87 Degeneracy, 234
Canonical systems, 87 Degree of estimability, 98
Compulsory measurements and upgrade, Degree of observability, 97
140 Degree of redundancy, 95, 97
Conformity, 74 Density instrumentation, 58
Connectivity, 83 Design for estimability, 128
Controlled variables, 18 Design for estimability efficiency, 139
Corrective maintenance, 281 Design of bilinear multicomponent Systems,
Cost-efficient nonrepairable sensor networks, 259
269 Digital protocols, 61
Maximum reliability model, 271 Dynamic data reconciliation, 170
Minimal networks, 270
Minimum cost model, 269 Efficiency of estimability, 108
Cutset, 103 Energy networks, 266
Determination, 107 Energy sensor networks, 151
Efficiency of estimability, 108 Entering variables, 230, 232, 239, 261
Estimable cutsets, 104 Error detectability, 297
343
344 Index
Errors in measurements, 65 Outliers, 292

Estimability, 25, 96 Resilience, 26, 298
Estimation availability, 217 Robustness, 25
Estimation reliability, 218 Size estimation, 297
Expected number of repairs, 280 Test for gross error presence, 293
GSENNET, 240
Failure density, 212
Failure intensity, 279 Hardware redundancy, 6, 87, 160, 176
Failure rate, 212 Hill climbing, 234
Fault detection, 2, 11, 26 Hysteresis, 72
Fault diagnosis, 26, 315
Fault observability, 327 Inaccuracy in gross error detection, 296
Fault resolution, 326, 332 Incidence matrix, 84
Fault trees, 316 Inferential control, 23
Flowrate exchange cost, 145 Influence variables, 65
Flowrate instrumentation, 31 Instrumentation design goals, 25
Coriolis flowmeter, 42
Differential pressure meters, 31, 44 Key variables, 15
Discharge coefficient, 34 Key variables in control, 17
Electromagnetic meters, magmeters, Knowledge-based fault detection, 322
35
Flow nozzle, 32 Leaks, 292
Mass meters, 31, 40, 45 Leaving variables, 230, 232, 238, 261
Orifice plate, 32 Level instrumentation, 46
Pitot tube, 39 Differential pressure meters, 46
Positive displacement meters, 31, 40, 45 Displacer meters, 46
Thermal flowmeter, 42 Radiation-based meters, 46, 50
Turbine meters, 35 Radio-frequency admittance meters, 46,
Ultrasonic meters, 37 49
Variable area meters, 31, 44, 45 Signal reflection meters, 46, 48
Velocity meters, 31, 35, 44 Strapping tables, 50
Venturi tube, 32 Linear data reconciliation, 162
Vortex meter, 36 Linearity, 73
Linearly estimable deficient networks,
General Networks, 137 144
Generalized maximum precision models, Linearly estimable sensor networks, 143
180
Genetic Algorithms, 307 Maintenance, 26
Global test, 293 Manipulated variables, 19
Graphs, 82 Matrix projection, 93
Gross errors Maximum precision models, 178, 180
Detectability, 25 Maximum reliability model, 271
Detection, 8 Mean time to failure, 216
Estimation, 9 Measured variables, 15, 19
Inaccuracy in gross error detection, Measurement test, 294
296 Minimal networks, 129, 270
Leaks, 292 Minimally estimable sensor networks, 143
Minimum cost model, 300 Model based monitoring, 4
Multiple gross error identification, Model-based fault detection, 322
296 Monitoring, 2, 4
Origin, 291 Monitoring variables, 16
Index 345
Multiple gross error identification, 296 Reliable linear nonrepairable sensor networks,
Multiple instruments, 176 225
Multivariate statistical methods, 324 Repair intensity, 280
Repairable sensor networks, 279
Networks with maximum reliability, 225 Minimum cost model, 286
Neural networks, 321 Reproducibility (see precision), 5
Nonlinear steady state data reconciliation, Residual precision, 26, 284
169 Resilience, 298
Nonrepairable sensor networks, 211 Resource reallocation, 198
Reliability, 211 Ring sum of cutsets, 227
Sensor service availability, 211 Robust sensor networks, 291, 299
Sensor service reliability, 212 Rule-based fault detection algorithms, 320
Observability, 85 SENNET, 232

Observability analysis Sensitivity, 71
Bilinear systems, 114 Sensor networks for bilinear systems, 142
Full nonlinear systems, 120 Energy sensor networks, 151
Linear systems, 87, 93 Flowrate exchange cost, 145
On Line analyzers, 60 Linearly estimable deficient networks, 144
On-line optimization, 3 Linearly estimable sensor networks, 143
Optimal bilinear sensor networks, 256 Minimally estimable sensor networks, 143
Outliers, 71, 292 Networks with splitters, 148
Sensor networks for fault diagnosis, 315
Parameter estimation, 2, 184 Sensor networks for fault resolution, 332
Planning, 2, 3 Sensor service availability, 211
Precise sensor networks, 175 Sensor service reliability, 212
Precision, 5, 7, 25, 66 Signed digraph, 317
Precision upgrade of Networks, 191 Singular value analysis, 20
Resource reallocation, 198 Smart sensors, 61
Pressure instrumentation, 56 Software redundancy, 6
Elastic element devices, 56 Span, 65
Electrical sensing devices, 56 Span shift, 70
Liquid column devices, 56 Spanning tree, 100, 146
Precision of estimates, 165, 185 Branch, 100
Process variables, 17 Chord, 100
Distance, 102
QR decomposition, 93 Speed of response, 71
Stopping criteria, 183
Range, 65 Subminimal networks, 132
Rangeability, 65 System availability, 222
Readability, 66 System connectivity, 83
Redundancy, 86, 95 System reliability, 222
Redundant and restricted networks, 246 Systematic error, 69
Redundant networks with maximum
reliability, 237 Temperature instrumentation, 51
Reliability, 5, 8, 25, 212 Electrical thermometers, 51, 53
Estimation reliability, 218 Radiation-based thermometers, 51, 55
Evaluation in bilinear systems, 254 Thermal expansion thermometers, 51
System reliability, 222 Tree, 100
Reliable bilinear nonrepairable sensor Tree-searching algorithm, 133
networks, 253 Tuy Duality, 182, 272
346 Index
Upgrade of instrumentation, 27, 127 Measured, 86

Compulsory measurements and upgrade, Nonredundant, 86
140 Observable, 86
Precision upgrade, 191 Redundant, 86
Upgrade of bilinear sensor networks, Unmeasured, 86
152 Unobservable, 86
Variance estimation, 168
Variables Variance of observable quantities,
Classification, 81, 110 165

Process Plant Instrumentation Design and Upgrade

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Process Plant Instrumentation Design and Upgrade

Uploaded by

Copyright:

Available Formats

Process Plant

Miguel J. Bagajewicz, Ph.D.

Boca Raton London New York

CRC Press is an imprint of the

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-56676-998-3 (Hardback)

Chapter 1. Plant Data Management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Chapter 2. Instrumentation Design Goals . . . . . . . . . . . . . . . . . . . . . . . 15

Chapter 4. Errors in Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

Chapter 5. Variable Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

Chapter 6. Design and Upgrade of Nonredundant and

Design for Estimability Efficiency 139

Chapter 7. Data Reconciliation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

Chapter 8. Design of Precise Sensor Networks . . . . . . . . . . . . . . . . 175

Chapter 9. Precision Upgrade of Sensor Networks . . . . . . . . . . . . 191

Chapter 10. Reliability of Nonrepairable

Sensor Service Reliability 212

Chapter 11. Design of Reliable Linear Nonrepairable

Chapter 12. Design of Reliable Bilinear Nonrepairable

Chapter 13. Design of Reliable and Cost-Efficient

Chapter 14. Design of Repairable Sensor Networks . . . . . . . . . . . 279

Chapter 15. Design of Robust Sensor Networks . . . . . . . . . . . . . . . 291

Chapter 16. Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307

Chapter 17. Design of Sensors for Process

Sensor Location for Fault Observability 327

The process of obtaining good estimates of process variables that conform to

Plant Data Management

Instrumentation is needed in process plants to obtain data that are essential

PLANT INFORMATION AND OPERATIONS MANAGEMENT

of the existing instrumentation are insufficient to obtain good quality esti-

Estimate parameters Update the model

Filtering of gross errors & Optimize & Update

Plant & Controllers

FIGURE 1.1. Real-time optimization.

Model-based monitoring consists of the use of a combination of models

—capability of identifying gross errors through statistical model-based tech-

Accurate and imprecise

Inaccurate and precise True value

FIGURE 1.2. Accuracy and precision.

Summarizing, precision is a desired property related to the quality of the

Software or Analytical Redundancy

In addition to direct measurements, there are indirect ways of determining

Precision and Software/Analytical Redundancy

Software or analytical redundancy increases the precision of estimates. It

Reliability and Software/Analytical Redundancy

If data are to be reliable, the instrumentation used has to be reliable too.

Detection of Gross Errors in Instrumentation

In the absence of gross errors, data reconciliation reduces typically to a

Estimation of Gross Errors

and eliminates the corresponding measurement until no gross error is de-

Many researchers have evaluated the performance of these approaches. Se-

Detection of Process Faults

Faults propagate throughout the process, altering the readings of instruments

Proceedings of the Process Systems Engineering (PSE) International Symposium, Kyongju,

Instrumentation Design Goals

In this chapter, the goals of instrumentation design are presented in further

MEASURED AND KEY VARIABLES

Traditionally, the fields of monitoring and process control relied heavily on

Definition: Measured variables are the variables whose measurements al-

Thus, the optimal choice of measurements can now be performed by mak-

SELECTION OF MONITORING VARIABLES

Monitoring of a process is performed for several purposes. Table 2.1 summa-

TABLE 2.1. Key Variables in Monitoring.