Professional Documents
Culture Documents
Instrumentation
Design and Upgrade
Process Plant
Instrumentation
Design and Upgrade
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 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
Index 343
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
INTRODUCTION
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
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
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.
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.
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.
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.
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
10 PLANT DATA MANAGEMENT
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.
12 PLANT DATA MANAGEMENT
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
INTRODUCTION
Definition: Key variables are the variables for which their estimated values
are of interest for control, monitoring and parameter estimation.
x1 y1
x2 y2
Input x3 PROCESS y3 Output
• •
variables • • variables
xM yN
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.
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 Key Variables in Control 19
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.
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.
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)
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
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
Selection of Key Variables in Control 23
Inferential 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.
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.
— 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:
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:
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:
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
NOMENCLATURE
Greek Letters
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).
30 INSTRUMENTATION DESIGN GOALS
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
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
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
Flow Rate Instrumentation 33
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
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
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
FIGURE 3.2. Variation of the discharge coefficient with the Reynolds number. Adapted
from Upp, 1993.
Flow Rate Instrumentation 35
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
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.
Flow Rate Instrumentation 37
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
FIGURE 3.6. Turbine meter and vortex meter signature curves. Adapted from Miller, 1996.
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.
Flow Rate Instrumentation 39
A B
a
d
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
Flow Flow
Low
DP Openings for static pressure
pressure side
High pressure
side
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 (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
Flow Rate Instrumentation 41
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.
Wires
• •• •
Sensor Tube
Q2
Q
Q1
Laminar By-pass
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 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
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.
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
Flow Rate Instrumentation 45
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.
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.
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
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.
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
Ultrasonic
Transmitter
Ultrasonic
Transducer
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
Radioactive
Source
Radioactive
Detector
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.
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
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).
FIGURE 3.21. Filled system. Reprinted from Perry, 1984, reproduced with permission of
The McGraw-Hill Companies.
Temperature Measurement 53
FIGURE 3.22. Bimetallic thermometers. Reprinted from Perry, 1984, reproduced with
permission of The McGraw-Hill Companies.
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).
Relative
Resistance
(Rt/Ro) 6
Nickel
Copper
4
Platinum
Radiation-Based 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
FIGURE 3.25. C-type bourdon manometer. Reprinted from Perry, 1984, reproduced with
permission of The McGraw-Hill Companies.
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.
DENSITY MEASUREMENT
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.
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.
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.
Differential
Temperature Pressure Pressure
signal signal signal
P
T DP
‚
D P* (P/T)
· K
P/T
Square True mass
flow
Root
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
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
INSTRUMENT PROPERTIES
Influence Variables
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
105 105
103 103
Temperature
Temperature
101 101
99 99
97 97
95 95
1 250 1 250
Time Time
(a) (b)
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
σ P = t p ŝ (4.3)
Origin of Fluctuations
105
104
103
Temperature 102
101
100
99
98
97
96
95
1 250
Time
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
70 ERRORS IN MEASUREMENT
pure substances. Flowmeters, in turn, need to be calibrated with the aid of some
other more precise instruments.
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.
Measurement Quality 71
104
103
Temperature 102
101
100
99
98
97
96
0 Time 250
Outliers
104
103
102
Temperature
101
100
99
98
97
96
0 Time 250
FIGURE 4.6. Gaussian signal with outliers.
72 ERRORS IN MEASUREMENT
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.
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.
Measurement Quality 73
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.
74 ERRORS IN MEASUREMENT
maximum deviations equal. These deviations from linearity are called static
errors.
Accuracy
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.
Flow Compensation
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.
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
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).
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).
78 ERRORS IN MEASUREMENT
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
MODEL
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
z = h(x) + ε (5.3)
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.
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
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
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
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
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)
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:
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
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.
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)
−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
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)
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
Canonical Representation of Linear Systems 91
S6
S8
U1 U2 U3 U4 U5
S1 S2 S3 S4 S5 S9
S7
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
92 VARIABLE CLASSIFICATION
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
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
dw4
= f4 − f7 − f8 − f9 (5.35)
dt
f5 − f8 − f9 = 0 (5.36)
S8
S4 U4 S5
U5
S9
S7
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)
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)
94 VARIABLE CLASSIFICATION
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
−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
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
96 VARIABLE CLASSIFICATION
S2 S4
S1 S2 S3 S4 S1 S5
S3 S6
(a) (b)
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.
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
S2 S4
U2
S1 S6
U1 U4
S3 S5
U3
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.
S1
U1 U2
S2
S3 S4
ENV
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.
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.
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.
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
S2 S4
U2
S1 S6
U1 U4
S3 S5
U3
ENV
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
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
S3 S4 S6 S5 S1 S2
1 -1 1
1 -1 (5.47)
1 -1
1 -1 1
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
ENV
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 ,
Graphs and Canonical Matrices 105
S2 S4 K4
U2 K5
S1 S6
U1
U4
K2
S3 S5
U3
K3
ENV
K1
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
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.
Graphs and Canonical Matrices 107
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,
108 VARIABLE CLASSIFICATION
K2
S2 S4
U2
S1 S6
U1 U4
K1
S5 K4
S3 U3
K5 K3
ENV
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.
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.
Graphs and Canonical Matrices 109
S2 S4
U2 K5
S1 S6
U1 U4
K2
S3 S5
U3
ENV
K1
S2 S4
U2
S1 S6
U1 U4
K2
S3 S5
U3
ENV
K1
NONLINEAR SYSTEMS
Component Balances
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
f j = f ⊗ cj (5.52)
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)
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
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
112 VARIABLE CLASSIFICATION
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
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)
Nonlinear Systems 113
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
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]
S3
S1 S2
S4
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)
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
Nonlinear Systems 115
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
S1 M M
S2 M M
S3
S4 M
S5 M
116 VARIABLE CLASSIFICATION
M MU
S4 S5 S4 S ˜3
-1 f2 - f1
1 -1 f2
-1 f1c1 - f2c2
c4 - c5 f2c2
(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 )
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 )
Nonlinear Systems 117
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
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
118 VARIABLE CLASSIFICATION
S4 S6
S1 S2 S3
S5 S7
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.
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)
S1 M M
S2 M
S3
S4 M M
S5 M
S6 M
S7 M
Nonlinear Systems 119
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)
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
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)
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
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
−0.0082
−0.0082
−0.0082
0.8164
Q2 =
−0.412
(5.118)
−0.404
0
0
0
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
124 VARIABLE CLASSIFICATION
Greek Letters
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).
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, 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).
126 VARIABLE CLASSIFICATION
INTRODUCTION
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.
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
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
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,
130 NONREDUNDANT AND REDUNDANT SENSOR NETWORKS
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.
Design for Estimability 131
S2 S4
U2
S1 S6
U1 U4
S3 S5
U3
( ):Measured flow rates
(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
U4
S8 (250) S7 (230)
S9 (130)
U5 U3
S11(150) S6 (160)
U7
U8
S13 (250) S5 (200)
U6 U2
S3 (180)
S2 (150) S1 (300)
U1
U4
S8 (250) S7 (230)
S9 (130)
U5 U3
S11(150) S6 (160)
U7
U8
S13 (250) S5 (200)
U6 U2
S3 (180)
S2 (150) S1 (300)
U1
( ):Measured flow rates
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
Design for Estimability 133
Tree-Searching Algorithm
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
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
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
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
136 NONREDUNDANT AND REDUNDANT SENSOR NETWORKS
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
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:
U4
S8 (250) S7 (230)
S9 (130)
U5 U3
S11(150) S6 (160)
U7
U8
S13 (250) S5 (200)
U6 U2
S3 (180)
S2 (150) S1 (300)
U1
General Networks
Example 6.4
U4
S8 (250) S7 (230)
S9 (130)
U5 U3
S11(150) S6 (160)
U7
U8
S13 (250) S5 (200)
U6 U2
S3 (180)
S2 (150) S1 (300)
U1
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).
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
U5
S8 S4
S7
S6 S5
U1 U6 U4
S1 S3
U2 S2 U3
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.
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 )
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.
142 NONREDUNDANT AND REDUNDANT SENSOR NETWORKS
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
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.
Sensor Networks for Bilinear Systems 143
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)
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.
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.
144 NONREDUNDANT AND REDUNDANT SENSOR NETWORKS
F3, y3
T, P, h
F1, y1
F2, y2
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
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)
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
Sensor Networks for Bilinear Systems 147
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.
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.
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 .
148 NONREDUNDANT AND REDUNDANT SENSOR NETWORKS
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)
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.
Sensor Networks for Bilinear Systems 149
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
shown in Figure 6.14.
150 NONREDUNDANT AND REDUNDANT SENSOR NETWORKS
U1
S1 S3
KS2 S KS1
c1
S4 c2
cT
S2
U2
Example 6.8
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.
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.
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.
Sensor Networks for Bilinear Systems 153
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.
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
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)
I GU
UO
GRO,1 GNR,1
D= I GM GRO,2 GNR,2 (6.27)
UO
GR
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)
156 NONREDUNDANT AND REDUNDANT SENSOR NETWORKS
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
Greek Letter
REFERENCES
Ali Y. and S. Narasimhan. Sensor Network Design for Maximizing Reliability of Linear Processes.
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).
158 NONREDUNDANT AND REDUNDANT SENSOR NETWORKS
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).
Madron F. Process Plant Performance, Measurement Data Processing for Optimization and
Retrofits. Ellis Horwood, West Sussex, England (1992).
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).
Serth R. and W. Heenan. Gross Error Detection and Data Reconciliation in Steam Metering Systems.
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
d x1
= g1 (x1 , x2 ) (7.1)
dt
g2 (x1 , x2 ) = 0 (7.2)
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
( )
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
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
−1
Z = {[e
xM (ti ) − z M,k ]T Q xM (ti ) − z M,k ]}
[e (7.8)
−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
162 DATA RECONCILIATION
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.
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)
dw R
BR = AR fR (7.14)
dt
CR f R = 0 (7.15)
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
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
164 DATA RECONCILIATION
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)
e
f N R = f N+R (7.22)
f O = CRO e
e f R + CNRO e
f NR (7.23)
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.
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):
Standard
Stream Measurement Deviation
S1 101.3 2.1
S6 102.7 1.9
Steady-State Linear Data Reconciliation 167
Reconciled/ Standard
Standard Estimated Deviation of
Stream Measurement Deviation Value Estimate
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
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.
168 DATA RECONCILIATION
Variance Estimation
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
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
170 DATA RECONCILIATION
NOMENCLATURE
A: incidence matrix
AR: incidence matrix for redundant systems
B: incidence matrix
BR : incidence matrix for redundant systems
172 DATA RECONCILIATION
C: incidence matrix
CR : incidence matrix for redundant systems
f: vector of flow rate of streams
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
x: vector of state variables
e
x: estimates of state variables x
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
INTRODUCTION
COST-OPTIMAL DESIGN
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
m
XX
ni
C( p) = ci,k pi,k (8.4)
∀i k=1
Multiple Instruments and Hardware Redundancy 177
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%.
Solution S1 S2 S3 S4
A — 2% 2% —
B — 2% — 2%
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
Solution S1 S2 S3 S4
C 3% 3% 2% —
D 3% 3% — 2%
Maximum Precision Models 179
Solution S1 S2 S3 S4
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.
180 DESIGN OF PRECISE SENSOR NETWORKS
(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.
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 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
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.
182 DESIGN OF PRECISE SENSOR NETWORKS
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.
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.
184 DESIGN OF PRECISE SENSOR NETWORKS
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
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
Example 8.2
components have the same vaporization efficiency η, that is, Equation (5.111)
is rewritten as follows:
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
Instrument Instrument
Measured Instrument Standard Measured Instrument Standard
Variable Cost Deviation Variable Cost Deviation
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
188 DESIGN OF PRECISE SENSOR NETWORKS
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
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
Greek Letters
η: vaporization efficiency
σ: precision
σ ∗: precision threshold
e
σ: precision of estimates
θ: vector of parameters
190 DESIGN OF PRECISE SENSOR NETWORKS
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).
Bagajewicz M. Design and Retrofit of Sensor Networks in Process Plants. AIChE J., 43, 9, pp. 2300–
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).
Chmielewski D., T. E. Palmer and V. Manousiouthakis. Cost Optimal Retrofit of Sensor Networks
with Loss Estimation Accuracy. AIChE Annual Meeting, Dallas (1999).
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).
Kretsovalis A. and R. S. H. Mah. Observability and Redundancy Classification in Multicomponent
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. Process Plant Performance, Measurement Data Processing for Optimization and
Retrofits. Ellis Horwood, West Sussex, England (1992).
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
INTRODUCTION
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
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.
Upgrade Models Based on Addition of Sensors 193
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
194 PRECISION UPGRADE OF SENSOR NETWORKS
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
Example 9.2
TABLE 9.2. Minimum Cost Optimal Solutions for Flash Sensor Upgrading.
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 — — —
Upgrade Models Based on Addition of Sensors 195
TABLE 9.3. Maximum Precision Optimal Solutions for Flash Sensor Upgrading.
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.
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
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
196
Upgrade Models Based on Addition of Sensors 197
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)
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.
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 à !
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
200 PRECISION UPGRADE OF SENSOR NETWORKS
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
u t,k,r ≤ u rR ∀r ∈ M R
t∈MT k∈K t
à !
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
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.
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.
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
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 )
204 PRECISION UPGRADE OF SENSOR NETWORKS
Variable
Index Variable Ni σi ,k ci ,k σi ,k
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
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 — — — —
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
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
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
208
References 209
NOMENCLATURE
Greek Letters
σ: precision
σ ∗: precision threshold
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).
Bagajewicz M. and M. Sánchez. Reallocation and Upgrade of Instrumentation in Process Plants.
Comp. & Chem. Eng. (2000).
210 PRECISION UPGRADE OF SENSOR NETWORKS
Reliability of Nonrepairable
Sensor Networks
INTRODUCTION
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:
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.
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
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.
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 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:
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
216 RELIABILITY OF NONREPAIRABLE SENSOR NETWORKS
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.
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.
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
Example 10.1
F1 = F2 + F3 (10.13)
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:
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:
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)
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
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:
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
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)
220 RELIABILITY OF NONREPAIRABLE SENSOR NETWORKS
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
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)
one more measurement in S6 . There is one new cutset containing S1 and only
measured variables:
(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.
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.
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
Greek Letters
σ ∗: precision threshold
τi : mean time to failure of sensor i
224 RELIABILITY OF NONREPAIRABLE SENSOR NETWORKS
REFERENCES
Ali Y. Sensor Network Design for Maximizing Reliability of Processes. Ph.D. Thesis, Indian Institute
of Technology. Kanpur, India (1993).
Ali Y. and S. Narasimhan. Sensor Network Design for Maximizing Reliability of Linear Processes.
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
INTRODUCTION
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.
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.
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.
228 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
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:
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,
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
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.
230 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
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 )}.
then
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,
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
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
Algorithm
233
234 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
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.
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
Nonredundant Networks Featuring Maximum Reliability 235
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
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:
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 }.
236 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
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).
{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 }
Redundant Networks Featuring Maximum Reliability 237
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.
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.
238 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
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
240 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
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
241
242 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
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
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
Redundant Networks Featuring Maximum Reliability 243
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.
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).
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.
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 .
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.
—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
248 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
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
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
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
249
250 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
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
NOMENCLATURE
REFERENCES
Ali Y. and S. Narasimhan. Sensor Network Design for Maximizing Reliability of Linear Processes.
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).
Kretsovalis A. and R. S. H. Mah. Observability and Redundancy Classification in Generalized
Process Networks. I. Theorems. Comp. & Chem. Eng., 12, 7, pp. 671–687 (1988a).
252 DESIGN OF RELIABLE LINEAR NONREPAIRABLE SENSOR NETWORKS
Madron F. and V. Veverka. Optimal Selection of Measuring Points in Complex Plants by Linear
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).
Serth R. and W. Heenan. Gross Error Detection and Data Reconciliation in Steam Metering Systems.
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
INTRODUCTION
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
Evaluation of Reliability
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:
where pi and pix are the failure probability of flow rate and composition sensor
of stream Si , respectively.
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.
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.
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
260 DESIGN OF RELIABLE BILINEAR NONREPAIRABLE SENSOR NETWORKS
Example 12.1
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.
262 DESIGN OF RELIABLE BILINEAR NONREPAIRABLE SENSOR NETWORKS
FIGURE 12.5. Ammonia plant network with unequal composition sensor failure proba-
bilities.
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:
264 DESIGN OF RELIABLE BILINEAR NONREPAIRABLE SENSOR NETWORKS
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
Example 12.3
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
266 DESIGN OF RELIABLE BILINEAR NONREPAIRABLE SENSOR NETWORKS
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
REFERENCES
Ali Y. and S. Narasimhan. Sensor Network Design for Maximizing Reliability of Linear Processes.
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
INTRODUCTION
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
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.
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
Generalized Maximum Reliability Model 271
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.
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:
272 DESIGN OF RELIABLE AND COST-EFFICIENT
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
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
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
Generalized Maximum Reliability Model 273
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
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
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.
Generalized Maximum Reliability Model 275
Using the duality according to Tuy, one can convert this model into maximum
reliability models.
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.
276
Generalized Maximum Reliability Model 277
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.
NOMENCLATURE
ci : cost of sensor i
cT : bound on total cost
Ei : estimability of variable i
E k∗ : threshold of estimability
M1 : set of variables where sensors can be placed
ME : set of key variables for estimability
MP : set of key variables for precision
MR : set of key variables for reliability
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
x: vector of state variables
xM : set of measured variables
Greek Letters
σ: precision
σ ∗: precision threshold
REFERENCES
Ali Y. and S. Narasimhan. Sensor Network Design for Maximizing Reliability of Linear Processes.
AIChE J., 39, 5, pp. 2237–2249 (1993).
Bagajewicz M. Design and Retrofit of Sensor Networks in Process Plants. AIChE J., 43, 9, pp. 2300–
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
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
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
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
Expected Number of Repairs 281
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:
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
282 DESIGN OF REPAIRABLE SENSOR NETWORKS
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
therefore,
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
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
availability of a repairable system does not go to zero, but rather levels off at a
certain time and stays constant.
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)
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 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,
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.
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,
Minimum Cost Model 287
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
S1 — — 0.9
S2 1.5% 2%
S5 2.5% 3%
S7 — — 0.9
Residual
Instrument Precision(%) Precision (%) Availability
Repair Measured Precision (S 2) (S 2) (S 1)
Rate Variables (%) Cost (S 5) (S 5) (S 7)
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
Greek Letters
REFERENCES
Bagajewicz M. Design and Retrofit of Sensor Networks in Process Plants. AIChE J., 43, 9, pp. 2300–
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
INTRODUCTION
—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.
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
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.
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)
294 DESIGN OF ROBUST SENSOR NETWORKS
Cov (r ) = C R Q R C RT (15.2)
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
eR
a = FR+ − F (15.4)
E(a) = 0 (15.5)
eR
Cov (a) = Q (15.6)
Gross Error Detection in Dynamic Data Reconciliation 295
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
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.
The following are the types of errors that one can incur while doing the
hypothesis testing above proposed.
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
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.
σi
ρ(m, α, β) ≤ κD (15.21)
(σi − σ̃i )1/2
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:
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
otherwise.
pi,k ≤ 1∀i
k=1
pi,k = 0,1 ∀i, ∀k = 1, n i m
Example 15.1
x1 x2 x3 x4
1% 3% — 2%
1% 3% 2% —
Minimum Cost Model for Robust Networks 301
Example 15.2
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).
302 DESIGN OF ROBUST SENSOR NETWORKS
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
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
x: vector of state variables
xα/2 : threshold value for measuremnt test
xM : set of measured variables
Greek Letters
REFERENCES
Albuquerque J. S. and L. T. Biegler. Data Reconciliation and Gross-Error Detection for Dynamic
Systems. AIChE J., 42, 10, pp. 2841 (1996).
Bagajewicz M. Design and Retrofit of Sensor Networks in Process Plants. AIChE J., 43, 9, pp. 2300–
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).
304 DESIGN OF ROBUST SENSOR NETWORKS
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).
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).
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
Retrofits. Ellis Horwood, West Sussex, England (1992).
Madron F. and V. Veverka. Optimal Selection of Measuring Points in Complex Plants by Linear
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).
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. 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 ).
Genetic Algorithms 309
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
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
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
312 GENETIC ALGORITHMS
Standard
Failure Deviation of
Variable Probability Error Cost
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
Performance
NOMENCLATURE
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).
314 GENETIC ALGORITHMS
Chmielewski D., T. E. Palmer and V. Manousiouthakis. Cost Optimal Retrofit of Sensor Networks
with Loss Estimation Accuracy. AIChE Annual Meeting, Dallas (1999).
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
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).
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
Fault Detection, Diagnosis and Alarms 317
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.
318 DESIGN OF SENSORS NETWORKS FOR PROCESS FAULT DIAGNOSIS
Node n i F0 L1 F1 L2 F2
Pattern p(n i ) + + + 0 0
Definition: A valid node in the signed digraph is any node which is abnormal
(i.e., p maps the node to + or −).
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.
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
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
322 DESIGN OF SENSORS NETWORKS FOR PROCESS FAULT DIAGNOSIS
Knowledge-Based Approach
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.
Fault Detection, Diagnosis and Alarms 323
FIGURE 17.7. Quantitative model-based fault diagnosis. Adapted from Patton, 1995.
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.
324 DESIGN OF SENSORS NETWORKS FOR PROCESS FAULT DIAGNOSIS
Other Techniques
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.
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.
326 DESIGN OF SENSORS NETWORKS FOR PROCESS FAULT DIAGNOSIS
Alarm Logic
FAULT OBSERVABILITY
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 Location for Fault Observability 327
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
328 DESIGN OF SENSORS NETWORKS FOR PROCESS FAULT DIAGNOSIS
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
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.
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
{R1 , R2 } A1 = {C6 , S N2 , C7 }
R3 A3 = {C7 , S N1 }
R4 A4 = {C7 , C8 }
R5 A5 = {C8 }
Nomenclature 337
Future Scope
NOMENCLATURE
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
Process Industries, Lyon, France (1998).
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
Process Industries, Lyon, France (1998).
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
Industries, Lyon, France (1998).
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).
340 DESIGN OF SENSORS NETWORKS FOR PROCESS FAULT DIAGNOSIS
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
Supervision in the Chemical Process Industries, Lyon, France (1998).
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
Supervision in the Chemical Process Industries, Lyon, France (1998).
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
Industries, Lyon, France (1998).
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
343
344 Index
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