Sie sind auf Seite 1von 48

Process Control and Automation

Richard D. Braatz * and Morten Hovd


A. Introduction
Automation and process control are vital for the safe and economic operation of refining and
petrochemical facilities. Some specific reasons implementing process control systems are:
1. Stabilizing the process. Many processes have integrating or unstable modes. These have
to be stabilized by feedback control, otherwise the plant will move into unacceptable
operating conditions. Feedback also may be needed when an unmeasured disturbance has
a sufficiently large effect on a process variable to cause unacceptably large variations in
the value of a process variable.
2. Regularity. Even if a process is stable, control is often needed to avoid shutdowns due to
unacceptable operating conditions. Such shutdowns may be initiated automatically by a
shutdown system, but also may be caused by outright equipment failure.
3. Minimizing effects on the environment. In addition to maintaining safe and stable
production, the control system should ensure that any harmful effects on the environment
are minimized. This is done by optimizing the conversion of raw materials, and by
maintaining conditions which minimize the production of any harmful by-products.
4. Obtaining the desired product quality. Control is often needed both for achieving the
desired product quality, and for reducing quality variations.
5. Achieving the desired production rate. Control is used for achieving the right production
rate in a plant. Ideally, it should be possible to adjust the production rate at one point in
the process, and the control system should automatically adjust the throughput of up- or
downstream units accordingly.
6. Optimize process operation. It is common in refining and petrochemical facilities to use
automation and control to achieve the most cost-effective production. This involves
identifying, tracking, and maintaining the optimal operating conditions in the face of
disturbances in production rate, raw material composition, and ambient conditions (e.g.,
atmospheric temperature, cooling water temperature). Process optimization often
involves the close coordination of several process units, and operation close to process
constraints.
Even plants of quite moderate complexity would be virtually impossible to operate without
process control. Even where totally manual operation is physically feasible, it is unlikely to be
economically feasible due to product quality variations and high personnel costs, since a high
number of operators will be required to perform the many (often tedious) tasks that the process
control system normally handles.
Usually many more variables are controlled than what is directly implied by the above
list; there are often control loops for variables which have no specification associated with them.
Some reasons for such control loops are
1. To reduce the effect of disturbances from propagating downstream. Even when there are
*

University of Illinois at Urbana-Champaign, 600 South Mathews Avenue, Box C-3, Urbana, Illinois 61801-3602,
U.S.A.

Department of Engineering Cybernetics, Norwegian University of Science and Technology, N-7491 Trondheim,
Norway.

no direct specifications on a process variable, variations in the process variable may


cause variations in more important variables downstream. In such cases, it is prudent to
suppress the effects of disturbances on the upstream variables.
2. Local reduction of the effect of uncertainties. By measuring and controlling a process
variable, it may be possible to reduce the effect of uncertainties with respect to
equipment behavior or disturbances. Examples of such control loops are valve positioners
used to minimize the effect of valve stiction, or local flow control loops which may be
used to counteract the effects of pressure disturbances up- or downstream of a valve,
changes in fluid properties, or inaccuracies in the valve characteristics.
This chapter provides an introduction to the automation and control of large facilities.
First the overall structure of control systems for large facilities and each of the specific
components are described, followed by a discussion of the design of experiments for collecting
data for the purposes of process modeling, control, and monitoring, and the filtering and
preprocessing of the resulting data. The subsequent topics are the derivation of dynamic models
and the estimation from experimental data of the parameters in models, and their use in the
design of feedback and model predictive control systems. The chapter ends with a description of
the design and implementation of process monitoring systems.
B. The Structure of Control Systems in the Process Industries
B.1. Overall Structure
Control systems in the process industries usually have the structure shown in Figure B1.
The lower level in the control system is the regulatory control layer. The regulatory control
system typically controls basic process variables such as temperatures, pressures, flowrates, and
concentrations based on measurements from sensors for temperature (using thermocouples),
pressure (using transducers), flow (using flowmeters), level (by floats, displacement meters, or
differential pressure transducers), and concentration (e.g., by conductivity, pH meters, gas
chromatographs, infrared probes). In some cases the controlled variable may be calculated based
on several measurements, e.g., a component flowrate based on the measurements of both
concentration and the overall flowrate. Most individual controllers in the regulatory control layer
are single-loop implemented in a layered manner, with some multivariable controllers. There are
many heuristics for selecting the order in which the individual controllers are designed and
implemented. Typically feedback control loops whose purpose is to stabilize unstable process
units are closed first, faster loops closed before slower loops, noninteracting loops closed before
interacting loops, loops for handling inventories closed before loops for controlling product
quality, etc. 1,2 Most controllers in the regulatory control layer manipulate a process variable
directly (e.g., a valve opening), but in some cases the manipulated variable may be a setpoint of
a lower level control loop. Most control functions that are essential to the stability and integrity
of the process are executed in the regulatory layer, such as stabilizing the process and
maintaining acceptable equipment operating conditions.
The supervisory control layer coordinates the control of a process unit or a few closely
connected process units. It coordinates the action of several control loops, and tries to maintain
the process conditions close to optimal while ensuring that operating constraints are not violated.
The variables that are controlled by supervisory controllers may be process measurements,
variables calculated or estimated from process measurements, or the output from a regulatory
controller. The manipulated variables are often setpoints to regulatory controllers, but process
variables also may be manipulated directly. Whereas regulatory controllers are often designed

and implemented without ever formulating any process model explicitly, supervisory controllers
usually contain an explicitly formulated process model. The model is dynamic and often linear,
and obtained from experiments on the plant. Typically, supervisory controllers use some variant
of model predictive control (see Section G).
The optimal conditions that the supervisory controllers try to maintain may be calculated
by a Real Time Optimization (RTO) control layer. The RTO layer identifies the optimal
conditions by solving an optimization problem involving models of the production costs, values
of products (possibly dependent on quality), and the process itself. The process model is often
static and nonlinear and derived from fundamental physical and chemical relationships.

Production
planning/scheduling

Real time optimization

Supervisory control

Regulatory control
To manipulated variables

From measurements

Process

Figure B1. Typical structure of the control system for a large plant in the process industries.
The highest control layer shown in Figure B1 is the production planning and scheduling
layer. This layer determines which products should be produced and when they should be
produced. This layer requires information from the sales department about the quantities of the
different products that should be produced, the deadlines for delivery, and possibly product

prices. From the purchasing department, information about the availability and price of raw
materials are obtained. Information from the plant describes which products can be made in the
different operating modes and the production rates that can be achieved.
In addition to the layers in Figure B1, there also should be a separate safety system that
will shut down the process in a safe and controlled manner when potentially dangerous
conditions occur. There are also higher levels of decision making that are not shown, such as
sales and purchasing, construction of new plants, etc. These upper layers do not influence the
design of control systems, and are not discussed further here.
There is a difference in time scale of execution for the different layers. The regulatory
controllers typically have sampling intervals on the scale of one second (or faster for some types
of equipment), supervisory controllers usually operate on the time scale of minutes, the RTO
layer on a scale of hours, and the planning/scheduling layer on a scale of days (or weeks). This
difference in time scales simplifies the required modeling in the higher layers. For example, if a
variable controlled by the regulatory control layer responds much faster than the sampling
interval of the supervisory control layer, then a static model for this variable (usually the model
would simply be variable value = setpoint) usually suffices for the supervisory control.
It is not meaningful to say that one layer is more important than another, since they are
interdependent. The objective of the lower layers are not well defined without information from
the higher layers (e.g., the regulatory control layer needs to know the setpoints that are
determined by the supervisory control layer), whereas the higher layers need the lower layers to
implement the control actions. In some plants human operators perform the tasks of some of the
layers shown in Figure B1, whereas the regulatory control layer is present (and highly
automated) in virtually all industrial plants.
Why has this multi-layered structure for industrial control systems evolved? It is clear
that this structure imposes limitations in achievable control performance compared to a
hypothetical optimal centralized controller which perfectly coordinates all available manipulated
variables in order to achieve the control objectives. In the past, the lack of computing power
would have made such a centralized controller virtually impossible to implement, but the
continued increase in available computing power could make such a controller feasible in the not
too distant future. Is this the direction industrial control systems are heading, in which an
industrial facility is controlled by one unstructured centralized controller? No. In the last two
decades control systems have instead moved towards an increasingly multi-layered structure, as
increased availability of computing power has made the supervisory control and Real Time
Optimization layers much more common. Some reasons for using such a multi-layered structure
are:
Economics. Designing a single centralized control system would require a highly
accurate dynamic model of nearly all aspects of process behavior. The required model
would be highly complex, and difficult and expensive to develop and maintain. In
contrast, the higher layers in a structured control system take advantage of the model
simplifications made possible by the presence of the lower layers. The regulatory control
level needs less model information to operate, since it derives most process information
from feedback from process measurements.
Redesign and retuning. The behavior of an industrial plant changes with time due to
equipment wear, variations in raw materials, changes in operating conditions in order to
change product qualities or which products are produced, and plant modifications. Due to
the sheer complexity of a single centralized controller, it would be difficult and timeconsuming to update the controller to account for all such changes. With a structured

control system, it is easier to see what modifications need to be made, and the
modifications themselves normally will be less involved.
Startup and shutdown. Common operating practice during startup is that many of the
controls are put in manual. Parts of the regulatory control layer may be in automatic, but
rarely will any higher layer controls be in operation. The control loops of the regulatory
control layer that are initially in manual are put in automatic when the equipment that
they control is approaching normal operating conditions. Once the regulatory control
layer for a process section is in service, the supervisory control system may be put in
operation, and so on. Shutdown is performed in the reverse sequence. Based on current
practice, there may be scope for significant improvement of the startup and shutdown
procedures of a plant, as quicker startup and shutdown reduces plant downtime.
However, a model capable of describing all operating conditions including startup and
shutdown is necessarily much more complex than a model which covers only the range
of conditions that are encountered in normal operation. Building such a model would be
difficult and costly. Startup and shutdown of a plant with a single centralized control
system will be more difficult than with a structured control system, because it may not be
difficult to put a large centralized control system gradually into or out of service.
Operator acceptance and understanding. Control systems that are not accepted by the
operators are likely to be taken out of service. A single centralized control system often
will be complex and difficult to understand. Operator understanding obviously makes
acceptance easier, and a structured control system, being easier to understand, often has
an advantage in this respect. Plant shutdowns may be caused by operators with
insufficient understanding of the control system. Such shutdowns are usually blamed on
the control system (or the people who designed and installed the control system). Since
operators are an integral part of the plant operation, the need for operators to understand
the control system well enough to safely and economically operate the facility should be
recognized during control systems design.
Failure of computer hardware and software. In structured control systems, the operators
retain the help of the regulatory control system in keeping the process in operation if a
hardware or software failure occurs in higher levels of the control system. A hardware
backup for the regulatory control system is much cheaper than for the higher levels in the
control system, as the regulatory control system can be decomposed into simple control
tasks (mainly single loops). In contrast, a single centralized controller would require a
powerful computer, and it is therefore more costly to provide a backup system. However,
with the continued decrease in computer cost this issue may become less of an issue.
Robustness. The complexity of a single centralized control system will make it difficult
to analyze whether the system is robust with respect to model uncertainty and numerical
inaccuracies. Analyzing robustness is not always trivial even for structured control
systems. The ultimate test of robustness will be in the operation of the plant. A structured
control system may be applied gradually, first the regulatory control system, then section
by section of the supervisory control system, etc. When a problem arises, it is easier to
analyze the cause of the problem with a structured control system than with a single
centralized control system.
Local removal of uncertainty. It has been noted earlier that one effect of the lower layer
control functions is to remove the influence of model uncertainty as seen from the higher
layers. Thus, the existence of the lower layers enables simpler yet more accurate models
to be used in the higher layers. A single centralized control system does not have this
5

advantage.
Existing control systems. Where existing structured control systems perform reasonably
well, it is more economical and safer to put effort into improving the existing control
system rather than to take the risky decision to design an entirely new control system.
Also, there is a huge experiential knowledge on how to design structured control systems
for various interconnections of process equipment, such as reactors and distillation
columns connected with recycle loops. Designing a structured control system for a new
industrial plant enables the control engineer to draw from past experience to design
various components of the plant instead of starting from scratch with an entirely new
centralized control design. Also, the design of any control system requires carrying out at
least some model identification and validation on the actual plant. During this period
some minimum amount of control will be needed to ensure safe and stable operations.
The regulatory control layer of a structured control system requires more limited
information, and therefore can be in operation during this model identification and
validation.
From this discussion it should be clear that control systems will continue to have a number of
distinct layers for the foreseeable future. Two prerequisites necessary for a structured control
system to be replaced with a single centralized control system are:
1. The existing control system gives unacceptable performance.
2. The entire process must be sufficiently well understood to be able to develop a process
model which describes all relevant process behavior.
Since it is quite rare that a structured control system is unable to control a process for which
detailed process understanding is available (provided sufficient effort and expertise have been
put into the design of the control system), it should follow that majority of control systems will
continue to be of the traditional structured type.
B.2. Common Control Loop Structures for the Regulatory Control Layer
The common control loop structures for the regulatory control layer are described here.
B.2.a Simple Feedback Loop
The simple feedback loop is by far the more common control loop structure for the
regulatory control level (see Figure B2). The controller acts on the difference between the
desired value for the process output (which is called the setpoint or reference value) and the
measurement of the process output. The controller manipulates an input to the process to make
the measurement follow the setpoint. Note that the measured value may not equal the actual
process output value, due to possible measurement noise or sensor malfunctions. Note also that
the manipulated variable is normally one of several process inputs which affect the value of the
process output; the additional process inputs which are not manipulated by the controller are
disturbances. The need for feedback of the process measurement arises from uncertainty both
with respect to the value of the disturbances, and with respect to the process response. Namely, if
we could know exactly the value of all disturbances, and the response of the process to both the
disturbances and the manipulated value, the measurement would be superfluous, since we would
know the exact value of the process output for a specific value of the manipulated variable. In
practice such exact process knowledge is unrealistic, and hence feedback of the measurement is
needed if accurate control of the process output is required.
Section F describes the design of feedback control loops.

Disturbances

Setpoint
_

Feedback
controller

Manipulated
variable

Process

Process
output

+
Noise

Measurement

Figure B2. Simple feedback control loop.


Disturbances
Feedforward
controller

Setpoint
_

Feedback
controller

+
Manipulated
variable

Process

Process
output

+
Noise

Measurement

Figure B3. Feedforward control for measured disturbances combined with feedback control.
B.2.b Feedforward Control
Feedforward control is used to counteract the effect of disturbances without first having
to wait for the disturbances to affect the process output (see Figure B3). Feedforward control
from measured disturbances combined with simple feedback control, that is,
u = u ff + u fb
(B1)
where u ff and u fb are the outputs of the feedforward and feedback controllers, respectively. The
ideal feedforward signal would exactly cancel the effect of the disturbance on the controlled
variable. For example consider a static single-input single-output nonlinear process in which
y = P(u ff ) + Pd (d )
(B2)

where y is the process output (which is the controlled variable), d is a disturbance, and P and Pd
are algebraic relationships for the process model and disturbance model, respectively. Then the
ideal value for the output of the feedforward controller is
y = P (u ff ) + Pd ( d ) = 0 u ff = P 1 ( Pd ( d ))

(B3)

assuming that the process model P is invertible for the full-range of process inputs and outputs.
This feedforward approach generalizes to take into account dynamics and multiple variables, in
7

which case more complex plant inverses are utilized. 3,4 For feedforward control to give good
performance, both the process model P and the disturbance model Pd must be quite accurate. 5
Since the feedforward controller cannot be expected to be perfect, it needs to be augmented with
a feedback controller, as shown in Figure B3, if accurate control is required.
The feedforward controller is always designed to be stable, that is, it produces a bounded
output for any bounded input. A pure and stable feedforward controller cannot by itself induce
instability of the closed-loop system, although it will result in poor performance if the model
used to design the feedforward controller is inaccurate. When the plant and disturbance models
are accurate and the measured disturbances have a large effect on the controlled variables, then
feedforward control can give significant improvements in closed-loop performance. On the other
hand, feedforward cannot be used to stabilize an unstable process.
B.2.c Ratio Control
Ratio control may be used whenever the controlled variable is strongly dependent on the
ratio between two inputs. Simple examples of control problems where this type of control
structure is appropriate are
Mixing of hot and cold water to get warm water at a specified temperature.
Mixing of a concentrated chemical solution with a diluent to obtain a dilute chemical
solution.
Ratio control may be considered as a special case of feedforward control. It is particularly
appropriate when one of the two inputs cannot be controlled, but varies rapidly. Measuring the
input that cannot be controlled and applying the other input in a specific ratio to the uncontrolled
one, essentially amounts to feedforward control. Figure B4 illustrates a typical application of
ratio control in mixing two streams.
Flowrate
controller
Setpoint

FC

Flowrate
measurement

XC

FT

Controlled stream

Multiplier

Property
controller

Property
measurement
XT

Control
valve
Mixer

Outflow to
downstream
operations
Uncontrolled stream

FT

Flowrate
measurement

Figure B4. Ratio control for mixing two streams to obtain some specific property for the mixed
stream. The property controller XC manipulates the multiplication factor, and thereby also the
ratio between the two streams.

B.2.d. Cascade Control


Cascade control is used in cases where an intermediate measurement can give an
indication of what will happen with the more important primary measurement being further
downstream (see Figure B5). Cascade control also was used in Figure B4, since the property
controller (via the multiplier) manipulates the setpoint to the flow controller instead of the valve
position itself. In Figure B4, the flow controller will counteract disturbances in the upstream
pressure and correct for a possibly nonlinear valve characteristic.
Inner
disturbances

Intermediate
setpoint

Primary
setpoint

Outer
feedback
controller

Inner
feedback
controller

Outer
disturbances

Manipulated
variable
Inner
process

Outer
process

Process
output

Intermediate measurement
Primary measurement

Figure B5. Cascaded control loops. The objective of the inner feedback controller is to cause the
intermediate measurement to track the intermediate setpoint while suppressing the effect of
disturbances entering the inner control loop. The objective of the outer feedback controller is to
cause the primary measurement to track the primary setpoint, which treats the setpoint for the
inner controller as its manipulated variable, while suppressing the effect of disturbances entering
the outer control loop.
In general, there may be more than two loops in cascade. For instance, a valve positioner
can get its setpoint from a flow controller, which in turn gets its setpoint from a level controller
(i.e., three loops in cascade). For cascade control to make sense, the inner loops must be
significantly faster than the outer loops since the intermediate process measurements are of
little interest. If the inner loops do not provide for faster disturbance suppression of at least some
disturbances, then the inner loops are not that useful. Fast inner loops also make the tuning of the
outer loops simpler, since then the outer loops can be designed with the assumption that the inner
loops are able to follow their setpoints.
B.2.e Auctioneering Control
Auctioneering control is a control structure where the worst of a set of measurements is
selected for active control, that is, the measurement that places the highest bid is used in the
control system. This type of control structure is particularly common in some chemical reactors
with exothermal reactions, where the process fluid flows through tubes filled with solid catalyst.
If the temperature becomes too high, the catalyst will be damaged or destroyed, therefore the
tubes are cooled by a cooling medium on the outside. On the other hand, if the temperature is too
low, the reactions will be too slow. Thus temperature control is very important. However, the
temperature will vary along the length of the reactor tubes, and the position with the highest
temperature will vary with operating conditions. Therefore several temperature measurements

10

along the reactor length are used, and the value of the highest temperature is chosen as the
controlled variable. This arrangement is illustrated in Figure B6.
A better approach may appear to be to use all of the temperature measurements as inputs
to an estimator that estimates the maximum temperature. Such an estimator could estimate the
maximum temperature when the maximum does not occur at the position of a temperature
measurement, and could also be made more robust to measurement malfunction (if properly
designed). However, this type of chemical reactor is normally highly nonlinear, and the estimator
would therefore need a nonlinear model, probably based on physical and chemical relationships.
The modeling work needed could be time consuming, and it also could be difficult to ascertain
that the estimator performs well in all operating regions. Thus, the auctioneering approach is
more often used in practice.
High select

>
TT1

TT2

TT3

TT4

TT5

TT6
Products

Reactants

Cooling medium
TC1
Temperature
controller

Figure B6. A chemical reactor with auctioneering temperature control.


B.2.f. Split-range Control
In split-range control, several manipulated variables are used to control one controlled
variable, in such a way that when one manipulated variable saturates, the next manipulated
variable takes over. In order to obtain smooth control, there is often overlap between the
operating ranges of the different manipulated variables. For example, manipulated variable 1
may take value 0% at a controller output of 0%, and value 100% at controller output 60%.
Similarly, manipulated variable 2 takes value 0% for controller outputs below 40%, and value
100% for controller output 100%.
It should be clear that there can be a lot of freedom in how to design the split-range
arrangement. To simplify the control design, a common industrial practice is to use this freedom
to make the response in the controlled variable to changes in the controller output as linear as
possible. Sections F2-F3 describe how to design split-range control systems.
B.2.g. Combining Basic Control Structures
Most of the basic control structures may be combined with each other. All the control
structures are variants of feedforward or feedback control, with feedforward control normally

10

11

combined with some form of feedback control. Feedforward and the basic feedback control
structures also can be combined with multivariable controllers, for example, usually the model
predictive controllers discussed in Section G are usually implemented above lower control loops
in a cascade control arrangement.
C. Experimental Design, Sensor Calibration, and Data Filtering
Most control systems are designed based on models, ranging from the very simple to
quite complex, so the closed-loop performance obtained by the control system depends on the
accuracy of the model. Models are constructed from experimental data, so that accuracy of the
model depends on the quality and quantity of data. The quality of data depends on the details of
the experimental design and any preconditioning of the data, such as sensor calibration and data
filtering.
C.1. Experimental Design
An experimental study should be designed to adequately cover the entire range of
operations. Covering a wide range of operations enables more accurate predictions to be
obtained using statistical data analysis and model building. This provides more confidence when
scaling up bench-top results to the factory production scale. Data should not be collected only
during best practices or around any prior assumed optimal operating conditions, since a goal of
a model constructed from the data is to provide accurate predictions even during abnormal
operating conditions.
There are numerous papers and texts that describe statistical experimental design in some
detail. 6,7 Such procedures are valuable when used to construct accurate models with the least
number of experiments. A general procedure for designing and carrying out an experimental
study can be organized into 8 major steps:
1. Clearly state the objectives of the experimental study.
(a) Use any reliable information gained from books, journal papers, and previous
experiments.
(b) State any assumptions or hypotheses that will be tested during the experimental study
(e.g., that there exists a linear relationship between two variables).
2. Draw up a preliminary experimental design.
(a) Choose experimental materials, procedures, and equipment (e.g., solvents, mixers,
actuators).
(b) Choose variables that will be fixed during the experimental study (e.g., the volume of
a sample). Choose variables that may be varied and determine the practical range of
these variables.
(c) Choose measurements and methods of measurement (i.e., sensors). Consider the
precision of the measurement methods. The more measurements and measurement
methods, the more reliable the results. Consider using multiple sensors for each
measured variable (e.g., measuring temperature with a thermometer and a
thermocouple), especially when additional sensors are cheap.
(d) Make sure that the entire region of interest is covered (i.e., that the full variable
ranges of interest are considered).
(e) Make sure that some of the experiments are repeated. This allows the measurement
noise to be quantified by calculating the variance, and helps to detect gross errors in
the measurements.
(f) Consider whether the order of the experiments should be randomized. Does
randomization justify any added costs?

11

12

3.

4.

5.

6.

7.

(g) List possible alternative outcomes.


(h) Consider the cost and time of experimentation versus the cost of reaching wrong
conclusions. If experiments are cheap, repeating all experiments will reduce biases
and increase the accuracy of the conclusions and models constructed from the
experimental data.
Review the design with all collaborators (e.g., team members, teaching assistant).
(a) Reach an understanding as to what decisions hinge on each outcome. Keep notes.
(b) Encourage collaborators to anticipate and list all experimental variables that might
affect the results (e.g., variations in room temperature, exposure of the samples to
air).
(c) Discuss the experimental techniques in sufficient detail to discover any procedures
that might lead to bias (e.g., one person reading the temperatures from a thermometer
on one day or experiments, another person reading the temperature during a second
day of experiments).
Draw up the final experimental design.
(a) Describe the experimental design in clear terms to ensure that its provisions can be
followed without any confusion.
(b) Include the methods of data analysis as part of the experimental design, ascertaining
that the conditions necessary for the validity of the data analysis methods will be
satisfied.
Carry out the experimental design.
(a) Record all data and any modifications to the experimental design. Make sure that all
data are recorded correctly during the experiment. Label the data by date, run
number, and other ancillary data (e.g., air temperature).
(b) During the course of carrying out the experiment, maintain constant communication
among all collaborators, so that questions arising from unforeseen experimental
conditions or results may be answered using your collective knowledge. This ensures
that each collaborators expertise and/or insights are brought to bear.
(c) If unforeseen experimental conditions or results occur, modify the experimental
design as necessary.
(d) Always think about the experiments while they are being carried out. DO NOT carry
out the experiments with no attention to the results being obtained. Ask the questions:
Do the measurements seem to make sense? If not, why not?
Analyze the data.
(a) Review the data with attention to recording errors, omissions, etc.
(b) Use graphics; plot the data, plot averages, plot simple graphs.
(c) Apply the data analysis methods outlined in Step 4.
(d) Use additional data analysis methods if required by a change in the experimental
design (Step 5c).
Interpret the results.
(a) Consider all the observed data.
(b) Confine initial conclusions to strict deductions from the experimental evidence.
(c) Elucidate the analysis in both graphical and numerical terms (e.g., plot relationships
between variables, quantify accuracy of numerical results).
(d) Arrive at conclusions as to the technical meaning of the results as well as their
statistical significance.

12

13

(e) Point out implications of the findings for application and for further work.
(f) List any significant data limitations discovered during the data analysis. Relate these
to any limitations on the experimental design that was conducted.
(g) Compare the results to any related results reported in the literature.
8. Write the report.
(a) Describe the work clearly, giving background, pertinence of problem(s) being
investigated, meaning of results.
(b) Present the data and results in clear tables and graphs, and describe their possible
future use.
(c) Compare the results with the stated objectives of the experiment (Step 1). Limit
conclusions to objective summary of the evidence.
(d) If the results suggest that further experimentation is desired or necessary, outline how
additional experiments should be carried out (repeat Step 4).
This procedure for experimental design is appropriate regardless of whether the data are used to
construct empirical models (that is, based only on fits to data) or first-principles models (that is,
based on material, energy, and/or momentum conservation equations). Readers interested in
further discussions on experimental design are referred to books devoted to the subject.6,8,9
C.2. Sensor Calibration
The quality of experimental data strongly depends on the quantity and quality of the
measurement data. Usually many of the important variables cannot be directly measured, but can
be indirectly measured. For example, directly measuring solution concentrations is rarely
possible but there are infrared or ultraviolet spectroscopy probes that can measure spectra, and
there may be a mapping from the spectra to the solution concentrations. Such calibration curves
(also called soft sensors or inferential models) are constructed from sensor data collected for
known samples, which then can be applied to unknown samples to obtain an indirect
measurement of the variables of interest.
Most calibration curves are linear functions of the sensor signals, that is, a linear model
relates a variable to predict, y, to the n sensor signals ai:
n

y = i ai + e = a + e = y + e

(C1)

i =1

where a is the row vector of sensor signals, is a column vector of calibration parameters, and e
is measurement noise typically assumed to be zero-mean and independent. For N known samples
the variables can be stacked into vectors and matrices
y1 a1
e1


Y = # = # + # = A + E
y N a N
e N

(C2)

where the superscript refers to the sample number. If the variance of each measurement noise is
equal for all measurements, then the calibration parameters should minimize the sum of squared
deviations between the calibration model, y = a , and the predicted variables, y,
N

min E 2j = min E T E = min (Y A )T (Y A )

j =1

13

(C3)

14

where the superscript


analytically:

is the matrix transpose. The calibration parameters can be computed

= [ AT A]1 ATY = AY

(C4)

where A is the pseudo-inverse of A. If the matrix A is square then this can be simplified to

= ( AT A) 1 ATY = A1 ( AT )1 ATY = A1Y ,

(C5)

using the expressions


(C6)
(AB) 1 = B 1 A1 , A1 A = I , and AI = A .
If the covariance matrix of the measurement noise cov{E} = V, then the best estimate of the
parameters is6

= ( ATV 1 A)1 ATV 1Y .

(C7)

As an example, consider a distillation column in which the concentration of a


hydrocarbon in the distillate is a product quality variable of importance that can be indirectly
measured using an infrared sensor probe inserted into a pipe. It is observed that the signal from
the infrared sensor varies with temperature, so a thermocouple is located next to the sensor so the
effects of temperature on the calibration can be taken into account. Two absorbance peaks are
observed in the infrared spectra, at 2000 cm-1 and 3000 cm-1, when the sensor probe is inserted
into 10 samples of known hydrocarbon concentration. Then the predicted variable y is the
hydrocarbon concentration, the vector of sensor signals is
a1 absorbance at 2000 cm 1

a
absorbance at 3000 cm 1
aT = 2 =

a3
temperature


1
a4

and the calibration model is written in terms of the sensor signals as

(C8)

y = hydrocarbon concentration = a = i ai = 1 (absorbance at 2000 cm 1 )


i =1

(C9)

+ 2 (absorbance at 3000 cm 1 ) + 3 (temperature) + 4

Note that the 1 as the last element of (C8) resulted in the inclusion of a constant bias term, 1,
in the calibration model (C9). The stochastic fluctuations in the sensor signals can be estimated
by measuring the same sample many times; let the variances of the sensor signals be written as
1, , 4, then
12 0
0
0

0
0 22 0

V=
0
0 32 0

0
0 42
0
The known hydrocarbon concentrations of the 10 samples are stacked to give

14

(C10)

15

y1

Y =#
y10

(C11)

and the 10 sets of sensor signals would be stacked to give


a11 a12 a31 a14
2

a1 a22 a32 a42

(C12)
A=
#
#
#
#
10
10
10
10
a1 a2 a3 a4
where the superscript refers to the sample number. Then the best calibration parameters are
given by (C7).
While the above method, known as weighted least squares, can produce good calibration
models when the number of calibration parameters is low, it performs poorly when the number
of calibration parameters is high, due to high correlations between the sensor signals. There are
many numerical algorithms for constructing correlations when such correlations occur, which go
by names such as principal component regression and partial least squares. Such algorithms are
referred to as chemometrics (for chemistry measurement), for which many textbooks and
software packages are available. 10,11
In some cases the relationship between the sensor signals and the predicted variables is
nonlinear, in which case the minimization
n

T
min E 2j = min E T E = min Y Y ( ) V 1 Y Y ( )

j =1

(C3)

is solved numerically (see Section E1 for more information on nonlinear optimization).


C.3. Filtering
Sensor signals nearly always contain short-time fluctuations. This means that the
measured quantity will show fluctuations that are much more rapid than the actual process is
varying. This section describes how to smooth out the signals from the sensors, so that it more
closely represents the actual variables being measured. This filtering step is often performed
before the data are used for modeling and process control.
C.3.a Exponential Filtering
A filter smooths fluctuations in noisy signals and can be implemented by processes,
digitally in computers, or in analog instrumentation. As an example of filtering by a process,
placing a surge tank between an unsteady process stream and the unit it feeds will smooth out
fluctuations in the feed (see Figure C1). An analog filter may consist of an RC electrical circuit,
which contains a capacitance (C) and a resistance (R) to flow of current, just like the tank
(capacitance) with a valve (resistance) at the outlet. A common digital implementation is called
an exponential filter, which is described by the ordinary differential equation (ODE):
dy (t )
F
+ y (t ) = x(t )
(C13)
dt
where x(t) is the raw unprocessed measurement, y(t) is the filtered output, t is time, and F is the
time constant of the filter which has the same units of time as t. Computers operate on digital
signals, with the derivative in this ODE typically approximated by a backward difference:

15

16

dy y (tn ) y (tn 1 ) yn yn 1
(C14)
=
=
dt
t
t
where t is the time between data points sampled at times t0, t1, , tn. Inserting (C14) into
(C13) and rearranging gives:
(C15)
yn = axn + (1 a ) yn 1
where = 1/[1 + (F/t)], and the filtered output is initialized by y0 = y(0) = x(0). This exponential
filter is also called a first-order filter. The order of the filter is the number of derivatives
required to describe the filter when written as an ODE, which is one for the ODE (C13). This
definition of order applies to processes, controllers, or closed-loop systems.

flow
time

Surge
tank

Reactor
(for example)

flow
time

Figure C1. A surge tank filters concentration and flow fluctuations.


The filter constant, , contains the time constant F of the hypothetical first-order process
that is being used as a filter. The filter constant can range from 0 to 1, with = 1 equivalent to no
filtering, and with smaller positive values providing strong filtering. For = 0, the present
measured value is not even used in the output, as the time constant of the filter is infinite, which
corresponds to an infinite amount of filtering. In practice the filter constant is selected so that
does not significantly change overall shape of the measured data, but reduces noise as much as
possible. This is a tradeoff between small values of (near 0), which bias the filtered signal y,
and large values of (near 1), which provide inadequate filtering of noise in y.
C.3.b Double Exponential Filtering
Placing two exponential filters in series produces a double exponential filter:
dy (t )
F1
+ y (t ) = x(t )
dt

16

(C16)

17

dz (t )
+ z (t ) = y (t )
(C17)
dt
where x(t) is the raw unprocessed measurement, y(t) is the filtered output from one exponential
filter, and z(t) is the filtered output of the double exponential filter. The time constants F1 and F2
in the double exponential filter are not required to be the same, but are usually the same in
practice. The double exponential filter provides better attenuation of high frequency noise than
does the exponential filter. The digital implementation of this filter is:
(C18)
yn = 1 xn + (1 1 ) yn 1
(C19)
zn = 2 yn + (1 2 ) zn 1

F2

where 1 = 1/[1 + (F1/t)] and 2 = 1/[1 + (F2/t)]. For F1 = F2 this can be written as
zn = 2 xn + 2(1 ) zn 1 (1 ) 2 zn 2

(C20)

Note that xn1 and xn2 do not appear in this equation.


For both the exponential filter and the double exponential filter, decreasing the filter
constant (increasing the time constant of the filter) gives a smoother output, at the cost of
having a more sluggish response. The choice of a filter constant is a design decision; ideally, a
large enough value will be chosen to dampen out high frequency noise, without significantly
altering the measured process dynamics. Below are the first several computations performed by
single and double exponential filters on a raw signal x(t) with = 0.5.
t
0
1

x( t )
5
6

Exponential Filtered y(t)

.5(4) + (1.5)5.5 = 4.75

(.5) 2 (4) + 2(1.5)6 (1.5) 2 (5) = 5.75

.5(4) + (1.5)4.75 = 4.375

(.5) 2 (4) + 2(1.5)5.75 (1.5) 2 (6) = 5.25

.5(5) + (1.5)4.375 = 4.6875

(.5) 2 (5) + 2(1.5)5.25 (1.5) 2 (5.75) = 5.0625

Double Exponential Filtered y(t)

.5(6) + (1.5)5 = 5.5

D. Dynamic Modeling
This section discusses the development of first-principles process models, that is, models
constructed from material, energy, and momentum balances. The balances can be written in
terms of a general conservation equation:
accumulation = in out + generation consumption
(D1)

where the terms are usually expressed in terms of rates. Since the dynamics are critically
important for controlling the process, these balance equations should include the accumulation
term. A steady-state balance should only be used when it is known without a doubt that the
accumulation term is negligible.
As an example of a conservation equation, the general energy balance is:
rate of accumulation of energy = rate of input of energy
rate of output of energy
+ rate of generation of energy
rate of expenditure of energy via work

17

(D2)

18

A typical method of generating energy is via chemical reaction. A typical method of energy
input is that associated with the enthalpy of a flowing inlet stream.
The material balance can be written in terms of the number of moles of a particular
species, the mass of a particular species, the total mass, or the number of particles (e.g., atoms,
molecules, crystals). For example, the mole balance for species A is
rate of accumulation

= rate of input of moles of A


of moles of A
rate of output of moles of A
+ rate of generation of moles of A
rate of consumption of moles of A

(D3)

The input and output of moles of species A are associated with the inlet and outlet streams; the
rates of generation and consumption of moles of species A are associated with chemical
reactions. The total mass balance for a system in which there are no nuclear reactions is:
rate of accumulation

= rate of input of total mass rate of output of total mass


of total mass

(D4)

The rates of generation and consumption of total mass are zero.


As an example, consider the mass and energy balances for the liquid in an electrically
heated well-mixed tank with level h and volumetric inlet flows or Fi and Fo. Assuming that the
tank is perfectly mixed and that the liquid density of the inflow and outflows are equal, the total
mass balance is
d
(D5)
( Ac h(t ) ) = Fi (t ) Fo (t )
dt
accumulation = in out

where Ac is the cross-sectional area of the tank, and is the liquid density. Assuming that the
density and cross-sectional area are constants gives
dh(t )
Ac
= Fi (t ) Fo (t ) .
(D6)
dt
With the assumptions of perfect mixing, equal liquid densities of the inflow and outflow streams,
and negligible work due to mixing, the total energy balance is
d
Ac h(t ) ( Cv (T (t ) Tref ) + U ref ) = Q (t ) + Fi (t ) ( C p (Ti (t ) Tref ) + H ref )
dt
(D7)
Fo (t ) ( C p (T (t ) Tref ) + H ref )

accumulation = in + in out
where T is the temperature, Q is the heat rate applied through the electric heater, Cv is the
constant-volume heat capacity on a per-mass basis, Cp is the constant-pressure heat capacity on a
per-mass basis, Tref is the reference temperature, Uref is the reference internal energy on a permass basis, and Href is the reference enthalpy on a per-mass basis. This equation assumes that the
heat capacities are constant over the range of temperature of interest.

18

19

Many more examples on the writing of dynamic material and energy balances are given
in textbooks on process controls, 12,13,14 chemical reaction engineering, 15 and transport
phenomena. 16 It is common in simulation and control software to write these models in statespace form:
dx(t )
= f ( x(t ), u (t ))
(D8)
dt
y (t ) = g ( x(t ), u (t ))
where y, u, and x are the vectors of model outputs, model inputs, and states, respectively, and f
and g are algebraic functions. It is common in the design of control systems to replace the
algebraic functions f and g by Taylor series expansions
dx(t )
= Ax(t ) + Bu (t )
(D9)
dt
y (t ) = Cx(t ) + Du (t )
where y, u, and x are the vectors of model outputs, model inputs, and states, respectively, and A,
B, C, and D are matrices of compatible dimensions. The linear model (D9) is convenient for
determining stability, that is, whether the model outputs are bounded for any bounded inputs. In
particular, the linear model (D9) is stable if and only if the real parts of all of the eigenvalues of
A are negative.
As an example, consider the well-mixed tank but with no electric heating. Then the sole
conservation equation is (D6) which can be written as
1
dx(t ) 1
1
1 F (t )
Fi (t ) Fo (t ) = 0N x(t ) +
=
i
dt
Ac
Ac
Ac Fo (t )
Ac
A


B
u (t )
y (t ) = x(t ) = 1x(t ) + 0u (t ) = Cx(t ) + Du (t )

(D9)

where the model output is equal to the state which is the height of the tank, y = x = h, and the two
flowrates are the model inputs collected into the vector u. Note that D = 0, which is true for
nearly all models of industrial processes. The eigenvalue of A is 0, so the linear model (D9) is
unstable any constant value for either model input ui results in the model output increasing or
decreasing without bound.
Many multivariable control systems use a discrete-time representation for (D9), in which
the values of the states and model outputs are computed only at discrete sampling instances. The
simplest method to derive such a discrete-time representation is to replace the derivative by its
forward difference approximation:
1
( x(t + t ) x(t ) ) = Ax(t ) + Bu (t )
(D10)
t
y (t ) = Cx(t ) + Du (t )
where t is the sampling interval, that is, the time between sampling instances. This can be
rearranged to give:
x(t + t ) = x(t ) + tAx(t ) + tBu (t ) = ( I + tA) x(t ) + tBu (t )
(D11)
y (t ) = Cx(t ) + Du (t )
or

19

20

(t ) + Bu
(t )
x(t + t ) = Ax
y (t ) = Cx(t ) + Du (t )

(D12)

The linear discrete-time model (D12) is stable if and only if the magnitude of all eigenvalues of
A is less than 1.
E. Parameter Estimation
While some model parameters are known and directly measured rather easily (e.g., mass,
density), other parameters require estimation from experimental data. For example, kinetic
parameters associated with chemical reactions usually are identified from experimental data
because the kinetics computed from molecular theories are not sufficiently accurate. This section
describes how to estimate parameters from dynamic data, and to quantify the accuracy of the
parameters so it can be determined whether enough experimental data have been collected to
obtain sufficiently accurate model parameters.
E.1 Formulation of Parameter Estimation as an Optimization
Parameter estimation is the process of fitting the simulation outputs to experimental data
to estimate the unknown parameters. Typically the parameters are estimated by minimizing a
weighted sum of squared errors between the model predictions and the measured variables:
Nm Nd

min wij ( yij y ij ( )) 2

(E1)

i =1 j =1

where is the vector of parameters, yij and y ij are the measurement and model prediction of
the i th measured variable at the j th sampling instance, wij is a weighting factor, N m is number
of measured variables, and N d is the number of sampling instances. To compute the best
estimates of the parameters, each wij in (E1) should be set equal to the inverse of the
measurement error variance i2 , where i is the standard deviation for the error in the i th
measurement. 17 This selection of weights has a lower weighting on the measurements that are
noisier.
The optimization (E1) is more general than that used for sensor calibration (D3), since
(E1) can be applied to experimental data from dynamic processes and the model in (E1) can be
nonlinear in the parameters. When the model y ij is linear in the parameters then (E1) can be
solved analytically in a similar manner as for (D3). That is, stack the yij and y ij into vectors

#

Y = yij ;
#

(E2)

#


Y = yij = A .
#

(E3)

Then (E1) can be written as

20

21

) (

T
T
min Y Y ( ) W Y Y ( ) = (Y A ) W (Y A ) ,

(E4)

where wij is the (i,j) element of the matrix W. The solution to this minimization is given by (C7)
with V replaced by W1.
The optimization (E1) is generalized to multiple experimental data sets by including an
additional summation in the objective function.17 When the model is nonlinear in the parameters,
then the optimization (E1) usually cannot be solved analytically, and hence numerical methods
for optimization are applied; commercial software for doing these calculations include Matlab, 18
IMSL, 19 or FFSQP. 20
E.2. Quantifying the Accuracy of the Parameters
The accuracy of the model parameters can be quantified using multivariate statistics. Due
to the stochastic fluctuations associated with measurements, the parameter estimates are also
stochastic variables with probability distributions. An approximate confidence region for the
parameters can be obtained by linearizing the model near the vicinity of the estimate:17
y j ( ) y j ( ) + Fj ( ) ( ) ,

(E5)

where y j = [ y 1, j ,, y Nm , j ]T is the vector of model predictions at the j th sampling instance, is


the vector of best parameter estimates, and Fj is the N m N p sensitivity matrix given by

Fj =

y j

(E6)

The sensitivity matrices Fj can be calculated using finite differences or by integrating the
sensitivity equations along with the model equations. 21 It is normally acceptable to assume that
the measurement errors are normally distributed and independent of each other, that is, the
measurement error covariance matrix V is diagonal with the diagonal entries, Vii = i2 . Then the
parameter covariance matrix V for the linearized problem is given by
Nd

V = FjT V 1 Fj
1

(E7)

j =1

The approximate 100(1 )% confidence region is the hyperellipsoid defined by


( )T V1 ( ) N2 p ( )

(E8)

where 2 is the chi-squared distribution which is reported in introductory books on statistics.8


This confidence ellipsoid generalizes the notion of a confidence interval used for single
parameters to multiple parameters. The eigenvectors of V1 give the direction and the
eigenvalues give the length of the axes of the hyperellipsoid. Because it is impossible to
visualize the hyperellipsoid for more than three dimensions, in these cases confidence intervals
are often reported:
i* t1 / 2 ( Nd N p ) V ,ii i i* + t1 / 2 ( Nd N p ) V ,ii

(E9)

where t1 / 2 ( Nd N p ) is the t statistic for Nd N p degrees of freedom and V ,ii is the (i, i ) element of

V . Note that these confidence intervals on each model parameter are not as good of a

21

22

quantification of the accuracy of the model parameters as the original confidence hyperellipsoid
(E8).
E.3 Chemical Reactor Example
Consider a reactor for which the reactant A forms the product B ( A B ) with kinetic
rate law,

rA = ke E /RT C A ,

(E10)

where C A is the molar concentration of species A , T is temperature, R is the ideal gas


constant, k is pre-exponential factor, E is the activation energy, and rA is the net rate of
generation of A. Assume that the experiments are carried out in a well-mixed batch reactor with
initial concentration CA0 , and that the volume remains constant throughout the reaction. A
thermocouple and an infrared sensor are inserted into the reactor to measure the temperature and
reactant concentration, respectively (Section C2 describes how to relate infrared spectra to
concentration). Assume that the concentration of A and the temperature can be measured once a
minute during the experiments, the total time for one batch run is one hour, and the temperature
measurement is twice as accurate as the concentration measurement (in terms of error variance).
The goal is to estimate the pre-exponential factor k and the activation energy E from the
experimental data from one batch run.
Since the measured variables are T and C A , N m = 2 . The measurements are taken every
minute for sixty minutes
t0 = 0 min
t1 = 1 min
(E11)
#
t60 = 60 min
so N d = 60 . The error variances of the measurements are related by:

T2 = C2 / 2 .

(E12)

The parameter vector is

= .
E

(E13)

The weighting factors are


w1 j = 1/ T2 ;

w2 j = 1/ C2 A .

The measurement and model prediction for the measured variables are
y =T ;
y = T ;
1j

1j

y 2 j = C A, j

y2 j = C A, j ;

Assume the tank is well-mixed, then the mole balance on A is


d 
(C AV ) = rAV
dt
accumulation = net generation consumption

22

(E14)

(E15)

(E16)

23

Since the volume V is constant,


dC A

= ke E /RT C A
dt

C A dC
t
E /RT ( t )
dt
CA0 C A = 0 ke

(E17)
(E18)

t
C

ln A = k e E /RT ( t ) dt
0
C A0

(E19)

t E /RT ( t )
dt

k e
C A = C A0 e 0

(E20)

t j E /RT ( t )
e
dt
0

k
C A, j = C A (t j ) = C A0 e

(E21)

Tj = T (t j )

(E22)

Inserting into (E1) gives


2

60

min wij ( yij y ij ) 2

(E23)

min w1 j ( y1 j y 1 j ) 2 + w2 j ( y2 j y 2 j ) 2

(E24)

k ,E

i =1 j =1

60

k ,E

j =1

Hence the best-fit parameters k* and E* are solutions to the minimization


60

min
k ,E

j =1

2
T


j

2
k e E /RT ( t ) dt
1

0


T j T (t j ) + 2 C A, j C Ao e

which can be solved numerically. To compute the confidence intervals, first note that

T (t j )

Tj ( * )
*

=
tj
y j ( ) = 

E/RT ( t )
*

k
e
dt

C A, j ( ) C A0 e 0

(E25)

(E26)

Hence the sensitivity matrices are


Tj

y j
1
Tj

Fj =
=
=


= * C A, j
= * C A, j
1

Tj

2
C A, j

2 = *

(E27)

Since the temperature T (t ) is not a function of the parameters k and E (it was specified by the
person who designed the batch experiment),
Tj
= [ 0 0] ;
(E28)

and

23

24

C A, j
1

C A, j
k

tj



k e E/RT ( t ) dt
0
C e

k A0
t j E /RT ( t )
e
dt
0

e
0

= C e k
A0
C A, j
2

C A, j

t j E /RT ( t )
e
dt

= C e k 0
A0
E

(E29)
tj

E /RT ( t )

dt

tj

E /RT ( t )
dt
k

0 E e

tj

1

dt
= C A (t ) k e E /RT (t )
 (t )
0
RT

Substituting these equations into (E27) gives


t
Fj = C A (t j )
j e E /RT (t ) dt
0
The error variance matrix is
V=

(E30)

12

2
2

.
k

e E /RT (t ) dt
RT (t )

tj
0

T2

2
C A

(E31)

(E32)

Inserting these expressions into (E7) and simplifying the algebra gives the parameter covariance
matrix:
tj


e E /RT ( t ) dt

t j E /RT (t )
k t j 1 E /RT (t )
(E33)
V1 =
e
dt
e
dt

t
R 0 T (t )
C2A j =1 k j 1 e E /RT (t ) dt 0

R 0 T (t )

and the confidence region and intervals are given by (E8) and (E9).
The parameter estimates and confidence region for this example was not a function of the
noise in the thermocouple readings because those temperature measurements were not used. If
the temperature profile T (t ) was not known, but had to be constructed from the temperature
measurements, then T (t ) could be estimated by filtering the temperature measurements, or by

C A2 (t j )

60

fitting a smooth function to the temperature measurements. The resulting parameter estimates
and confidence region would be a function of the temperature measurements, and the
temperature noise level.
F. Feedback Control Design
Feedback controllers are usually implemented in digital form, which means that the
controller input signals are at discrete time instances and the controller output signals are at
discrete time instances. The time interval between sampling instances is known as the sampling
time. A good rule-of-thumb is to select the sampling time equal to 1/10 to 1/20 of the fastest time
constant of the process that is of interest for control. Some heuristics for the order in which
feedback control loops are closed were summarized in Section B. Many analytical methods are
available for selecting which manipulated variables to pair with which controlled variables
(called the pairing problem), 22,23,24 based on steady-state or dynamic models between the

24

25

process inputs and process outputs. A commonly used method for pairing variables is using the
relative gain array whose elements are defined by:12,13,14
(F1)
ij = Pij H ij
where Pij is the steady-state gain between the i th plant output and j th manipulated variable:
y
Pij = i
u j

(F2)

and
H = ( P 1 )T

(F3)

The pairing rules are:


(1) For ij 0, do not pair the j th manipulated variable with the i th plant output,
(2) ij 1 indicates that it is good to pair the j th manipulated variable with the i th plant
output,
(3) Either ij < 0.6 or ij > 2 indicates that it is bad to pair the j th manipulated variable
with the i th plant output.
If no pairing satisfies all three rules, then redesign the process or implement a full multivariable
controller such as a model predictive controller.
As an example, a model for a fluid catalytic cracking unit has the steady-state gain
25,26
matrix
2.7 0.105
P=
(F4)

1.85 0.085
where the manipulated variables are the flowrates of regenerated catalyst to the reactor and air to
the regenerator, and the controlled variables are the riser exit and regenerator cyclone
temperatures. Then the relative gain matrix is
0.2006 4.3658
0.5416 0.4584
(F5)
H = ( P 1 )T =
=

0.2478 6.3717
0.4584 0.5416
Pairing rule (3) is violated for either pairing, indicating that multivariable control would be
preferred over using multiple single-loop feedback controllers. If two single-loop controllers
were implemented, then it would be better to pair the regenerated catalyst flowrate to the riser
exit temperature and the air flowrate to the regenerator cyclone temperature, since that pairing is
closest to satisfying pairing rule (2).
F.1. Proportional-Integral-Derivative (PID) Control
Proportional-integral-derivative (PID) controllers constitute a significance proportion of
the control systems implemented for most refining and petrochemical facilities. The simplest
type of feedback controllers is the Proportional (P) controller,
u (t ) = u (0) + K c e(t ),
(F6)

where u is the manipulated variable (controller output), the control error e is the difference
between the setpoint yref and measured variable ym, and Kc is the controller gain which is a
tuning parameter. Small values of the controller gain lead to offset between the setpoint and
measured variable and sluggish closed-loop response to disturbances and setpoint changes.
Larger controller gains lead to small offset and fast closed-loop response. For hydrocarbon
25

26

processes, very large controller gains result in large oscillations in the measured variable, or
instabilities. A common tuning approach is to first set the controller gain fairly small, and then
increase its value by factors of 2 until some overshoot is observed in the closed-loop response to
a step change in setpoint.
The offset between the setpoint and measured variable obtained by a P controller is nonzero for most processes, irrespective of controller gain, and a common closed-loop specification
is that the offset resulting from constant changes in the setpoint and disturbances approach zero.
This motivates the use of a Proportional-Integral (PI) controller:

1
u (t ) = u (0) + K c e(t ) +
I

e(t )dt ,

(F7)

where I is an additional tuning parameter known as the integral time constant. It can be shown,
under mild conditions on the process, to produce zero offset to constant changes in the setpoint
and disturbances. Basically, any offset (nonzero e) causes the integral in (F7) to grow, which
increases the manipulated variable u until the offset is removed. For most hydrocarbon processes
too small of an integral time constant results in closed-loop oscillations. A large integral time
constant reduces the effect of the integral action. A popular tuning rule for PI controllers is to set
the integral time constant equal to the slowest time constant of the process, and then vary the
controller gain until a small amount of overshoot is observed in the closed-loop response to a
step change in setpoint. The slowest time constant of the process can be estimated as the time it
takes for the process output to reach 63% of the difference between its initial and final values for
a step change in the manipulated variable. PI controllers are the most common of PID controllers
used in the refining and petrochemical industries.
There are numerous variations of implementation of PID controllers, with a common
time-domain expression for a PID controller being

du
de 1
+ u (t ) = u (0) + K c e(t ) + D
+
dt
dt I

e(t )dt ,
0

(F8)

where D is the derivative time constant and F is a filter time constant. The purpose of the
derivative term is to enable the controller to anticipate changes in the control error e, so that the
control action is larger if the control error e is increasing with time. Usually the derivative time
constant D is set somewhere between 0 and a quarter of the integral time constant I. A common
heuristic tuning rule is to set the derivative time constant to zero unless the best tuning achieved
by a PI controller is too sluggish to satisfy the closed-loop performance specifications, in which
case set D = 0.25I. The purpose of the filter time constant F is to reduce the effect of
measurement noise on controller output, arising from the use of a derivative of the control error
in the controller. Typically the filter time constant is set between 8 and 20 times the derivative
time constant, and for hydrocarbon processes this usually results in controller outputs that do not
have too much stochastic fluctuations resulting from the measurement noise.
A weakness of the PID controller in (F8) is that the derivative of the error signal becomes
very large when there is a step in the setpoint, which results in a large rapid change in the
controller output referred to as derivative kick. 27 Derivative kick can be avoided by replacing the
derivative of the error e = yref ym with the derivative of the measured variable ym to give

dy
du
1
+ u (t ) = u (0) + K c e(t ) D m +
dt
dt I

26

e(t )dt .
0

(F9)

27

By approximating the integral by a summation and the derivatives by a first-order backward


difference, and rearranging, the digital form of the control algorithm is obtained:
ym ,n ym,n 1 t n
K c
F /t

un =
+
un 1 +
en D
+ ek ,
t
1 + F /t 1 + F /t
1 + F /t
I k =1
u0

(F10)

where t = tn tn 1 is the sampling time, un is the value of the manipulated variable which is
held constant between times tn and tn +1 , and ym ,n and en are the measured and error variables at
time tn .
Many methods for tuning PID controllers have been developed over the years, including
Ziegler-Nichols, Cohen Coon, BLT tuning, Internal Model Control, and direct synthesis.12,13,14,27
Some of these tuning rules produce controller parameters that are very sensitive to disturbances
or model uncertainties, whereas there are sets of other tuning rules that give similar closed-loop
performance. A tuning rule applied in the refining and petrochemical industries that allows a
tradeoff between closed-loop performance and robustness to model uncertainties is the Internal
Model Control (IMC) method. Consider a stable process well-approximated by a first-order
model with time delay, which is very common in hydrocarbon processes. Such models have
three parameters: the time delaythe steady-state gain K, and the time constant . The time delay
is the time it takes for a step change in the manipulated variable to begin to affect the process
output, the steady-state gain is the difference between the initial and final values of the process
output divided by the magnitude of this step change, and the time constant is the time it takes for
the process output to reach 63% of the way to its final value. The IMC PID controller parameters
are 28
2 +
Kc =
2 K ( + )
I = + / 2
(F11)

D =
2 +

F =

2( + )

where is a tuning parameter that defines a tradeoff between the closed-loop speed of response
and the robustness of the closed-loop system to measurement noise and inaccuracy in the model.
For larger values of the tuning parameter, is approximately the time constant of the closed-loop
system. Tables of PID tuning rules for other process models are available, 29 including rules
designed to suppress the effects of both output and load disturbances on the closed-loop
performance. 30,31
F.2. Antiwindup, Bumpless Transfer, and Split-range Control
Industrial control systems must account for the discrete process changes such as occur
during the startup and shutdown of any refining or petrochemical facility, hard limits on actuator
movements, switching from manual to automatic control, and switching between control systems
designed for different operating conditions. Providing fast and smooth transitions during discrete
process changes is of high industrial importance to ensure safe operating conditions. Many a
plant explosion occurred during the startup procedure shortly before or after a discrete transition.

27

28

The simplest form of discrete transition occurs when a controller output saturates. Reset
windup is said to occur when the controller continues to integrate the error signal during
saturation, causing large overshoots and oscillations. Discrete process changes also occur during
split-range control, in which different manipulated variables become active in different
operating regimes. Split-range control is useful when more than one manipulated variable is
required to span the whole range of setpoints. Controllers that provide smooth transitions during
discrete process changes are said to provide bumpless transfer.
One method to provide bumpless transfer involves the use of model predictive control,
which is described in Section G. A simple method to provide bumpless transfer that is applied in
single-loop feedback control systems is to implement the velocity form for the controller
(equation (F10) is an example of a controller said to be in position form). To derive the velocity
form for (F10), write (F10) for the n 1 sampling instance and subtract to give:
un = un 1 +

ym ,n 2 ym ,n 1 + ym ,n 2 t
K c
F /t

u
+
e

+ en .

( n1 n2 )
n
n 1
D
1 + F /t
1 + F /t
t
I

(F7)

The main advantage of implementing the controller in this form is that it will not integrate the
controller error when the manipulated variable reaches a constraint (for example, 0 or 100%
Power). For this reason, the controller will also perform better during transitions between
different operating conditions, that is, will provide bumpless transfer.
F.3. Split-range Control Example
Precise control of the temperature T of a sample containing a thin polymer film was
required to provide accurate measurements of diffusion coefficients. 32,33 The manipulated
variable was the power to a heating tape which surrounds the polymer sample. Heat sinks
allowed the temperature of the sample to be reduced quickly. For temperatures below 30 o C , the
heat sink was distilled water. For higher temperatures the heat sink was gaseous nitrogen. The
advantage of the gaseous nitrogen heat sink over the liquid water sink was that it allowed a wide
range of temperatures to be covered by only manipulating the heating power. The distilled water
sink provided a more stable response for temperatures under 30 o C .

The heat sinks were at room temperature, which was ~ 21 o C with slow variations up to
1 o C. For each heat sink, temperature responses to step changes in heating power were taken at
a variety of operating conditions along the desired temperature trajectory in order to estimate the
importance of nonlinearity. The process responses were linear for each heat sink, which were
well modeled as first-order-plus-time-delay with model parameters
K1 = 1.0; 1 = 9.5; 1 = 2.4;
(F8)
for the gaseous nitrogen heat sink ( T > 30 o C ) and
K 2 = 0.068; 2 = 1.7; 2 = 1.4;

(F9)

for the liquid water heat sink ( T < 30 o C ), where the time constants i and time delays i were
in units of minutes and the process gains K i were in units of o C /%Power. The heating power
was constrained between 0 and 100%. At steady state, the sample was at room temperature when
the heating power was turned off.
The goal of the closed-loop system was to smoothly ramp the temperature from stable
operations at 120 oC to 25 oC (see Figure F1a). For reproducible collection of diffusion data, the

28

29

120

Temperature (Centigrade)

Temperature (Centigrade)

temperature had to stay within 0.5 oC of the setpoint 70 minutes before and 50 minutes after the
ramp, and within 1.5 o C throughout the ramp. The control algorithm was required to provide
bumpless transfer between the radically different process behaviors (F13)-(F14) that results
when the temperature crosses 30 oC , while satisfying the constraints on heating power.

100
80
60
40
20
0

200
400
600
Time (minutes)
(a)

100
80
60
40
20
0

Temperature (Centigrade)

% power

80
60
40
20
200
400
600
Time (minutes)
(c)

120

800

100

0
0

140

800

200
400
600
Time (minutes)
(d)

800

1
0.5
0
0.5
1
1.5
0

800

200
400
600
Time (minutes)
(b)

Figure F1. (a) Setpoint; (b) closed-loop temperature tracking (-) and transition line between heat
sinks (--); (c) power output from the nitrogen (--) and water (-) heat sinks; (d) difference between
setpoint and controlled variable.
d
p1

k1

yref

_
k2

p2

ym

Figure F2. Block diagram. The setpoint signal is yref , the error signal is e , the measured
temperature is ym , and the effect of the disturbances on the temperature is d. The selector S
switches between the controllers k1 and k2 depending on the value of the measured temperature.
29

30

The sampling time was selected as t = 0.1 minute which is within 1/10 to 1/20 of the
fastest time constant, 2 = 1.7. This split-range control problem was solved by implementing the
PID controller (F12) in velocity form, using the IMC tuning rules (F11), switching between the
two sets of model parameters (F13) and (F14) depending on whether the temperature was above
or below 30 o C (see Figure 2). The IMC tuning parameter was selected to be = 1.0 minute to
give fast uniform closed-loop response throughout the temperature ramp. The closed-loop
temperature response to programmed step and ramp trajectories is shown in Figure 1bd, with the
controller output shown in Figure 1c. This demonstrates that two digital IMC-based PID
controllers implemented in velocity form that switch during transitions between operating
regimes satisfied all of the closed-loop specifications for this problem.
G. Model Predictive Control
Most advanced controllers being implemented in the refining and petrochemical
industries today are Model Predictive Control (MPC) algorithms, 34,35 of which there are many
variants including RMPC (Honeywells implementation), and DMC (dynamic matrix control).
These algorithms share the common trait of using an explicitly formulated process model to
predict the future and then compute the future control trajectory that optimizes a performance
objective based on the model predictions (see Figure G1). One reason for the popularity of MPC
is its ability to directly account for constraints in the manipulated, state, and controlled
variables. 36 For linear processes the optimization is in the form of a linear program (LP) or
quadratic program (QP), which must be solved at each sampling instance, with the constraints
written directly as constraints in the LP/QP. 37 After the value for the control move at the current
sampling instance is implemented, new measurements are collected and the control calculation is
repeated. These steps update the control move calculations to take into account the latest
measurement information. This section describes a typical modern formulation for a model
predictive controller, followed by a discussion of some technical issues and an example.
past

future
Setpoint

y(k+1)
y
Projected outputs

u(k)
Manipulated variables

k+1 k+2

k+nu

k+np

Control horizon
Prediction horizon

Figure G1. A simplified model predictive control scheme.

30

31

G.1 MPC Formulation


Recall the linear discrete-time state-space model (D12)
xk +1 = Axk + Buk

yk = Cxk

(G1)

where uk, yk, and xk are the vectors of controller outputs, controlled variables, and states at
timestep k and the matrices A, B, and C are assumed to be controllable and observable. The
subscript k + 1 refers to the sampling instance one sampling interval after sampling instance k.
Note that for discrete-time models used in control, there is normally no direct feedthrough term,
that is, the measurement yk does not depend on the input at time k , but it does depend on the
input at time k 1 through the state xk . The reason for the absence of direct feedthrough is that
normally the output is measured at time k before the new input at time k is computed and
implemented.
The state x , input u , and measurement y in (G1) should be interpreted as deviation
variables. This means that they represent the deviations from some consistent set of variables
{xL , uL , yL } around which the model is obtained. To illustrate the notion of deviation variables, if
yL represents a temperature of 330 K, a physical measurement of 331 K would correspond to the
deviation variable y = 1 K. The model (G1) is typically the result of identification experiments
performed around the values {xL , uL , yL } or the result of linearizing and discretizing a nonlinear
first-principles model around the value {xL , uL , yL } . For a stable process, the set {xL , uL , yL } will
typically represent a steady state often the steady state that defines the desired process
operating point.
A typical optimization problem in MPC takes the form
min
ui

i =0,...,n 1

n 1

{( x x
i =0

ref ,i

)T Q( xi xref ,i ) + (ui uref ,i )T R(ui uref ,i )T }

(G2)

+ ( xn xref ,n )T S ( xn xref ,n )

subject to constraints
x0 given
xi +1 = Axi + Bui
yi = Cxi
U L ui UU

for 0 i n 1

YL Hxi YU

for 1 i n + j

(G3)

where i is the time index, n is the control horizon. In the objective function (G2), both the
deviation of the states xi from some desired reference trajectory xref ,i and the deviation of the
inputs ui from some desired trajectory uref ,i are penalized. These reference trajectories, which
may be constant or vary with time, are provided to the model predictive controller by a process
operator or a higher level control system. The constraints on the achievable manipulated
variables or acceptable states are usually not dependent on the reference trajectories, and
31

32

therefore these reference trajectories do not appear in the constraint equations (G3). Usually, the
state constraints represent constraints on the process measurements (that is, H = C), but
constraints on other combinations of states are also possible (including constraints on
combinations of inputs and states). Also, it may be desired to place constraints on the rate of
change of the inputs, giving additional constraints of the form U L ui ui 1 UU . The
output constraints in (G3) are enforced over an extended horizon of length n + j . A sufficiently
large value for j ensures that the constraints will be feasible on an infinite horizon if they are
feasible up to the horizon n + j . 38 For the output constraints in (G3) to be well defined, how the
inputs ui should behave in the interval n i n + j 1 should also be specified. Typical choices
for this time interval are either ui = uref ,i or (ui uref ,i ) = K ( xi xref ,i ) where the matrix K is
designed to stabilize the closed-loop system. For the latter choice, the input constraints should be
included in the problem formulation for the time interval n i n + j 1 . These input constraints
then effectively become an additional set of state constraints over that period.
The symmetric matrices Q , R, and S are design parameters that weight the relative
importance of the three terms in the objective function (G2). The state penalty matrix Q is
positive semidefinite, R is a positive-definite matrix that penalizes the magnitude of the
controller outputs during the control horizon, and S is a positive-definite matrix that penalizes
the magnitude of the states at the end of the control horizon. A good way to specify S is as the
solution to the discrete-time Lyapunov equation:
S = ( A + BK )T S ( A + BK ) + Q

(G4)

where the matrix K is the infinite-horizon linear-quadratic-optimal controller for the weights Q
and R.38 In many applications it is more natural to place a weight (or cost) on the actual
, where Q is the
measurements rather than the states, which can be done by choosing Q = C T QC
weight on the vector of measurements.
It is common to introduce integral action in model predictive controllers by using the
changes in the manipulated variables at time i as free variables in the optimization, rather than
directly using the manipulated variables. Then the actual manipulated variables are computed by
integrating, or summing, the changes in the manipulated variables. This approach can be
implemented within the same framework and model structure as above, by using the model

xk +1
k +1 + B uk
= Ax
u
k

xk +1 =

(G5)

yk = C xk

where uk = uk uk 1 , and

A B
B
(G6)
A =
; B = ; C = [C 0] .

0 I
I
With a lot of algebra, (G2) and (G3) can be written in the standard form for the quadratic
program
+ cTv
(G7)
min v T Hv
v
subject to the constraints

32

33

Lv b
where v is the vector of free variables in the optimization,

(G8)

u uref ,0

n2

n 1

ref ,1

ref , n 2

ref , n 1

u u
#
v=
u u
u

(G8)

H is a positive-definite Hessian matrix, the vector c describes the linear part of the objective
function, and the matrix L and vector b describe the linear constraints. There are many standard
solvers for computing the optimum for (G7).18,19
Many early MPC formulations used an objective function of the form
min
ui

i =0,...,n 1

np

nu

i =0

i =0

( xi xref ,i )T Q( xi xref ,i ) + (ui uref ,i )T R(ui uref ,i ) ,

(G9)

with the prediction horizon being greater than the control horizon, n p > nu , and the typical
assumption being that ui = uref ,i for nu < i < n p . It was also common in industrial practice to use
models based on step response descriptions, in which the discrete values of the process output at
future time instances are reported for steps in each process input. Whereas step response models
have no theoretical advantages, they have the practical advantage of being easier to understand
for engineers with little background in first-principles modeling or control theory. The MPC
formulation can be written directly in terms of a step response model 39 or the step response
model can be expressed in state-space form followed by application of a state-space MPC
formulation such as (G2)-(G3). 40 Note that step-response models have finite memory, and
hence only should be used for asymptotically stable processes, that is, processes where the effect
of past manipulated variable moves vanish over time. Most industrially successful MPC
controllers based on step response models also have been modified to handle integrating
processes (processes whose outputs integrate at least one process input), but not to handle
exponentially unstable processes. Handling unstable processes using step response models would
require more complex modifications to the controllers and model description, and would thereby
remove the step response models advantage of being easy to understand. Partly due to these
reasons, model predictive controllers are seldom directly implemented on unstable processes. If
the underlying process is unstable, it is usually first stabilized by some basic feedback control
loops, with the model predictive controller treating the setpoints of the lower level control loops
as manipulated variables.
The MPC formulation is based on the process model. Good control of the true process is
only obtained if the process model is able to predict the future behavior of the true process with
reasonable accuracy. Model errors and unknown disturbances must always be expected, and
therefore it is necessary to update the process model to maintain good quality predictions of the
future process behavior. The most general way of doing this is through the use of a state
estimator, typically a Kalman filter. 41 The Kalman filter also may be modified to estimate
unmeasured disturbances or model parameters that may vary with time. The Kalman filter is
described in advanced control engineering textbooks, and will not be described further here.

33

34

For asymptotically stable systems, a particularly simple model updating strategy is


possible for MPC formulations that only use process inputs and measurements in the formulation
(that is, do not have unmeasured states in the objective function or constraints). In such cases, it
would be natural to calculate the predicted deviations from the desired output trajectory, rather
than the predicted deviations from the desired state trajectory. Then, the model can be updated
by simply adding the present difference between process output and model output to the models
prediction of the future outputs. This is known as a bias update, and is widespread in industrial
applications.
With MPC it is very simple to include feedforward from measured disturbances, if a
model is available on how the disturbances affect the states/outputs. Recall that feedforward is
used to counteract the future effects of disturbances on the controlled variables (see Section B2).
Thus, feedforward in MPC only requires that the effect on disturbances on the controlled
variables are taken into account when predicting the future state trajectory in the absence of any
control action. Recall from Section B2 that effective feedforward requires both the disturbance
and process model to be accurate.
It is also worth noting that many MPC formulations have been derived that directly take
into account nonlinearities and model uncertainties. 42,43,44,45 Taking model uncertainties into
account is especially important in the context of the control of batch and semibatch processes, in
which the process models are usually based on nonlinear first-principles models, and the target is
a product quality specification at the end of the batch or semibatch run.
G.2. Feasibility, Constraint Handling, and Ill-conditioning
For any type of controller to be acceptable, it must be very reliable. For model predictive
controllers, there is a special type of issue with regard to feasibility of the constraints. An
optimization problem is infeasible if there exists no set of values for the free variables in the
optimization for which all constraints are fulfilled. Problems with infeasibility may occur in
certain MPC formulations, for instance if the operating point is close to a constraint and a large
disturbance occurs. In such cases, it need not be possible to satisfy the constraint at all times.
This also can occur during the startup of model predictive controllers, as the operations may be
far from the desired operating point, and in violation of some constraints. Naturally, it is
important that the model predictive controller should not terminate with providing a solution
when faced with an infeasible optimization problem. Rather, it is desirable that the performance
degradation is predictable and gradual as the constraint violations increase, and that the model
predictive controller should effectively move the process into an operating region where all
constraints are feasible.
Usually the constraints on the manipulated variables result from true physical constraints
that cannot be violated. For example, a valve cannot be more than 100% open. On the other
hand, constraints on the states/outputs often represent operational desirables rather than
fundamental operational constraints. State/output constraints often may be violated for short
periods of time (although possibly at the cost of producing off-spec products or increasing the
need for maintenance). It is therefore common to modify the MPC optimization problem in such
a way that output constraints may be violated if necessary. There are (at least) three approaches
to doing this modification:
1. Remove the state/output constraints for a time interval in the near future. This is simple,
but may allow for unnecessarily large constraint violations. Furthermore, it need not be
simple to determine for how long a time interval the state/output constraints need to be
removed this may depend on the operating point, the input constraints, and the assumed
maximum magnitude of the disturbances.

34

35

2. To solve a separate optimization problem prior to the main optimization in the MPC
calculations. This initial optimization minimizes some measure of how much the
output/state constraints need to be moved in order to produce a feasible optimization
problem. The initial optimization problem is usually a LP problem, which can be solved
very efficiently.
3. Introducing penalty functions in the optimization problem. This involves modifying the
constraints by introducing additional variables such that the constraints are always
feasible for sufficiently large values for the additional variables. Such modified
constraints are termed soft constraints. At the same time, the objective function is
modified, by introducing a term that penalizes the magnitude of the constraint violations.
The additional variables introduced to ensure feasibility of the constraints then become
additional free variables in the optimization. Thus, feasibility is ensured by increasing the
size of the optimization problem.
The two latter approaches are both rigorous ways of handling the feasibility problem. Approach
3 has a lot of flexibility in the design of the penalty function. The implementation can be
designed to ensure that the constraints are violated according to a strict list of priorities, that is, a
given constraint only will be violated when it is impossible to obtain feasibility by increasing the
constraint violations for less important constraints. Alternatively, the constraint violations can be
distributed among several constraints. Although several different penalty functions may be used,
depending on how the magnitudes of the constraint violations are measured, two properties are
desirable:
That the QP problem in the optimization problem can still be solved efficiently. This
implies that the Hessian matrix for the modified problem should be positive definite, that
is, there should be some cost on the square of the magnitude of the constraint violations.
That the penalty functions are exact, which means that no constraint violations are
allowed if the original problem is feasible. This is usually obtained by putting a
sufficiently large weight on the magnitude of the constraint violations (i.e., a linear term)
in the objective function.
The use of penalty functions is described in standard textbooks on optimization, 46 and is
discussed in the context of MPC in many papers. 47,48,49
Another issue of concern is that the model predictive controller, or any multivariable
controller for that matter, can provide poor closed-loop performance if the matrices in the model
are ill-conditioned, that is, have a high condition number. Well-implemented MPC algorithms
for large-scale systems compute the condition number of suitably scaled matrices and revise the
matrices to remove the effect of ill-conditioning on the computed manipulated variables. 50,51 For
example, Honeywells Robust Multivariable Predictive Control Technology computes an
optimally diagonally scaled condition number 52 as an intermediate step to assess illconditioning. 53
G.3 Example
Consider a process model with the discrete-time state-space matrices (reported to 4
significant figures):

35

36

0.9280 0.0024 0.0031 0.0040


0.0415 0.9538 0.0119 0.0065

A=
0.0521 0.0464 0.8957 0.0035

0.0686 0.0508 0.0318 0.9346

(G10)

0.0000 0.3355
0.1826 0.0074

B=
0.0902 0.0093

0.0418 0.0123

(G11)

0.0000 0.0000 0.0981 0.2687


C=
(G12)

0.0000 0.0000 0.0805 0.3271


which has a sampling interval of 0.5 minutes. The open-loop step responses are shown in Figure
G2.
Step Response
From: U(2)

From: U(1)
1

To: Y(1)

0.5

Amplitude

0.5

To: Y(2)

0.5
0
0.5
1
1.5
2
0

20

40

60

80

100

20

40

60

Sample

Figure G2. Open loop unit step responses for (G10)-(G12).

36

80

100

120

37

The objective of the controller is to drive the controlled variable y to zero given the initial
condition
0
0

x0 =
(G13)
20

1
which violates the output constraint at t = 0 , and the manipulated and output variables are
constrained between 1 . To ensure feasibility, the output constraints were addressed by
penalizing output constraint violations in the MPC formulation:
n 1

min

uk ,uk +1 ,",uk + N 1

i =0

xkT+i Qxk +i + ukT+i Ruk +i + xkT+ N Sxk + N + T P + pT

(G14)

subject to
uL uk +i uU ,
1
Cxk +i i U yU ,
r
1
Cxk +i + i L yL ,
r
U 0 , L 0 ,

where selecting r = 2.5 results in the penalty on the output constraints increasing rapidly into the
future. The model predictive controller tuning parameters were selected as Q = C T C , R = I 2 (the
identity matrix), P = 0.01 , p = 100 , and n = 20 , with S computed from (G4). An excess
constraint horizon j = 70 was shown in simulations to be sufficiently large to guarantee
feasibility of the constraints on the infinite horizon. Fairly fast satisfaction of the constraints on
the controlled variables is seen in the closed-loop responses (see Figure G3).
2.5
1

0.5

1
0.5

y1 (actual)
y1 (predicted)
y2 (actual)
y2 (predicted)

0
0.5
1
1.5

inputs

outputs

1.5

10

20
30
sample

0
u1 (actual)
u1 (predicted)
u2 (actual)
u (predicted)

0.5
1
40

50

10

20
30
sample

Figure G3. Controlled and manipulated variables for MPC example.


37

40

50

38

H. Process Monitoring
In the refining and petrochemical industries, there has been a large push to produce
higher quality products, to reduce product rejection rates, and to satisfy increasingly stringent
safety and environmental regulations. To meet these standards, modern industrial processes
contain a large number of variables operating under closed-loop control. The standard process
controllers (PID controllers, model predictive controllers, etc.) are designed to maintain
satisfactory operations by compensating for the effects of disturbances and changes occurring in
the process. While these controllers can compensate for many types of disturbances, there are
changes in the process, known as faults, which the controllers cannot handle adequately. More
precisely, a fault is an unpermitted deviation of at least one characteristic property or variable of
the system. Examples of faults include stuck valves, poisoned catalyst in a reactor, fouled heat
exchangers, and biased sensors.
To ensure that the process operations satisfy the performance specifications, the faults in
the process need to be detected, diagnosed, and removed. These tasks are associated with
process monitoring, also known as statistical process control. The goal of process monitoring is
to ensure the success of the planned operations by recognizing anomalies of the behavior. The
information not only keeps the plant operator and maintenance personnel better informed of the
status of the process, but also assists them to make appropriate remedial actions to remove the
abnormal behavior from the process. As a result of a well designed process monitoring system,
downtime is minimized, safety of plant operations is improved, and manufacturing costs are
reduced.
Process monitoring is typically organized into four steps of a process monitoring loop:
fault detection, fault identification, fault diagnosis, and process recovery (see Figure H1). 54,55
Fault detection is determining whether a fault has occurred. Early detection may provide
invaluable warning on emerging problems, with appropriate actions taken to avoid serious
process upsets. Fault identification is identifying the observation variables most relevant to
diagnosing the fault. The purpose of this procedure is to focus the plant operators and engineers
attention on the subsystems most pertinent to the diagnosis of the fault, so that the effect of the
fault can be eliminated in a more efficient manner. Fault diagnosis is determining which fault
occurred, in other words, determining the root cause of the observed out-of-control status.
Process recovery is counteracting or eliminating the fault.
This section covers control charts, since these are the most prevalent form of process
monitoring practiced in the refining and petrochemical industry. Control charts are based on
statistical theory which relies on the assumption that the characteristics of the data variations are
relatively unchanged unless a fault occurs in the system. This implies that the properties of the
data variations, such as the mean and variance, are repeatable for the same operating conditions,
although the actual values of the data may not be very predictable. The repeatability of the
statistical properties allows thresholds for certain measures, effectively defining the out-ofcontrol status, to be determined. Violating these thresholds indicates that a fault has been
detected, and tracking which variables violate the thresholds provides the simplest approach for
fault identification. This information can be used to assist the process operator or process
engineer in diagnosing the fault, and using engineering judgment to determine the most effective
approach to process recovery.

38

39

No

Fault
Detection

Yes

Fault
Identification

Fault
Diagnosis

Process
Recovery

Figure H1. The process monitoring loop.


Below is described the pretreatment of historical data before using the data to design
control charts and the design of univariate and multivariate control charts including the
determination of statistically-derived thresholds.
H.1 Data Pretreatment
To extract the information in the data relevant to process monitoring effectively, it is
often necessary to pretreat the data in the training set, which contains historical data available
for analysis prior to the on-line implementation of the process monitoring scheme that are used
to determine the thresholds representing the in-control and out-of-control operations. Most fault
detection methods require that the training set contain data only collected during process
operations in which no faults occur, that is, during normal operating conditions.
The training set should not contain any data that have no information relevant to
monitoring the process. For instance, it may be known a priori that the output from a certain
sensor regularly exhibits spikes or may slowly drift with time irrespective of the process
behavior, or some of the sensor readings may be physically separate from the portion of the
process that is being monitored. In these instances, the number of false alarms can be reduced
and fault identification improved by removing such signals from the training set.
Since the training data are intended to define normal operating conditions, it is important
that all data that arose during faulty operations be removed. Obvious outliers can be removed by
plotting and visually inspecting the data for outlying points. A common method to remove
outliers is to calculate the standard deviation for all of the historical data for each variable and
remove data points that are farther than 4 standard deviations from the mean. To detect outliers
in data while taking correlations between variables into account, first stack the training data,
consisting of m measurement variables and n observations for each variable, into a matrix
x11
x
X = 21
#

xn1

x1m
x22
x2 m
.
(H1)
#
#

xn 2 " xnm
Then the sample covariance matrix of the training data is defined by
1
(H2)
S =
X TX.
n 1
Assuming that the sample covariance matrix is invertible, then with 100(1)% confidence a
vector of measurement variables x is an outlier if
x12

"
"

39

40

(n 1) 2 (m/ (n m 1)) F (m, n m 1)


1

x S x>
n(1 + (m/ (n m 1)) F (m, n m 1))
T

(H3)

where F (m, n m 1) is the upper 100 % critical point of the F-distribution with m and
n m 1 degrees of freedom (this method cannot be applied when S is not invertible). 56
H.2 Univariate Control Charts
The most common control chart is known as the Shewhart chart (see Figure H2). 57 The
values of the upper and lower control limits on the Shewhart chart are determined so as to
minimize the rate of false alarms and the rate of missed detections (faults that occur but not
detected). There is always an inherent tradeoff between minimizing the false alarm and missed
detection rates. Tight threshold limits for an observation variable result in a high false alarm and
low missed detection rate, while limits which are too spread apart result in a low false alarm and
a high missed detection rate.
Given certain threshold values, statistical hypothesis theory can be applied to predict the
false alarm and missed detection rates based on the statistics of the data in the training sets. Let
represent normal operating conditions and i represent operations in which a specific fault, i,
occurred. The null hypothesis is that a particular measurement x corresponds to normal operating
condition and the alternative hypothesis is that the measurement x corresponds to the fault i.
Then false alarm rate is equal to the type I error, and the missed detection rate for fault i is equal
to the type II error, as defined in textbooks on statistics.57 This is illustrated graphically in Figure
H3.
52
Upper Control Limit

+3

51.5

51

50.5

50

49.5

49

48

Lower Control Limit

48.5

10

15

20

Time (hr)

Figure H2. Shewhart chart for some Mooney viscosity data.12

40

25

41

p(x | )

p(x | i)

Threshold

Type II Error

Type I Error

Figure H3. The type I and type II error regions for the null hypothesis (assign x as ) and the
alternative hypothesis (assign x as i ). The probability density function for x conditioned on
is p( x | ) ; the probability density function for x conditioned on i is p( x | i ) . The
probability of a type I error is and the probability of a type II error is . Using Bayesian
decision theory, these notions can be generalized to include a priori probabilities of and i . 58
Increasing the threshold (shifting the vertical line to the right in Figure H3) decreases the
false alarm rate while increasing the missed detection rate, and decreasing the threshold
decreases the missed detection rate while increasing the false alarm rate. This inherent tradeoff
cannot be avoided, although it can be improved by basing decisions on more data points,
installing better sensors, and modifying the process to reduce the normal process variability. The
value of the type I error, also called the level of significance , specifies the degree of tradeoff
between the false alarm rate and the missed detection rate. As a specific example, assume for the
null hypothesis that any deviations of the process variable x from a desired value are due to
inherent measurement and process variability described by a normal distribution with standard
deviation :
p ( x) =

1
(x )
exp
.

2 2
2

(H4)

The alternative hypothesis is that x . Assuming that the null hypothesis is true, the
probabilities that x is in certain intervals are
Pr{x < c / 2 } = / 2
(H5)
Pr{x > c / 2 } = / 2

(H6)

Pr{ c / 2 x + c / 2 } = 1

(H7)

where c / 2 is the standard normal deviate corresponding to the (1 / 2) percentile, reported in


Table H1. These values are calculated using the cumulative standard normal distribution.

Table H1. Some commonly used values for the standard normal deviate. 59 These values are
calculated using the cumulative standard normal distribution.

41

42

c / 2

0.0027
0.005
0.01
0.02
0.05

3.00
2.81
2.58
2.33
1.96

The lower and upper thresholds for the process variable x are c/ 2 and + c/ 2 ,
respectively. Figure H2 shows the application of Shewhart chart to monitor the Mooney
viscosity of an industrial elastomer.12 The mean value was 50.0, with a standard deviation
value of = 0.5 known to characterize the intrinsic variability associated with the sampling
procedure. Since all the data points fall inside the upper and lower control limit lines
corresponding to c/ 2 = 3.0 , the process is said to be in control.
As long as the mean and standard deviation of the training set accurately represent the
true statistics of the process, the thresholds in (H4) should result in a false alarm rate equal to
when applied to on-line data. If 20,000 data points were collected during in control operation
defined by c/ 2 = 3.0 , 27 data points would be expected to fall above the upper control limit,
while 27 data points would be expected to fall below the lower control limit. Some typical
values for fault detection are 0.005 , 0.01 , and 0.05 . It has been suggested that even if x does
not follow a normal distribution, the thresholds in (H4) are effective as long as the data in the
training set are an accurate representation of the variations during normal operations. 60
When Shewhart charts do not provide adequate false alarm and missed detection rates,
these rates can be improved by employing measures that incorporate observations from multiple
consecutive time instances, such as the cumulative sum (CUSUM) and exponentially-weighted
moving average (EWMA) charts.12,57 For a given false alarm rate, these methods can increase the
sensitivity to faults and decrease the missed detection rate, but at the expense of increasing the
detection delay, which is the amount of time expended between the start of the fault and time to
detection. This suggests that the CUSUM and EWMA charts are better suited for faults
producing small persistent process shifts, with the Shewhart charts being better for detecting
faults producing sudden large process shifts.
Univariate statistical charts (Shewhart, CUSUM, and EWMA) determine the thresholds
for each measurement individually without considering the information contained in the other
variables. These methods do not take into account correlations between measurements, which
tend to be very strong in modern refining and petrochemical facilities, due to heat integration,
mass recycle loops, etc. Improved process monitoring can be achieved using multivariate control
charts, which take these correlations into account.
G.3 Multivariate Control Charts

The simplest multivariate control chart is based on Hotellings T 2 statistic 61

T 2 = xT S 1 x

(H8)
2

where S is the covariance matrix of the measurement vector x, S = cov(x). The T statistic is a
scaled measure of the distance between the measurement vector x from its mean. Appropriate
thresholds for the T 2 statistic based on the level of significance, , can be determined by
assuming the observations are randomly sampled from a multivariate normal distribution. If it is
42

43

assumed additionally that the sample mean vector and covariance matrix for normal operations
are equal to the actual mean vector and covariance matrix, respectively, then the T 2 statistic
belongs to the 2 distribution with m degrees of freedom56
T2 = 2 (m) .

(H9)

The set T 2 T2 is an elliptical confidence region in the measurement space, as illustrated in


Figure PM4 for two measurement variables (m = 2), with data points in the interior of the
ellipsoid classified as being in control and data points outside of the ellipsoid classified as
being out of control. For a given level of significance , Figure PM4 shows the difference
between applying the multivariable control chart and applying two univariate control charts to
each measurement variable. By taking the correlations between the two measurements into
account, the multivariate control chart has a much lower false alarm rate with a somewhat
improved missed detection rate, for the same level of significance . The stronger the correlation
between measurement variables, the more elongated the ellipsoid, and the larger difference in the
process monitoring performance between the multivariate control chart and the individual
univariate control charts.
x2

y1

T2 Statistical
Confidence
Region

y2
x1
Univariate
Statistical
Confidence
Region

Figure H4. A comparison of the in-control status regions using T 2 statistic (H9) and the
univariate statistics (H5) and (H6) for two measurements.
When the actual covariance matrix for the in-control status is not known but instead
estimated from the sample covariance matrix S computed from (H2), then the threshold on the
statistic is56
m(n 1)(n + 1)
(H10)
T2 =
F (m, n m)
n( n m)
For a given level of significance, the upper in-control limit in (H10) is larger (more conservative)
than the limit in (H9), and the two limits approach each other as the amount of data increases
( n ). 62
The thresholds in (H9) and (H10) assume that the measurements at one time instance are
statistically independent from measurements at other time instances. If there are enough data in
the training set to capture the normal process variations, the T 2 statistic can be an effective tool
for process monitoring even if there are mild deviations from the normality or statistical
independence assumptions.6,60 More sophisticated control charts55 can provide improved process

43

44

monitoring when the data are highly temporally correlated, which occurs when data are collected
at short sampling intervals, due to process dynamics.
There are several extensions that are not often applied in industry, but for which there are
rigorous statistical formulations. For example, lower control limits can be derived62 for T 2
which can detect shifts in the covariance matrix (although the upper control limit is usually used
to detect shifts in mean, it also can detect changes in the covariance matrix). 63 The CUSUM and
EWMA control charts can be generalized to the multivariable case in a similar manner as the
Shewhart control chart was generalized to the multivariable case using the T 2 statistic.60,64,65,66,
67,68
As in the scalar case, the multivariable CUSUM and EWMA charts can detect small
persistent changes more readily than the multivariable Shewhart chart, but with increased
detection delay. Multivariate statistical methods also have been generalized to fault isolation and
diagnosis.55,69
H.4. Data Requirements
The quality and quantity of the data in the training set have a large influence on the
effectiveness of the T 2 statistic as a process monitoring tool. The process monitoring
performance depends on having enough data to statistically populate the covariance matrix for
m observation variables. One way of assessing this is in terms of the amount of data needed to
produce a threshold value sufficiently close to the threshold obtained by assuming infinite data in
the training set.
For a given level of significance , a threshold based on infinite observations in the
training set, or equivalently an exactly known covariance matrix, can be computed using (H9),
and the threshold for n observations in the training set is calculated using (H10). A large relative
difference between these two threshold values,
m(n 1)(n + 1)
F (m, n m) 2 (m)
n( n m)
(H11)
=
2 (m)

implies that more data should be collected. The data requirements for various quantities of data
are reported in Table H2.55 The table indicates that the required number of measurement vectors
is approximately an order-of-magnitude larger than the number of measurements. The data
requirements given in Table H2 do not take into account sensitivities that occur when some
eigenvalues of the covariance matrix are small, in which case the accuracy of estimates of those
eigenvalues will be poor, which can give erratic values for T 2 for some measurement vectors x.
Such ill-conditioning is resolved by dimensionality reduction methods for process monitoring,
such as principal component regression, partial least squares (also known as projection to latent
structures), canonical variate analysis, and Fisher discriminant analysis.55,70
Acknowledgements
Contributions from Evan L. Russell, Leo H. Chiang, and Serena H. Chung are acknowledged.

Table H2. The quantity of data required for various numbers of measurements, m, where
= 0.10 and = 0.5 , which implies that the medians of the T 2 statistic using (H9) and (H10)
differ by less than 10%.55
number of measurements, m
1

number of measurement vectors, n


19
44

45

2
3
4
5
10
25
50
100
200

30
41
52
63
118
284
559
1110
2210

References
1

Page S. Buckley. Techniques of Process Control. Wiley, New York, 1964.


W. L. Luyben and M. L. Luyben. Essentials of Process Control. McGraw-Hill, New York,
1997.
3
M. A. Henson and D. E. Seborg, editors. Nonlinear Process Control. Prentice Hall PTR, Upper
Saddle River, New Jersey, U.S.A., 1997.
4
M. Hovd, D. L. Ma, and R. D. Braatz. On the computation of disturbance rejection measures.
Ind. Eng. Chem. Res., 42:2183-2188, 2003.
5
J. G. Balchen and K. I. Mumm. Process Control. Structures and Applications. Blackie
Academic & Professional, Glasgow, Scotland, 1988.
6
G. E. P. Box, W. G. Hunter, and J. S. Hunter. Statistics for Experimenters - An Introduction to
Design, Data Analysis, and Model Building. Wiley, New York, 1978.
7
B. H. Gunter. How statistical design concepts can improve experimentation in the physical
sciences. Computers in Physics, 7:262-272, 1993.
8
E. L. Crow, F. A. Davis, and M. W. Maxfield. Statistics Manual. Dover Publications, Inc., New
York, 1960.
9
J. M. Juran and F. M. Gryna, editors. Jurans Quality Control Handbook. McGraw-Hill, New
York, 4th edition, 1988.
10
H. Martens and T. Naes. Multivariate Calibration. Wiley, Chichester, 1989.
11
J. E. Jackson. A Users Guide to Principal Components. Wiley, New York, 1991.
12
B. A. Ogunnaike and W. H. Ray. Process Dynamics, Modeling, and Control. Oxford
University Press, New York, 1994.
13
G. Stephanopoulos. Chemical Process Control - An Introduction to Theory and
Practice. Prentice Hall, Englewood Cliffs, New Jersey, 1990.
14
B. W. Bequette. Process Control - Modeling, Design, and Simulation. Prentice Hall PTR,
Englewood Cliffs, New Jersey, 2003.
15
H. S. Fogler. Elements of Chemical Reaction Engineering. Prentice-Hall PTR, Englewood
Cliffs, New Jersey, 3rd edition, 1997.
16
R. B. Bird, W. E. Stewart, and E. N. Lightfoot. Transport Phenomena. Wiley, New York,
1960.
17
J. V. Beck and K. J. Arnold. Parameter Estimation in Engineering and Science. Wiley, New
York, 1977.
18
Matlab. The Mathworks, Inc., Natick, Massachusetts, computer software, 2005.
19
IMSL. Visual Numerics, Inc., computer software, 1997.
20
J. L. Zhou, A. L. Tits, and C. T. Lawrence. FFSQP. University of Maryland, College Park,
computer software, 1989.
2

45

46

21

M. Caracotsios and W. E. Stewart. Sensitivity analysis of initial value problems with mixed
ODEs and algebraic equations. Comp. & Chem. Eng., 9:359-365, 1985.
22
T. J. McAvoy and R. D. Braatz. Controllability limitations for processes with large singular
values. Ind. Eng. Chem. Res., 42:6155-6165, 2003.
23
R. D. Braatz, J. H. Lee, and M. Morari. Screening plant designs and control structures for
uncertain systems. Comp. & Chem. Eng., 20:463-468, 1996.
24
S. Skogestad and I. Postlethwaite. Multivariable Feedback Control: Analysis and Design.
Wiley, Chichester, United Kingdom, 1996.
25
J. Alvarez-Ramirez, J. Valencia, and H. Puebla. Multivariable control configurations for
composition regulation in a fluid catalytic cracking unit. Chem. Eng. J., 99:187-201, 2004.
26
M. Hovd and S. Skogestad. Procedure for regulatory control-structure selection with
application to the FCC process. AIChE J., 39:1938-1953, 1993.
27
D. E. Seborg, T. F. Edgar, and D. A. Mellichamp. Process Dynamics and Control. John Wiley,
New York, 1989.
28
D. E. Rivera, S. Skogestad, and M. Morari. Internal model control 4: PID controller design.
Ind. Eng. Chem. Proc. Des. Dev., 25:252-265, 1986.
29
M. Morari and E. Zafiriou. Robust Process Control. Prentice-Hall, Englewood Cliffs, New
Jersey, 1989.
30
I. G. Horn, J. R. Arulandu, C. J. Gombas, J. G. VanAntwerp, and R. D. Braatz. Improved filter
design in internal model control. Ind. Eng. Chem. Res., 35:3437-3441, 1996.
31
R. D. Braatz. Internal model control. In W. S. Levine, editor, The Control Handbook, CRC
Press, Boca Raton, Florida, pages 215-224, 1995.
32
P. A. Drake. Surface-Enhanced Raman and Surface Plasmon Resonance Measurements of
Case II Diffusion Events on the Nanometer Length Scale. PhD thesis, University of Illinois,
Urbana, Illinois, 1995.
33
S. H. Chung and R. D. Braatz. Teaching antiwindup, bumpless transfer, and split-range
control. Chem. Eng. Edu., 32:220-223, 1998.
34
C. E. Garcia, D. M. Prett, and M. Morari. Model predictive control: Theory and practice - A
survey. Automatica, 25:335-348, 1989.
35
J. B. Rawlings. Tutorial: Model predictive control technology. Proceedings of the American
Control Conference, IEEE Press, Piscataway, New Jersey, pages 662-676, 1999.
36
Methods to quantify the potential closed-loop performance degradation caused by constraints
are described by D. L. Ma, J. G. VanAntwerp, M. Hovd, and R. D. Braatz. Quantifying the
potential benefits of constrained control for a large scale system. IEE Proceedings - Control
Theory and Applications, 149:423-432, 2002.
37
Methods to reduce the computational expense of MPC algorithms are reviewed by J. G.
VanAntwerp and R. D. Braatz. Model predictive control of large scale processes. J. of
Process Control, 10:1-8, 2000.
38
J. B. Rawlings and K. R. Muske. The stability of constrained receding horizon control. IEEE
Trans. on Automatic Control, 38:1512-1516, 1993.
39
C. E. Garcia and A. M. Morshedi. Quadratic programming solution of dynamic matrix control
(QDMC). Chem. Eng. Comm., 46:73-87, 1986.
40
M. Hovd, J. H. Lee, and M. Morari. Truncated step response models for model predictive
control. J. of Process Control, 3:67-73, 1993.
41
B. D. O. Anderson and J. B. Moore. Optimal Filtering. Prentice Hall, Upper Saddle River,
New Jersey, 1979.

46

47

42

Z. K. Nagy and R. D. Braatz. Robust nonlinear model predictive control of batch processes.
AIChE J., 49:1776-1786, 2003, and citations therein.
43
M. A. Henson. Nonlinear model predictive control: Current status and future directions.
Comp. & Chem. Eng., 23:187-202, 1998.
44
F. Allgower, R. Findeisen, and Z. K. Nagy. Nonlinear model predictive control: From theory
to application. J. of the Chinese Institute of Chemical Engineers, 35:299-315, 2004.
45
D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert. Constrained model predictive
control: Stability and optimality. Automatica, 36:789-814, 2000.
46
R. Fletcher. Practical Methods of Optimization. 2nd edition, John Wiley & Sons, New York,
1987.
47
N. M. C. de Oliveira and L. T. Biegler. Constraints handling and stability properties of modelpredictive control. AIChE J., 40:1138-1155, 1994.
48
P. O. Scokaert and J. B. Rawlings. Feasibility issues in linear model predictive control. AIChE
J., 45:1649-1659, 1999.
49
M. Hovd and R. D. Braatz. Handling state and output constraints in MPC using timedependent weights. Modeling, Identification, and Control, 25:67-84, 2004.
50
J. G. VanAntwerp and R. D. Braatz. Model predictive control of large scale processes. J. of
Process Control, 10:1-8, 2000.
51
R. D. Braatz and J. G. VanAntwerp. Fast Model Predictive Ellipsoid Control Process, U.S.
Patent #6,064,809, May 16, 2000.
52
R. D. Braatz and M. Morari. Minimizing the Euclidean condition number. SIAM J. on Control
and Optimization, 32:1763-1768, 1994.
53
RMPCT: A new robust approach to multivariable predictive control for the process industries.
Preprints of the Control Systems 96 Conference, SPCI, Halifax, Nova Scotia, Canada, pages
53-60, 1996.
54
A. C. Raich and A. Cinar. Statistical process monitoring and disturbance diagnosis in
multivariable continuous processes. AIChE J., 42:995-1009, 1996.
55
L. H. Chiang, E. L. Russell, and R. D. Braatz. Fault Detection and Diagnosis in Industrial
Systems. Springer-Verlag, London, 2001.
56
J. F. MacGregor and T. Kourti. Statistical process control of multivariate processes. Control
Eng. Practice, 3:403-414, 1995.
57
D. C. Montgomery. Introduction to Statistical Quality Control. John Wiley & Sons, New
York, 1985.
58
R. O. Duda and P. E. Hart. Pattern Classification and Scene Analysis. John Wiley & Sons,
New York, 1973.
59
W. W. Hines and D. C. Montgomery. Probability and Statistics in Engineering and
Management Science. John Wiley & Sons, New York, 3rd edition, 1990.
60
T. Kourti and J. F. MacGregor. Multivariate SPC methods for process and product monitoring.
J. of Quality Technology, 28:409-428, 1996.
61
J. E. Jackson. Quality control methods for several related variables. Technometrics, 1:359-377,
1959.
62
N. D. Tracy, J. C. Young, and R. L. Mason. Multivariate control charts for individual
observations. J. of Quality Control, 24:88-95, 1992.
63
D. M. Hawkins. Multivariate quality control based on regression-adjusted variables.
Technometrics, 33:61-67, 1991.
64
C. A. Lowry, W. H. Woodall, C. W. Champ, and S. E. Rigdon. A multivariate exponentially
weighted moving average control chart. Technometrics, 34:46-53, 1992.
47

48

65

R. S. Sparks. Quality control with multivariate data. The Australian J. of Statistics, 34:375390, 1992.
66
S. J. Wierda. Multivariate statistical process control - Recent results and directions for future
research. Statistica Neerlandica, 48:147-168, 1994.
67
J. J. Pignatiello and G. C. Runger. Comparisons of multivariate CUSUM charts. J. of Quality
Technology, 22:173-186, 1990.
68
S. S. Prabhu and G. C. Runger. Designing a multivariate EWMA control chart. J. of Quality
Technology, 29:8-15, 1997.
69
L. H. Chiang, E. L. Russell, and R. D. Braatz. Fault diagnosis in chemical processes using
Fisher discriminant analysis, discriminant partial least squares, and principal component
analysis. Chemometrics & Intelligent Laboratory Systems, 50:243-252, 2000.
70
E. L. Russell, L. H. Chiang, and R. D. Braatz. Fault detection in industrial processes using
canonical variate analysis and dynamic principal component analysis. Chemometrics &
Intelligent Laboratory Systems, 51:81-93, 2000.

48

Das könnte Ihnen auch gefallen