0% found this document useful (0 votes)
153 views96 pages

Process Automation (Anewcerefresher) A 14-Part Series

This document provides an overview of basic concepts, terminology, and techniques for process control, including: 1) It defines key terms like controlled variables, manipulated variables, and load variables that are important for understanding process control. 2) It describes feedback control systems, which are more common than feedforward systems, and how a basic feedback loop works to control a process variable through comparison of measured and desired values. 3) It explains that the goal of process control is to balance the manipulated and load variables to maintain the controlled variable at its desired setpoint through a feedback controller's trial-and-error adjustments.

Uploaded by

Jared Tapia
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
153 views96 pages

Process Automation (Anewcerefresher) A 14-Part Series

This document provides an overview of basic concepts, terminology, and techniques for process control, including: 1) It defines key terms like controlled variables, manipulated variables, and load variables that are important for understanding process control. 2) It describes feedback control systems, which are more common than feedforward systems, and how a basic feedback loop works to control a process variable through comparison of measured and desired values. 3) It explains that the goal of process control is to balance the manipulated and load variables to maintain the controlled variable at its desired setpoint through a feedback controller's trial-and-error adjustments.

Uploaded by

Jared Tapia
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 96

Process Automation

A 14-Part Series
CONTENTS
Page

Basic concepts, terminology, and techniques


for process control ....................................... 3

Feedback control modes ...................................... 12

Tuning process controllers ..................................... 19

Feedback methods for process control systems ........................ 25


Direct synthesis and adaptive controls ............................. 31
How to use feedback loops to meet process
conditions............................................
36

Scaling converts process signals to


instrulnent ones ......................................... 43

Advanced control methods ..................................... 49

Feedforward methods for process


57
control systelns .........................................

Advanced control techniques for


disth]ation columns ...................................... 63

Microprocessor controllers ..................................... 70

Sequential control .......................................... 78

Fault-tolerant processes ...................................... 84

Optimizing process and economic aims ............................. 90

Reprinted by special permission from CHEMICAL ENGINEERING, May 30,


August 8, September 19, and November 14, 1983; February 6, April 30,
June 25, August 20, October 15, December 10, 1984; and February 4, April 29,
September 16, and December 9/23, 1985.
Copyright © 1983, 1984, 1985 by MCGraw-Hill, Inc., New York, NY 10020.
A new CE REFRESHER starts in this issue on
instrumentation and techniques applicable
to the control of processes. The subjects to
be covered are:
I Basic concepts
I Basic control modes
I Tuning process controllers
I Techniques of feedback control
I Combining feedback control loops
I Instrument scaling
I Advanced colitrol techniques
I Advanced Control
I Microprocessor regulatory control
I Process control computers

Basicconcepts,terminologyand
techniquesforprocesscontrol
How the interplay among the measured,load and control
variables is established in order to achieve required
objectives for controlling process operations.

Lewis M. Cordon, The Foxboro Co.

I Any study of process control must begin by investigat- In this example, the control system manipulates the
ing the concept of a ``process." From a production position of a steam valve. However, the temperature of
viewpoint, it is generally thought of as a place where the water depends not only on the position of this valve
materials and, most often, energy come together to but also on the flowrate of the' water, its inlet tempera-
ture, the enthalpy of the steam, the degree of fouling in
produce a desired product. From a control viewpoint,
the meaning is more specific. A process is identified as the exchanger, and the ambient temperature.
having one or more variables associated with it that are This simple example illustrates controlled, manipu-
important enough for their values to be known and for lated and load variables-the three categories associated
them to be controlled. with every process under control (Fig. 1b). The parame-
Initially, in this new CE REFRESHER (see accompany- ters that indicat,e product quality or the operating condi-
ing box for series topics), we will concentrate on proc- tion of the process are called controlled variables, such as
esses having only one controlled variable, such as the pressure, level, temperature, pH, specific gravity or den-
heat-exchange process shown in Fig. Ia. To maintain the sity, composition, moisture cont,ent, weight and speed,
temperature of the product (hot water) in this process, and other variables, depending on the process.
another variable influencing the variable being con- Manipulated variables include valve position, damper
trolled must be available for manipulation by the control position, motor speed and blade pitch. Further, one
System. control loop is often manipulated for controlling another
variable in more complicated control schemes. For exam- signal based on the difference between the actual and
ple, a flow variable is manipulated to control a tempera- reference-measurement values. For /ccd/o7'tt;ord systems,
ture or a level. the control signal is generated from values based on the
All variables affecting a controlled variable, other than various load variables as they affect the process.
the one being manipulated, are defined as loads. Both
loads and the manipulated variable may influence a Feedback systems
controlled variable from either the supply side or the Feedback systems are more common than feed for-
demand side of the process. For example, the outlet ward ones. The structure of a feedback loop is shown in
temperature of a heat exchanger can be controlled by Fig. 2. Here, the value of the controlled variable re-
manipulating the steam valve, while tank level can be sponds to the net effect of the loads and the manipulated
controlled by manipulating a valve on the outflow from variable. A sensor/transmitter measures the current
the tank. Often, a controlled variable in one process is a value of the controlled variable and sends a signal to the
load variable for another. For example, the temperature feedback controller, where the signal is compared (by
of the outlet stream from a heat exchanger will almost subtraction) to a reference value. The control function
certainly affect other plant variables-otherwise, it would within the controller generates a signal, which positions a
not be important enough to control. valve on the basis of the sign and magnitude of the dif-
ference between the measurement and the reference or
The control problem setpoint values.
The relationship among controlled, manipulated and In the example for the heat exchanger, a temperature
load variables qualifies the need for process control. The transmitter continuously generates a signal that repre-
manipulated variable and the various load variables may sents the actual temperature of the hot water. At the
either increase or decrease the controlled variable, de- controller, this signal is subtracted from an operator-set
pending on the design of the process. Changes in the
value that represents the desired temperature. If these
controlled variable reflect the balance between the loads values are the same, the current position of the steam
and the manipulated variable. valve is correct, and the controller will not change its
For the heat exchanger, increases in steam-valve open- output. However, if the actual value is below the refer-
ing, steam enthalpy, inlet temperature and ambient tern- ence value, the controller will change its output in the
perature tend to raise the product temperature, while it direction that opens the steam valve and raises the actual
is lowered by increases in flowrate and exchanger foul- temperature. Conversely, if the actual temperature is
ing. The temperature responds to the net effect of these above the desired one, the controller will change its
influences. If the positive influences are greater than the output in the direction that closes the steam valve, to
negative, the temperature will rise. If the reverse is true, lower the actual temperature.
the temperature will fall. If all the load variables were to Thus, a feedback controller solves the control prob-
remain constant, the steam valve could then be adjusted lem through a trial-and-error procedure. Assume that a
until the product temperature was constant at the de- change in the load variables upsets the temperature, and
sired value, and would remain there indefinitely. a new valve position is required. The controller becomes
Process control equipment is needed because these aware of the upset when the imbalance between the
variables do not remain constant. For example, varia- loads and the manipulated variable begins to change the
tions in inlet temperature and flowrate both upset prod- controlled variable. The controller immediately begins
uct temperature, and require a different steam-valve to make corrective changes in its ouputs-even as it
position in order for water temperature to be maintained monitors the effect of these changes on the controlled
at the desired value. Thejob of the control system is to variable. When the controller sees that its correc(ions
determine and continuously update this valve position as have returned the controlled variable to the desired
load conditions change. value (i.e., difference equals zero), it holds the output
Generally, the control problem is to determine the one steady and continues to observe the controlled variable,
value of the manipulated variable that establishes a and waits for the next upset.
balance among all the influences on the controlled vari-
able and keep the variable steady at a desired value. Feedforward systems
Other factors such as speed of response, shape of re- While feedback control is reactive in nature and re-
sponse, and operator interface are also important in sponds to the effect of an upset, feed forward schemes
designing control systems. respond directly to upsets and, thus, offer improved
No matter how complicated, every control system control.
solves this same basic problem, and for a given process The block diagram of a feed forward-control scheme is
and load conditions must arrive at the same result. shown in Fig. 3. Transmitters measure the values of the
The control problem can be solved in only two ways, load variables, and a calculation unit computes the cor-
each of which corresponds to a basic control-system rect control signal for the existing load conditions and
design philosophy. Fcec!bcLcfe systems generate the control reference value. In this way, changes in load conditions

4
Inside a feedback controller
Regardless of the hardware used for implementation,
the concept of feedback control remains the same. The
first feedback mechanisms were mechanically connected
directly to the process and the manipulated variable.
When pneumatic and electronic transmission made cen-
tral control rooms possible, pneumatic and electronic
controllers were developed.
The state of the art today is distributed control
through digital systems, and controllers now often exist
in software. Digital systems may have an extensive selec-
tion of features such as automatic alarming, output
clamps, and built-in linearization or signal compensa-'
tion. However, none of these change the basic function
of the feedback controller-to solve the control
problem.
All feedback controllers must have certain common
elements (Fig. 4). The feedback-control function always
has two inputs and one output. One input will be the
measurement signal from the transmitter; the other, the
reference value. For feedback controllers, the reference
signal is called the setpoint, which usually represents the
desired value of the measurement.
For simple loops, the reference signal may be entered
directly by the operator and is called a "local" setpoint.
In complicated schemes, this signal can come from
cause a direct change in the control signal without another instrument and is defined as a "remote" set-
waiting for the controlled variable to be upset. point. Often, the controller can accept both types of
In general, this technique is more complicated and setpoints, and a remoteflocal switch is available for the
more expensive. It requires greater process understand- operator to select which one the controller will use.
ing than trial-and-error feedback. Therefore, feed for- Within the controller, measurement and setpoint val-
ward control is usually reserved for difficult and critical ues are compared by subtraction. The difference is
applications. called the error and is the input to the mechanism, circuit
or algorithm that generates the output. Generally, this
response contains proportional, integral and derivative
(PID) components, although they may not all be present
in every controller. Proportional or integral responds to
error, while derivative usually responds directly to mea-
surement. The sum of the individual responses forms
the automatic control signal.

Startups and emergencies


For startup and emergency conditions, the controller
will also include a manual control-signal generator that
can be driven by the operator. When the output comes
from the PID response generator, the controller is said
to be in "automatic." When the output comes from the
manual generator, the controller is said to be in ``man-
ual." The procedure for switching between these two
outputs will range from fairly involved to virtually trams-
parent, depending on the sophistication of the control-
ler. The important thing is not to "bump" the output
signal and cause an upset to the process.
In simple loops, this signal will directly position a
valve, while in more-complicated schemes, the signal will
be an input to another instrument. Typically, the con-
troller will have an associated operator interface. As a
minimum, this interface will display the setpoints, mea-
surement, current output, and the remote/local and
automatic/manual status.
Just as all feedback controllers have certain elements
in common, so do all feedback-control loops share three
important concepts: open vs. closed loop, positive vs.

5
negative feedback, and oscillation. Let us now examine
in some detail the significance of these characteristics for
feedback loops.

Open vs. closed loop


Fig. 2 also illustrates the first of these concepts. Once a
feedback controller is installed on a process and placed
in automatic, a closed loop is created. The controller
output affects the measurement, and vice versa. This
closed loop creates the possibility of control through
feedback.
Should this effect be broken in either direction, the
loop is said to be open, and feedback control no longer
exists. Several events can open a feedback loop:
I Placing the controller in manual. This causes the
output to remain constant (unless changed by the opera-
tor) even if the measurement changes.
I Failure of the sensor or transmitter. This ends the
ability of the controller to observe the controlled
variable.
I Saturation of the controller output at 0 or 100% of
scale. This ends the ability of the controller to influence
the process. reinforced the change in measurement. This is positive
I Failure of the valve actuator because of friction or feedback.
debris in the valve. For a feedback loop to be successful, it must have
When a control loop does not seem to be operating negative feedback. The controller must change its out-
properly, the first thing to check is whether or not the put in the direction that opposes the change in measure-
loop is closed. Often, a great deal of time is wasted trying ment. Fig. 5b shows the same loop, except that the
to adjust a controller when the problem is elsewhere in controller has been set to increase-decrease action. The
the loop. controller then responds to increases in temperature by
closing the valve. A decrease in temperature causes the
Positive vs. negative feedback controller to open the valve. These responses tend to
Connecting a controller to a process, as shown in Fig. drive the measurement back toward the setpoint. Select-
2, creates a closed feedback loop. However, feedback can ing the proper control action is as fundamental as mak-
be either positive or negative, and the difference is ing sure the loop is truly closed. The wrong choice
crucial to the loop's performance. destroys control.
Eve.ry feedback controller will have a means of chang- The correct choice for feedback will depend on the
ing the controller action, which defines the direction of Tt

the controller response to a change in the measurement.


Increase-increase (or, direct) action causes the controller
to increase its output in response to an increasing mea-
surement. Increase-decrease (or, reverse) action causes
the controller to decrease its output when the mea-
surement increases. Choosing the wrong action will
make control impossible.
Fig. 5a shows a possible record of an output-tempera-
ture control loop installed on the heat exchanger of Fig.
2. The steam valve is set air-to-open (i.e., fail closed).
This means that an increasing control signal will open
the valve to increase steam flow. The controller action is
set to increase-increase, which is incorrect.
The measurement may bel brought to the setpoint
under manual control, but as soon as the controller is
placed in automatic, the loop becomes unstable. Any
small disturbance that increases the temperature will
also cause an increase in controller output. This opens
the valve, causing the temperature to increase further
and the valve to continue opening. The result is a
runaway temperature. If a small disturbance caused the
temperature to drop, the controller would close the
valve, and the temperature would fall even more. In turn,
this would cause the valve to close even more.
In both cases, the response of the controller has

6
cause of the lags within the process, the outlet tempera-
?=-iifet3£=--ri.`.I=t'.
ture does not respond immediately. In fact, it continues
to move away from the setpoint. The controller then
---:`: -+:-rlr-, -.-:. -:--•..;..........:.:.:..-........;.:...gr.I.,.'.as;..:.-...,.:..... .'12`

continues to change its output until the measurement


turns around and begins to return to the setpoint.
When the measurement reverses itself, so will the
controller output, but the effect of this reversal will also
be delayed. Later, the measurement may reverse a sec-
ond time and cause another reversal in the controller
output. In turn, this causes another reversal in the
measurement, and so on. The result is an oscillation in
both the measurement and the controller ouput.
Time---=
Thus, the combination of negative feedback and lags
Time =
in the process means that oscillation is the natural re-
a. Positive feedback causes instability
sponse of a feedback control loop to an upset. The
characteristics of this oscillation are the primary means
for evaluating the performance of the control loop.
Specifically, an instrument engineer will be interested in
the period and the damping ratio of the cycle.
Fig. 5c shows a typical oscillation. The period of this
cycle may be measured as the time (usually in minutes)
between any two analogous points, such as between two
positive or negative peaks. Fig. 5c also shows another
oscillation that is steadily decaying to a constant signal.
The damping ratio measures the rate of decay.
Although there are mathematical definitions of the
Time = Time - damping ratio, practically it may be measured as the ratio
of the deviations of any two successive peaks from the
b. Negative feedback causes stability
estimated final or average value. These measurements
are usually taken from a record of the controlled variable
because it is often i`ecorded. However, the same cycle
can be observed in the controller output, or in any
measurement directly affected by the control signal. For
example, if a record were kept of the steam flow to the
heat exchanger, the cycle would also appear there. Fre-
quently, other variables will provide a more sensitive
representation of the cycles within a loop, and these will
Time--- = Time - allow more accurate evaluation of loop performance to
c. Oscillating signals
be made.

Characteristics of the oscillation


The exact characteristics of the oscillation in a particu-
lar loop will mainly depend on the adjustments to the
proportional, integral and derivative responses within
application. For example, if tank level is controlled by the controller. Incorrect adjustments can make this pe-
manipulating an air-to-open valve on the outflow, in- riod too long or too short. Even worse, they can make the
crease-increase action will be needed. Moving the same cycle grow larger instead of smaller.
control valve to the inflow requires increase-decrease For good control, the cycle in the measurement signal
action. Reversing the action of the valve to air-to-close should steadily decay, and end with the measurement
(i.e, fail open) can reverse the required control action. returned to the setpoint. Simultaneously, the cycle in the
A controller taken out for maintenance might not be controller output should also steadily decay, and end
set correctly when it is reinstalled. Sometimes, position- with the output at the new value. This re-establishes
ers on valves can reverse the response of the valves to a balance among the load variables and the manipulated
change in the control signal. The penalty for not think- variable.
ing this out is a control loop that drives the measurement In fact, this oscillation represents the trial-and-error
to one of its range limits. search for the new solution to the control problem. The
controller is not aware of the load variables. Hence,
Oscillation when it sees the measurement begin to change, it tries
While negative feedback is necessary for control, it new output values until it narrows in on the one value
also leads to oscillation within the loop. Once again, let that returns the measurement to the setpoint.
us consider the temperature control loop in Fig. 2. When If the controller in a particular loop responds to an
the measurement begins to move away from the set- upset with an oscillation in which each successive peak is
point, the controller begins to change its output. Be- one-fourth as large as the preceding one, the loop is said

7
L,:,: , , , ., , ,. , ,1, , 1, , ,: : ,
to have quarter-wave damping (i.e., B/A = 1/4 in Fig.
5c). Depending on the period, a loop having quarter-
wave damping stabilizes fairly quickly following an upset.
Often, this is taken as an indication of good control.
Determining proper controller adjustments is somewhat
more complicated than achieving this one objective.
Nevertheless, quarter-wave damping may be used for a
rough evaluation of controller performance.

Process characteristics
The existence of lags in the process has a fundamental
effect on the performance of the feedback loop. Without
understanding the causes and characteristics of these
lags, it is impossible to evaluate which control modes
(proportional, integral, derivative) will be required, or
whether feedback control will be successful in any partic-
ular application. Basically, lags may be considered in two
categories: deadtime and capacity.

Deadtime
A process that has essentially pure deadtime response
is shown in Fig. 6a. A hopper valve deposits material on a
moving belt. A weight transmitter measures the amount
of material. How does the weight measurement respond
to changes in the control signal to the hopper valve?
As shown in Fig. 6a, a step change in the control signal
will immediately begin to deposit more material on the
belt. This step change will appear in the measurement
after a delay (deadtime) that corresponds to the time
necessary for the material to travel from the hopper to
the sensor.
In general, deadtime is defined as the time delay
between a change in the control signal and the beginning
of its effect on the measurement. The shape of the
change in the control signal is not relevant. Fig. 6a also
shows an oscillating control-signal input delayed by the
same time interval.
Because deadtime is often caused by the time required
to move material from one point to another, it may be
referred to as transport lag or distance/velocity lag. The
actual time depends on the distance traveled and the
velocity of the material.
Delay in the process response can be created in other
ways. The performance of mixers (i.e., agitators) has a
large influence on the deadtime in loops monitoring
composition, such as pH, density, or oxidation-reduction
potential. The sampling operation of a chromatic ana-
lyzer will also create delay in the perceived measurement.
And, significantly, a combination of a number of capac-
ity-lag elements will also create deadtime.
From a control point of view, what is important is the
length of the delay. Deadtime represents an interval
during which the controller has no information about `
the effect of a control action already taken.
Deadtime does not slow down the rate at which the
measurement can change. Except for the delay, the
measurement changes at the same rate as does the
control signal. Still, the longer the delay, the more
difficult it will be to control. As will be shown, the
amount of deadtime in the process has a strong effect on
the controller adjustments and on the perforinance that
can be expected from the loop.
Because deadtime interferes with good control, every

8
T| I

I -\,.

attempt should be made to reduce this delay by properly 2. The capacity inhibits the rate at which the measure-
locating transmitters, specifying sufficient mixing, de- ment can change.
signing proper tankage, and minimizing transmission Because level is a measure of the liquid stored in the
lags. tank, and because the rate of accumulation (positive or
negative) responds to the difference between inflow and
Capacity and its effects outflow, level cannot change instantly even if the control
Pure deadtime processes are rare, and virtually every signal does. The bigger the tank in comparison with the
control loop will include, and will be dominated by, flows, the slower the level will change. Therefore, the
capacity elements. capacity element in the process tends to attenuate distur-
A capacity element is that part of the process system bances. This makes control easier, whereas deadtime
where material or energy can accumulate. The tank makes control more difficult.
shown in Fig. 6b represents a single capacity (material The size of a capacity is measured by its time constant.
storage). Flow into the tank is manipulated to affect the Fig. 6c shows, in more detail, the level response of Fig.
level; flow out of the tank is the load variable. Initially, 6b. Since the two flows (in and out) approach equality
the level remains constant because inflow and outflow asymptotically, they never quite become equal-at least`
are equal. How does the response of this process differ in theory. The level never stops changing and, therefore,
from that of a deadtime element? the response cannot be measured by the time to
Let us assume that the valve and flow respond in- completion.
stantly to changes in the control signal. When a step Instead, the response is quantified by a time constant
change occurs in this signal, the difference between in- that is defined as the time required to complete 63.2% of
flow and outflow will immediately cause an increase in the total response. (This number is not`arbitrary. It has
level. However, as level increases, the gradually increas- significance in terms of the differential equations that
ing pressure across the drain valve raises the outflow. model the process.) As a first approximation, the time
This tends to bring the two flows back into balance, with constant of a capacity element will be roughly equal to its
the net result that level rises more rapidly at first, then residence time, which is defined as the volume divided
more slowly, and finally stops as the flows become equal. by the throughput (in consistent units). Thus, if the tank
The other vessel shown in Fig. 6b also represents a in Fig. 6b holds 1,000 gal, and flow through the tank is
single capacity (energy storage). Temperature responds 100 gpm, the residence time becomes 1,000/100 = 10
to the accumulation of energy in a process just as level mln.
responds to the accumulation of material. The response Fig. 6b also shows the response of a capacity element
of the temperature to a step change in heat input will be to a cycling control signal. If the signal cycles the inflow,
the same as the response of the level to a step change in the outflow will approach the average value of the inflow.
flow input. The level will rise while the inflow is greater than the
The responses of these capacity elements differ from outflow; and it will fall while the inflow is less than the
that of the deadtime element in two significant ways: outflow. In short, for a cycling input, the measurement
I. No delay occurs before the measurement begins to signal from a capacity element will also cycle at the same
change-i.e., no deadtime is associated with a single- period.
capacity element. The variation in the measurement signal, in compari-

9
son with `the variation in the control signal, depends
strongly on the period. If the control signal cycles very
rapidly (with a short period), the swing in the level will be
very small. Conversely, if the same variation in the
control signal occurs at a much longer period, the swing
in the level will be much greater.

Modeling the process


Single-capacity and pure deadtime processes exist
only in theory. Any real processes will include a number
of each of these dynamic elements. For example, the heat The open-loop response of a heat exchanger to a step
exchanger, shown in Fig. 1a, includes a deadtime associ- change in the controller output is shown in Fig. 8.
ated with the time it takes for the hot water to flow from Initially, the temperature remains constant but later
the exchanger to the sensor. In addition, the identifiable begins to rise and approaches a new steady-state value.
capacities include : Although a process may actually be an intricate collec-
I Volume of the air actuator for the control valve. tion of deadtime and capacity elements, it can usually be
I Volume of the exchanger shellside. represented by a deadtime-plus-capacity model in order
I Energy stored in the tubes. to design the feedback loop. The parameters for this
I Energy stored in the water in the tubes. model may be taken as the apparent deadtime and the
I Energy stored in the thermowell and sensor. apparent time constants.
If the controls are pneumatic, an effective deadtime While this representation may be obvious to the de-
and capacity are also associated with each transmission signer, the controller cannot tell the difference. Since
line. This is a typical situation-one or two identifiable deadtime makes control difficult while capacity makes it
deadtimes, and a number of large and small capacities. easier, an estimate of the difficulty of control can be
Deadtimes in series are additive-a 1-min delay fol- made by calculating the ratio of the apparent deadtime
lowed by a 2-min delay combine to form a 3-min delay. to the apparent time constant. This ratio, TD7/7.I, will
However, the combined effect of a number of capacities also have a strong effect on the control adjustments.
in series is not so obvious. Fig. 7 shows a series of three The behavior of feedback control loops can be under-
capacities having an equal time Constant, T7ic, along stood from practical or theoretical points of view. Al-
with the responses at various points to a step input. The thoughwehavethusfaremphasizedthepractical,under-
step input appears at Point I . Point 2 shows the response standing the two mathematical concepts of gain and
of a single capacity to a step input, as shown in Fig. 6c. phase is essential to a fundamental knowledge of feed-
Points 3 and 4 show the effect of subsequent capacities. back control.
The net effect is that a sequence of capacities looks (to
the controller) like the combination of a deadtime delay,
Gain and phase
followed by a single capacity with a time constant, 7-I, that An element from a feedback control loop is repre-
is larger than the time constant of the individual sented in Fig. 9. This element could be the process, the
capacities. valve, the transmitter or the controller. Each of these ele-

10
ments has an input and an output. The first parameter, Betlnning at any point in the loop, let us consider the
gain, describes the amount of change in the output that effects on that signal as it travels once around the loop.
will be caused by a stven change in the input. Both The signal is made larger or smaller as it passes through
s[cady-state and dynamic gains must be considered. For each element, according to the gain of that element. At
a step input, the output of the element begins to change the same time, the signal will be somewhat displaced,
and approaches a new value. The steady-state gain, Gas, is according to the magnitude of the phase angle associated
defined as the ratio of the final change in the output to with that element.
drge in the input, or: For the cycle to continue, the total effect of these
displacements must equal 360 deg., so that the signal
Gss = A(Out)/ A(In) (I)
returns to the berinning point. Therefore: a feedback
of the uhits for gain. For control loop will cycle at that period which makes the
kispi:pi:r:::t::e::eps
example, if the steady=-asta gain of the valve in the sum of the phase angles equal to 360 deg.
temperature loop were being determined, the output More importantly, the net effect on the size of the
would be in units of steam flow, while the input would be signal depends on the product of the individual gains, or
percentage. Thus, if a 10% change in controller output the open-loop gain, Gal:
caused a change of 200 lbth in steam flow, the steady-
COL --(GD)c(GD)v(GD)p(GD)t (6)
state gain becomes:
where (GD)c is the dynamic gain of the controller, (GD)u is
Gss = 200/10 = 20 (lb/h)7%; (2)
the dynamic gain of the valve, (GD)p is the dy-
However, the signals traveling around a control loop namic gain of the process, and (GD), is the dynamic gain
usually vary cyclically. The sensitivity of an element to a of the transmitter.
cycling input is measured by its dynamic gain. When the The dimensional units for the individual gains must be
input cycles, the output will also cycle at the same period specified in such a way so that they cancel when the
(see Fig. 6a and 6b). The dynamic gain may be computed open-loop gain is calculated from Eq. (6). If that gain is
as the ratio of the size of the output swing, Acha„ to the greater than 1.0, the signal will arrive at the beginning
size of the input swing, A7„ or: larger than when it started. As it continues to travel
around the loop, it will continue to grow. At any one
GD -AowJ Aln (S)
point in the loop such as at the measurement input to the
For the heat exchanger, let us suppose that a 200 lbth controller, the signal will appear as an ever-increasing
variation in steam flow caused a 20°F variation in outlet oscillation. Therefore, a feedback control loop will be
temperature. The dynamic gain for this situation stable only when the product of the dynamic gains in the
tHomes: loop is less than 1.0.
Adjustments to proportional, integral and derivative
GD = 20°F/200 lb/h = 0.1°F/(lbth) (4)
responses affect the gain and phase parameters of the
The second parameter of the response of an element controller and, in turn, the behavior of the entire loop.
to a cycling input is the phase angle, which is illustrated These concepts will be explored in subsequent articles of
in Fig. 9. Because of the lags (i.e., delays) within the this series.
element, the peak of the output does not coincide with
the peak of the input. The phase angle, ¢, of an element Su-any
measures this displacement. One complete cycle in any The purpose of every control loop is to find the one
periodic signal is considered to be composed of 360 value for the control signal that holds the measurement
degrees. If the peak of the output cycle occurs one- at the setpoint for the existing load conditions. A feed-
qualter of the way through the input cycle, the phase back or feed forward approach may be used. In this
angle is: article, we have concentrated on feedback techniques. In
a later article, we will cover the feed forward approach.
¢ = (360)(-1/4) = -90° (5) The next article in this CE REFRESHER will appear in
In Eq. (5), the negative sign indicates that the peak of the Aug. 8 issue, and will analyze the actions and reac-
the output occurs after the peak of the input. This is tions for basic feedback-control modes.
termed a phase lag. It is also possible for the output peak Steven Deyndos , Editor
to occur before the input peak; and this is called a phase
had.
The author
CHosed-loop applications IIewis M. Cordon is a systems
application enctneer at The Foxboro
Theparametersofgainandphasearefundamentalfor Co., Foxboro, MA 02035. Currently, he
is involved in advanced control systems,
understanding the behavior of a feedback loop. They are
especially important in the study of controller tuning fy::::indgesEi=edseacl:Le::guqoonn,a:odntrol-
startup. He also conducts seminars for
because both are functions of the period of the imp.ut personnel at Foxboro and offplant sites.
rfu.
When a feedback control loop is upset by a change in
Previously, Mr. Cordon spent seven
years with Foxboro Educational
Services. He has a B.S. and and M.S. in
either the load conditions or setpoint, it will begin to mechanical enctneering from the
University of California (Davis).
usrillate at some period characteristic of that loop. Every
element in that loop sees an input signal varying at that
pernd.
11
Feedback control modes
Control modes are specific responses to a change in the
measured variable or error signal. The analysis of control
modes and their combinations will show how to improve the
stability and speed of response for closed feedback loops.

Lewi,s M. Cordon, The Ft]xbt)ro Co.*

I Understanding the individual lnodes in a controller is tified in different units. The de]`ivative response may bc
essential to successfully apply feedback control. These generated in several ways~and varying degrees of inter-
modes involve: on-off, proportional-only, integral, and action are possible among the proportional, integral zlnd
derivative actions. Each possible combination represents derivative modes.
a tradeoff between cost and performdnce. For specific situations, many specidl features have
A feedback controller must be connected in a closed been added to improve control, such as external integral
loop, and appropriate control action selected, to estab- feedback, batch switches, tracking, and output biasing.
lish negative feedback. Given these essentials, the con- In the future, the flexibility inherent in digital feedback-
troller can solve the control problem by a trial-and-error algorithms will increase the specialization and variety of
search for the output that establishes a balance among all feedback controllers. Nevertheless, control systems will
the influences on the controlled variable. still be built ()n the foundation pr()vided by the basic
Selecting the proper control action establishes nega- responses.
tive feedback by defining the direction of the controller A controller is a nonthinking device-its responses are
response. The next objective is to determine the magni- built in. It is up to the designer to select those zippropri-
tude of this response. ate lo the apt)licalion. Specifying the wrong combination
of c()ntr()I modes leads to poor system performance,
Control modes increzises the complexity of the tuning pr()blem, and may
A controller in a feedback loop is in a difficult position. add unnecessary c()st.
Unpredictable forces can influence the measurement it
is trying to control. Even worse, the dynamic characteris- On-off control
tics of the rest of the loop will delay and distort the On-off or two-position response is the simplest form
output variations used by the controller to reduce error. of feedback control loop. Fig. I shows the performance
In this environment, it is misleading to believe that of this loop for a process in which liquid is being heated.
control is imposed on the process. Instead, the relation- An on-off control function has only two p()ssible out-
ship between a controller and the process is interactive. puts (on,10097o; or off, 097o), and only considers the sign
Here, the size, shape and rate of the variations in the of the error. In the example, the controller closes the
controller's output are crucial as the controller restores fuel valve when the measurement rises above the set-
the measurement to the setpoint value following an point (Fig. Ib). Because ()f deadtime and/or lags in the
upset. process, the temperature continues to rise before revers-
A control mode is a particular controller response to a ing and moving toward the setpoint. When the tempera-
change in the measurement or error. The four basic ture falls below the setpoint, the controller opens the
responses are: fuel valve. Deadtime and/or lags in the proc.ess again
I On-off (two-position). create a delay before the temperature begins to rise. As it
I Proportional. crosses the setpoint, the controller again shuts off fuel
. Integral. flow, and the cycle repeats.
I Derivative. Cycling is the normal condition for a loop under on-
Variations on these basic responses exist among man- off control. This limitation arises because with only two
ufacturers of control instrumentation. Sometimes, these possible outputs the controller is unable to solve the
responses are identified by different names, or are quan- control problem exactly. The output is either too high or
*To meet the author, see Che7%. E7.g., May 30, p. 66.
too low to establish a balance among all the influences
Articles published thus far in this CE REFRESHER: Part I, May 30, p. 66. on vessel temperature. A 100% output supplies too

18
depends on the length of the period and the rate at which
the measurement changes. Since capacity inhibits mea-
surement change, the amplitude is inversely proportion-
al to the time constant, 7.1, of the process. [See Part 1 for a
discussion of period, amplitude and deadtime.]
On-off control should be applied to those situations
where three conditions are present:
1. Precise control must not be required, because the
measurement will constantly cycle.
2. Deadtime must be moderate to prevent excessive
valve wear because of too short a period.
3. The ratio 7.D7i/7-1 must be small to prevent too large
an amplitude in the measurement cycle.
When these conditions apply, the simplicity and econ-
omy of on-off control offer significant advantages.
A variation of on-off control that reduces wear on the
final operator, and that may be described as differential-
gap or gap-action control, is shown in Fig. 1c. Instead of
changing the output in both directions at a single point,
the control function may take action only at specified
high and low limits. As long as the measurement remains
within the gap, the controller holds the last output state.
As Fig. 1c illustrates, the effect of this variation is to
extend the period, and to increase the amplitude.
Often, the size of the gap will be adjustable and need
not be symmetrical, hence some acceptable compromise
can be achieved. Typically, an on-off controller will have
a very small gap designed into its mechanism.

Controller response: open vs. closed loop


As was discussed in Part 1 of this series, feedback
control requires a closed loop. The closed-loop response
much heat, causing the temperature to rise. A 0% output is to a change in the setpoint or in the measurement
supplies too little heat, allowing the temperature to fall. caused by a load upset. The simplicity of the on-off
Negative feedback causes cycling between the two function allows it to be presented in terms of its closed-
conditions. loop response. However, the interaction between the
controller and the process in this configuration obscures
Applying twolposition control the properties of the proportional, integral and deriva-
The principal disadvantage of on-off control is con- tive control-modes.
stant cycling; the principal advantage is low cost. Be- A controller is isolated from a process in order to
cause of its simplicity, on-off control will be the least study its open-loop responses (Fig. 2). Here, the control-
expensive approach to feedback control. It does not even ler receives an artificial measurement and a setpoint.
require a controller; the same function can be created The difference between these values generates an error
with alarms, contacts, digital outputs, and relays. signal, and the controller output is merely recorded. In
Acceptability of on-off control depends on the charac- this configuration, the effect of a change in the control-
teristics of the cycle in the measurement. If the ampli- ler's output does not appear at the measurement pbint
tude of the swing is too large, unacceptable variations in where it would cause further changes in the output. Any
product quality, or upsets to other process units, may desired measurement or setpoint change may be ap-
occur. If the period of the cycle is too short, the wear on plied, and the controller's response observed on the
the valve and/or upsets to the fuel distribution system recorder.
(Fig. Ia) may be unacceptable. An on-off control function cannot solve the control
The period of the cycle depends on how long it takes problem, because it has only two possible outputs. For
for the measurement td turn around after a change in the example, the reponse of an on-off controller to the
val`'e position. Thus, the period is directly proportional smallest error is to drive its output to one extreme. To
tD dcadtime, 7.Dr. If` the deadtime were reduced to zero, solve the control problem, the controller must also be
the measurement would instantly reverse itself with each able to generate any output between the extremes. Pro-
dlange in controller output. Since the output reverses portional, integral and derivative modes have this capa-
cach time the measurement crosses the setpoint, both bility, but each is based on a different concept of how the
the period and the amplitude would be reduced to zero. controller should respond to an error.
Control would be very good, but the valve wear would be
crde and unacceptable. Proportional control
Amphitude`of the cycle depends on how much the Proportional control is based on the principle that the
in-Irement changes before it reverses. In turn, this size of the controller response should be proportional to

13
the size of the error. To achieve this, proportional
control ties the change in output to the change in error, Controller boundary `` .:;

with both values being usually expressed as percent of I:+f:'.'.SS±

range. setpoint
Fig. 3 is a graphical representation of proportional
ri:..` .<.

}: ;S:i,: f: ;*:£i

action. Regardless of how proportional action is created


(pneumatic, electronic or digital), this effect may be
imagined as a double-ended pointer, pivoted in the Art lclal
measurement
middle [for a proportional band = 100%], and moving :!integral+divatecions±
PID = proportional +
along an error scale and an output scale. Changes in
either the measurement or the setpoint create changes in
the error, which drives the lefthand end of the pointer.
The righthand end indicates the corresponding output.
•;....;::..a:::....::..:.:..-.,.,.;.::`i*i,::::;:::,;,:;:::;:.:%:;:;:.: `.=i:l:l:-.i-:if'..i?±i=

-
As shown in Fig. 3, the output scale describes increase-
decrease (J/D) action. Changing to increase-increase ac-
tion simply reverses the output scale. Dynamic properties of proportional action
Fig. 3 also illustrates two properties of proportional
Measuring proportional action action that have the most influence in a closed loop. Pro-
Fig. 3 illustrates several important concepts about portional action is both immediate and specific.
proportional action-the first of which is proportional I . The linkage between the error and the output, rep-
band, PB, or gain, G. These adjustable parameters define resented by the pointer, means that the output change
how strongly the controller reacts to changes in the occurs simultaneously with error change. No delays
error. The location of the pivot, as shown in Fig. 3, fixes occur in the proportional response.
the amount of output change` for a given error change. 2. Each value of the error for a given proportional
With the pivot in the middle, a 100% change in measure- band generates a unique value of the output. The pro-
ment (from 50% below the setpoint to 50% above it) will portional-response generator is incapable of any other
cause the output to change from 0 to 100%. Moving the combination. This one-to-one relationship between the
pivot to the left can reduce the measurement change error and the output places severe limitations on the
required for a 100% output change to 50%, i.e., from closed-loop performance of proportional-only control,
25% below to 25% above the setpoint. In the same way, as will be described shortly.
moving the pivot to the right will increase the percent Fig. 4 presents another graphical representation of
change in error required for full-valve travel. proportional action. Each value of the proportional band
The proportional band, PB, is defined as the percent
change in measurement (at a constant setpoint) required
to cause 100% output change. Gain, G, is defined as the
ratio of the output change to error change. Both quantify
the same thing-the sensitivity of the controller to
changes in the error, and each can be expressed in terms
of the other:
G -100%/PB (1)
The relationship of Eq. (1) can also be expressed in PB = 1000/o

the form of a matched scale:

PB = 500/o

Fig. 3 also illustrates the concept of the proportional


bias. Regardless of the value of the proportional band,
the output will _be 5097o when the input error is zero
(measurement at the setpoint). This bias on the output
gives the controller a value around which it can vary its
output to reduce error. As the error increases (or be-
PB = 200%
comes negative), the output changes from 50%, accord-
ing to the value of the proportional band. Normally, the PB = proportional band
bias in a proportional controller is adjusted at the instru-
ment factory but some manufacturers make this adjust-
ment available to the operator.

14
band equals 50%. Then, for increase-decrease action:
07D = (60 -40)(100/50) + 50 = 90%
here O+D = output for increase-decrease action.
Increase-increase action is achieved by reversing the
calculation of the error for Eq. (2) within the controller.
Then:
0,, = (40 -60)(100/50) + 50 = 1097o

The straight-line relationship between error and out-


put identifles a proportional-only controller as a linear
or constant-gain device. In this representation, the spe-
-50 -25 0 +25 +50
cific character of proportional action means that the
Error, e; or offset, eo, °/o
PB = proportional band
coordinates of error and output must identify a point
falling on the given proportional-band line, and the op-
Crating point for the controller can only move along this
line.
As the proportional band is decreased, proportional
action is concentrated into a narrower band around the
defines a specific relationship between error, c, and setpoint. From a gain point-of-view, the same change in
output, 0, which may be expressed as: error causes larger changes in output. In the limit, the
O = c(100)/PB + 5097o (2) proportional band equals zero (gain equals infinity), and
the smallest error causes the output to go full scale. On-
where 0 is output, % ; e is error, % ; and PB is proportional off control, then, becomes a limiting case of propor-
band' %. tional-only control. At the other extreme, when the pro-
For example, assume that setpoint is at 60% of scale,
portional band equals infinity (gain equals zero), the con-
measurement at 40% of scale, and that proportional troller simply does not respond to changes in error.

Applying proportional-only control


A level process under proportional-only control is
shown in Fig. 5a, where the outflow is the load on the
process. To control the level, the controller must bal-
ance the outflow by manipulating inflow. This requires
increase-decrease action. Both flows vary from 0 to
100%, the setpoint of the controller is 50%, and the
proportional band equals 100%.
As a starting point, assume that the load equals 50%
and that the level is at the setpoint. Then, the controller
output will also be 50%, inflow will equal outflow, and
level will remain constant.
Next, assume an upset in the form of a load decrease
to 2597o, i.e., outflow is reduced. How will the loop
respond to this upset?
Since the outflow is less than the inflow initially, level
will begin to rise, and the error will begin to go negative.
By referring to the 100%-PB line in Fig. 4, it will be seen
that controller output (for increase-decrease action) will
simultaneously begin to decrease as the operating point
moves toward the upper lefthand corner of the chart.
This action gradually restricts inflow until it equals 25%
when the level has risen to 7597o (Fig. 5a). Then, inflow
equals outflow, and the level will remain constant.
The controller cannot return the measurement to
setpoint. When the proportional band equals 100%,
output equals 25%, but only when the error is -25%.
Therefore, a steady-state deviation is required to balance
the load on the process. In the same manner, if the load
increases to 75%, level will fall until it is 25% below set-
point, where the 75% output from the controller again
balances the load on the process.
This steady-state deviation from setpoint is called
``offset." It arises because the bias, 8 (output when the
measurement equals setpoint), is fixed. If the loading

15
\\,,,,`````
`` ` `tt i

conditions require an output different from the bias, \,'

\```` ` `t`

+Ia,LE
some steady-state error must be present. Each variation \,

in the load will require a different output and a different -.-..fro ---tffi*iE`-i`` - tF'`1`\`i` `

offset. The amount of offset, co, is a function of the Ei !Ti-iPolntA8C-=±_I, \\\`..`,-.``,:,,,,\


t`tl,,,,,`',\

required output and the proportional band, according to


the following equation describing a proportional-only
controller: a) + t---tan---E5 -- RE,i*`!.+a.,i.,=ffi- .+
" --tt ffita

O = e(100/PB) + B
eo = (PB/100)(O -B)
(3a)
(3b) a. Integral action responds to'® ign, size and duratio',,n of er
e- r \\,`,\ ',\

Thus, the purpose of an adjustable bias becomes clear. ----givpeif}t i


-aErfu-
By changing the bias on the proportional response to :i-------
equal the required output, the measurement can be -::S r- 1
returned to the setpoint. This adjustment is often called

-
rfeSRES
"manual reset."
55=----i- -th
t'

`1
a ce © © 'jl6.a
espQ\ se`;a/repea
tS \,&`

Assuming the bias remains fixed at 50%, the offset for --#+lr,t ` ``,`

\`

f¥i FTt-±J:
t!

a required output is also seen to vary with the propor-


tional band. Referring to Fig. 4, if the loading conditions T
require a 75% output and the proportional band is b. Integral time determines rate of response \ ``\` `

200%, offset will be 50%. Reducing the proportional 100 \\


\ ``\
```::`````:`'`

band to 50% reduces the required offset to 121/2%. :o50J25


75±h-uPSSt ;
t t,

However, reducing the proportional band also increases


the gain of the controller and reduces the damping in the
closed-loop response. a004§ oseaa ffiae`

®~ 0 4Gas €- ^^-
Fig. 5b shows the effect of narrowing the proportional ®Eoo=E-2oio*ao'aoofroaiQ?o® ``Oc'a•©,©',,3 a• oo. .rfu#i#t I ````\tir
``` ``

'\

band on the closed-loop response to a load upset:


I Case A-The controller does not respond. The mea- 100se1--
surement falls to a new steady-state value. 0=` § § 75 ---rty -, --
i-ife*n®i#*\t '

I Case B-The proportional response is too weak,


leading to excessive offset.
+ a ®
6±% 25
50 \ Iime-
I Case C-The proportional band is correct. The `,`,I

response of the controller is just strong enough to cause c. Integral action shifts the bias to balance the lo
`,\ , ri:;f.: : Fi:i;i:i;. ` `, \'fi.`,`,`\, `````,'`'
quarter-wave damping. \''
I

I Case D-The proportional band is too narrow. The I tntngra* astlSn *mprnveSth # *¢#*#¢1\
++
#ffiasffiRE`!`RES
a,,

overreaction causes excessive swing in the measurement,


which takes too long to even out.
If the proportional band is reduced too much, the gain
in the controller will become high enough to make the proportional action, integral action also responds to the
open-loop gain greater greater than 1. Instead of decay- error. However, integral action is based on the principle
ing, the cycle for both the measurement and the control- that the response should be proportional to both the size
1er output will grow until the valve cycles between its and duration of the error.
limits, as in on-off control. The open-loop response in Fig. 6a shows how integral
For every process under proportional-only control, action is related to the error. Initially, while the error
one particular proportional band (i.e., gain) creates the equals zero, the output remains constant at a value that
desired closed-loop response. The exact value will de- depends on the history of the error. Errors in the mea-
surement will produce the following:
pend on the other elements in the loop, each having
individual gains. In general, where process gains are low I Point A-A constant error appears. The integral
because of a small 7-D7/7i ratio, the required proportion- responds by driving the output at a constant rate, pro-
al band will also be low. Once tuned, however, offset will portional to the size of the error, as long as the error
vary with the load on the process, as in Fig. 5a. remains constant.
Proportional control is a major improvement over on- I Point B-The size of the error increases. The inte-
off control because of its ability to stabilize the loop. Its gral responds by driving the output at a faster rate.
main disadvantage is the inevitable offset. Where the I Point C-The sign of the error changes. The inte-
loads are fairly constant and the required proportional gral responds by driving the output in the opposite
band is narrow, offset will not be a problem. The set- direction.
I Point D-The error returns to zero. The integral
point can be adjusted until the measurement is at the
desired value. The setpoint is then no longer the desired action stops at the existing output value.
measurement value but simply a reference for propor- I Point E-The error increases at a constant rate. The
tional action. integral responds by driving the ouput at an ever-in-
creasing rate.
Integral action I Point F-The error returns to zero. The integral
Integral action may be combined with proportional action ceases at that output.
action to eliminate offset where it is unacceptable. Like These responses illustrate the most significant prop-

16
elty of integral action. Whereas proportional action ties
the output to the measurement through the error, inte-
gral action can achieve any output value-stopping only
when the error is zero. This is the property that enables
integral action to eliminate offset. Integral action is only
sadsfied when the measurement has returned to the
setpoint. As long as an error exists, integral action will
drive the output in the direction that reduces error.
The open-loop response in Fig. 6b shows how propor-
tional and integral actions are combined in a controller.
Initially, the output is constant because the error is zero.
When a step change in the error appears, a simultaneous
step change occurs in the output because of proportional
action (see Fig. 3). The size of this response depends on
the proportional band. At the same time, the integral
action begins to drive the output, as shown in Fig. 6a.
For a constant error, the adjustment to integral action
changes the rate at which the output is driven. This rate
f quantified in terms of the time required for the change
in output (due to integral action) to equal or repeat the
response caused by proportional action.
Some instrument manufacturers use dimensional
units of minutes/repeat, referred to as integral time.
Others use units of repeats/minute, referred to as inte-
gral gain. Each is simply the inverse of the other, as
shown in the chart:

term as a function of the error. When the response is


complete, the bias term has increased to 75%, and the
proportional term has returned to zero. The 75% bias
means that the proportional band has shifted so that the
range of proportional action extends from 1097o below to
30% above the setpoint. Thus, integral action continu-
ously performs the manual-reset function, described
Increasing the integral time, or lowering the integral earlier.
gain, reduces the strength of the integral action. The ability of integral action to eliminate offset is very
advantageous, and integral action is almost always speci-
Applying integral action fled for feedback control. However, this action does have
The combination of proportional-plus-integral action a significant disadvantage: To create its gradual re-
can also be expressed in equation form: sponse, a capacity-like lag is built into the controller.
This causes a phase lag across the controller and length-
O=(i#)co+(i#)(+)/codf (4) ens the period of oscillation of the loop, as a function of
the relative contribution of proportional and integral
When Eq. (4) is compared to Eq. (3a), which describes actions.
a proportional-only controller, the only difference is in Typically, the period of oscillation for a loop under a
the bias term. When proportional-only control is limited
properly tuned proportional-plus-integral controller will
by a fixed bias, integral action [Eq. (4)] uses the integral be 5097o longer than if the controller were proportional-
of the error to adjust the bias-stopping when the error only. For relatively fast loops such as flow control, this
equals zero. will not be significant. However, for slower loops, exten-
Fig. 6c is a representation of how integral action sion of the period can be a serious limitation. For loops
eliminates offset, following a load upset. Initially, at 50% where the exact value of the measurement is not critical
load, a 50% output holds the measurement at setpoint. (as in level control), the shorter period of a proportional-
In the steady state. this is also the value of the variable only controller can be an advantage.
bias, since the error equals zero. The controller has a Like proportional action, integral action increases the
40% proportional band. The 50% bias indicates that the
gain of the controller. Too much of either can cause the
40% variation in measurement over which proportional loop to cycle. In general, integral time should be propor-
action will occur is centered around the setpoint. When tional to how fast the process responds to control action.
the measurement begins to fall, following a load in- If the time is too short, it will drive the final operator to
crease, proportional and integral actions return the mea- its limit before the measurement is able to respond.
surement to the setpoint via a quarter-amplitude Then, when the measurement does respond, it will
damped response. overshoot the setpoint-causing the integral to drive the
The contribution of integral is to increase the bias operator to its opposite limit.

17
Applications in which sustained errors are likely (batch vanced in time. The size of this advance is the derivative
processes or those having large setpoint changes) can time, D„ min. Derivative action is sometimes erroneously
lead to integral (or reset) windup. Although the integral referred to as "anticipating" action. (Note: The control-
time may be correct for normal control in such situa- ler can only respond to a real error, and cannot antici-
tions, the output may saturate during the sustained pate the arrival of an error.) Increasing the derivative
error, and lead to overshoot when the measurement time will generate a larger derivative response that will
finally approaches the setpoint. appear as a larger time difference between the two
In these applications, a "switch" may be added to the responses in Fig. 7b.
integral circuit (whether electronic or pneumatic) of the Following the techniques for proportional and inte-
controller. This switch has b;come known as a "batch gral actions, earlier controllers applied derivative action
switch" because the windup problem is primarily associ- to the error. However, this causes the derivative action to
ated with discontinuous or batch processes. Newer con- respond to both measurement and setpoint changes.
trollers and control algorithms are designed to avoid the Since setpoint changes are usually made stepwise, this
integral saturation or windup problem. approach often "bumped" the process with large output
spikes, as shown in Fig. 7a.
Adding derivative action Almost universally today, controllers are designed so
Proportional and integral actions share one serious that the derivative-response generator looks only at the
limitation. A significant error must be present before measurement signal. Initially, only the proportional and
either of these modes generates a strong response. integral actions respond to changes in the setpoint.
Derivative action is based on the principle that the When derivative action is combined with proportional
controller should also respond to the rate at which the and integral actions the total response is given by:
measurement is changing-even though the actual error
is still small.
O=#[co++/codj-D,(#)] (5)
The open-loop response in Fig. 7a shows how deriva-
Live response is related to measurement. (The rate of
change may be computed as an amount of change divid- where c, a controlled variable, represents the measure-
ed by the time over which the change takes place.) For ment signal.
example, in Fig. 7a: Eq. (5) describes an ideal, noninteracting controller.
I Point A-A step change appears. Because the In most three-mode controllers, some interaction occurs
change takes place in zero time, its rate is infinite, and among the control modes, so that changing any one of
derivative action responds with an output spike. The the adjustments has some effect on all the responses.
response direction will be determined by the controller
action. Fig. 7a shows the response for increase-increase Applying derivative action
action. Since the measurement is steady after the step Incorporating derivative action can significantly im-
change, the derivative contribution immediately returns prove control for processes having large lags. Derivative
to zero. action is the opposite of integral action. To generate the
I Point B-A second, negative step appears. The de- derivative response, the dynamic inverse of a lag (i.e„ a
rivative contribution responds with a negative spike. lead) is built into the controller. Although derivative
I Point C-The measurement begins increasing at a action also increases the gain of the controller, its lead
constant rate. Derivative responds with a constant, posi- characteristics can effectively cancel a lag elsewhere in
tive contribution that is proportional to the rate of the control loop, and therefore shorten the period of
change. oscillation. This can more than cancel the increase in the
I Point D-The change in the rate of measurement period caused by integral action, even though offset is
undergoes an increase. The derivative contribution in- still eliminated.
creases proportionately. The main disadvantage of derivative action is sensitiv-
I Point E-The measurement stops changing. The ity to noise. Because it reacts to the rate-of-measurement
derivative contribution returns to zero. change, even very low-amplitude noise can cause large
The derivative response is unrelated to the absolute variations in controller output. In effect, the derivative
value of the measurement. Whenever the measurement tries to control the noise-an impossible task.
stops changing, the derivative contribution returns to Since noisy measurements are usually responsive mea-
zero. When it starts to change, derivative action opposes surements, the reduction in the period offered by deriva-
that change whether the measurement is moving away tive action will not be a significant benefit. Hence, deriv-
from or toward the setpoint. ative action should not be applied to noisy loops.
The open-loop response in Fig. 7b shows how propor- Controlled variables that are slow enough to benefit
tional and derivative actions are combined in a control- from derivative action (e.g., temperature) are usually not
ler. When the measurement starts to change, derivative noisy. One exception is the output of sampling analyzers
action generates an immediate response proportional to such as chromatographs. This signal, which changes
its rate of change. As the measurement continues to stepwise, must be filtered before it is applied to a con-
change, the output changes because of proportional troller having derivative action.
action. Because of derivative action, the output immedi- The next article in this CE REFRESHER will appear in
ately reaches a value that it would not have reached until the Sept. 19th issue, and will review principles and pro-
sometime later. cedures for tuning control loops.
In effect, the proportional response has been ad- Steven Danatos , Editor

18
Tuningprocesscontrollers
A review of the basic principles and procedures of controller
tuning will enable engineers to tune a variety of control loops
so as to achieve stability in the loops, and thus the process.

Thomas 8. ELrmey, The Foxboro Co.

I Controller tuning is accomplished by measuring cer-


tain control-loop characteristics. The techniques for
these initial measurements are critical because they are
the basis for subsequent controller settings.
A generalized tuning guide for determining controller
settings requires the use of a model. The model usually
chosen-representative of many processes and their
control systems-is one having a first-order lag with
deadtime.
The control modes that we will consider are the most
frequently encountered combinations of proportional,
integral or reset, and derivative actions. Our discussion
will focus on procedures for determining the final set-
tings for each mode, based on the fundamental process-
control characteristics of capacitance, deadtim`e, and nat-
ural period.
To determine controller settings, two methods-
open-loop step response and closed-loop cycling-are
used to measure their characteristics. The former will
yield the capacitance, 7.1, and deadtime, TOT; the latter,
the natural period, 7o. (See Part 1 of this series* for the
discussion of capacitance and deadtime.)
measurement begins to rise is the deadtime. It can be
Open-loop method calculated by measuring the distance (in.) on the chart
To determine capacitance and deadtime via open-loop and dividing it by the chart speed (in./min).
response (also known as the reaction method), a record- The measurement (Fig. 1) rises to a final value that is
ing device having a fast chart-speed (say, 3/4 in./min) is the new steady-state, which results from the step change
connected to the measurement signal. The test is then made in controller output. From this curve (approximat-
performed by: ing a response having a first-order lag for a single capaci-
1. Placing the recorder in the high-speed mode, with tance system), the time constant, deadtime, and the proc-
the controller in the manual position and the measure- ess-response rate or slope of the control loop can be
ment lined-out at a constant value. measured.
2. Making a step change to the controller's output at The units of measurement for calculating the slope are
some fixed value, such as 5 to 10%; and, at the same the controller settings-usually expressed as percent or
time, making a mark on the recorder chart so that dead- time. The slope of the response curve should be in units
time can be determined. of percent/time, and is expressed as:
3. Removing the chart from the recorder when the
measurement has reached a final value. mE
A typical response from a process having deadtime
and a first-order lag is shown in Fig. 1. The elapsed time where Rjt is response rate, I/min; AM is the change in
from the point where the step change was made until the measurement, %; f is time, min; and AO is the change in
Articles published thus far in this CE REFRESHER: Part 1, May 30, p. 66; Part 2,
output, %.
Aug. 8, p. 79. The controller settings for obtaining a specific closed-

19
Steam

loop response can be predicted from the results of this The final step is to draw a tangent to the maximum
method by using algorithms developed by Ziegler and rate of rise, and lo measure the slope of this line to find
Nichols /I/, Cohen and Coon /2/, Shinskey /3J, et al. the resp()nse rate. The slope is determined as follows:
As an example of the open-loop lnethod, let us con-
sider temperature control of the heat-exchanger process y-axis: #=25%changeininput
sh()wn in Fig. 2a. Assume that the temperature of the
water leaving the exchanger is 100°F, the temperature
transmitter has a span ranging fr()in 0 to 200°F, and the
X-axis: #=0.33min
steam pressure remains constant. A chart recorder is at-
tachcd to the measurement signal. The mezlsuremcnt Since the slope is prt>portional to the size of the output
bcf()re and after the step change w()uld appear ()n the step, the units must be normalized before the slope is
recorder as shown in Fig. 2b.
computed, to account for the percent change made in
the ()utput. Hence, the response rate, fzft, becomes:
For this example, the rcc()rder has a chart speed of :i/4
in./min. At Point A (Fig. 2b), a step change of +20% is 25%/0.33 min
made to the output ()f the controller, and at the same RR- - 3.7/min
20% output
time A mark f()r the measureinent signal is made on the
chart. At Point 8, the measurement begins to rise, and Closed-loop cycling
reaches a final value of 15:OOF. This temperature increase The closed-loop cycling method is popular because
corresponds to the increase in steam flow. The chart only one parameter is measured. Its disadvantage is that
travel through the recorder (after the step change is somc- online processes cannot be allowed to cycle for
made and before the measurement rises) is I/2 in. Since even a short period of time. By causing a control loop to
the chart moves at 3/4 in./min, the deadtime is calculated cycle at a constant amplitude and period, its naturdl pc-
froln: ri()d, 7„ can be determined. An example of a measure-
7.fjr;` = 0.5/().75 = 2/3 min = 40 S ment that is cycling sinusoidally is shown in Fig. 3.
To induce constant-amplitude cycling in a process-
The next step is t,o locate the point at which 63.2% of control I()op, it is necessary to:
the rise* between the initial and final measurements oc- `Step / . Make sure that the loop is in a stable condition.
curs. This corresponds to a temperature of 131.6°F. At S'/c¢ 2. Adjust the integral (I) and/or derivative (D)
this temperature, draw a horizontal line on the chart to modes to minimum aiction if the controller has more
intersect the measurement curve at Point C. Next, drop than one mode (i.e., proportional plus integral, or pro-
down to the line extended from the initial measurement
portional plus integral plus derivative).
to locate Point D. Measure the distance between Points 8 LSTfc¢ 3. Make a step change in the controller's setpoint,
and D (which is I/I in.). Then, the time constant for ca- and observe the resulting measurement cycle.
pacitance is calculated as: S`£¢ 4. Reduce the pr()portional band further if the
7.I = 0.25/0.75 = 0.33 min measurement cycle damps out to a steady-state value,
*Lsee Pal`t 1 ol` this series ((,'/mw J`j`ntJ.. Mzi)I :io, I). 64) for a (lisc`ussioli ol time. and, following this, make another change in the control-
{.t>nst:tnt antl (rit> 6:i.2t/ ixi]nt. ler's setpoint.

ao
S/ep i. Continue repeating Steps 3 and 4 until the
measurement cycles at a constant amplitude and period,
as shown in Fig. 3. Also, be certain that the final actuator
is not oscillating between its limits, because such oscilla-
tion will yield erroneous results.
S/c¢ 6. Measure the peak-to-peak distance from the
recorder's output of the measurement. Convert this
measurement to time by dividing it by the speed of the
chart in the recording device.
As an example of this procedure, consider the heat
exchanger and control loop in Fig. 2a. The temperature
controller (TC) was switched to the automatic mode, and
the proportional band reduced until the measurement
cycled continuously, as represented in Fig. 3. The peak-
to-peak distance is measured as I.65 in. Therefore, the
natural period for this cycling becomes:

1.65 in.
- 2.2 mill sponse will produce a larger total error than QAD but
0.75 in./min
may be acceptable, depending on the particular process
The proportional band that produced the cycling was requirements.
140%, and will be referred to as /PB)*, i.e., (PB)* = If the gain of the controller is increased further, pro-
140%. longed cycling will occur from an upset. This type- of
response is referred to as "underdamped," and results in
Errors in measurement a smaller deviation from the setpoint (see curve in Fig.
A chart recorder or trending device may be used I.o 4b). C()nversely, if the gain is reduced, the response lo an
track the response of the measurc`d variable for a corre- upset will be reduced, resulting in a large dc`viation fr(`tm
sponding change in controller output. Since the control-
ler settings are based on measurements from the record-
ing element, any errors in rec()rding and measuring are
passed directly to those settings.
In addition, adjustments on the controller frequently
have a resolution no better than 15% of the value indi-
cated, because the adjustment dials have large gradua-
tions. Hence, one cannot expect an accuracy of better
than +15% of the value desired for mechanical dials.
Digital controllers, on the other hand, permit exact set-
tings. Therefore, only the errors arising from the ()verall
system used to capture the responses need be considered
in using digital controllers.

Tuning objectives
After measuring the natural period and/or the dead-
time and capacity time-constant, controller settings can
be determined by using relationships developed by
Ziegler and Nichols /I/, Cohen and Coon /2J, Shinskey
/3J, and others. The objective of these investigators was
to determine settings that would minimize the inlegrflted
error over time.
The methods for deriving such relationships are
highly analytical and, therefore, difficult to use. In gen-
eral, the response that results from an optimally tuned
controller will be that of quarter-amplitude damping
/QAD). Fig. 4a is characteristic of QAD. This type of re-
sponse correlates well with the minimized-error re-
sponse, and is practical to use.
Some processes cannot tolerate an oscillation about the
final value, and so another type of response must be cho-
sen. An example of a response curve for a proportional-
only controller, approaching the final value, is shown by
the curve in Fig. 4b, and is called a ``critically damped"
response. If gain were added to this controller, the re-
sponse path would cross the final value of the measure-
ment more than once. A critically damped controller re-

al
setpoint (see curve), and a response referred to as "over-
damped.„
The formulas developed by Ziegler and Nichols /JJ for
predicting controller settings to produce QAD are based
on a process model hav`ing a capacity that is purely inte-
grating. In the example of Fig. 5, the level in the tank
corresponds to the integrated value of flow. If a change
in inflow occurs and the outflow remains constant, the
tank will either empty or overflow. The steady-state gain,
Gss, of this process is infinity, and the process is said to be
non-selfregulating.
If the outflow from the tank is affected by changes in •.. .. --.®

trie inflow, the level in the tank will likely reach a steady -./,.----

state if the inflow upset is not too large. This type of


2 4 6 8 10
process response is said to be selfregulating. Cohen and
Coon /2J developed relationships for predicting control- Patio of deadtime to Capacity, 7-D7-/7-1

ler settings to account for selfregulation. a. Chart for natural period, 7.a
However, it is recommended that the Ziegler and
Nichols relationships be used rather than those of Cohen
and Coon, unless the ratio for 7D7/7-1 becomes greater
than 0.1.
Procedures and guidelines for tuning the common
combinations of proportional, integral and derivative where (PB)* is the proportional-band setting that pro-
modes, along with criteria for their evaluation, will fol- duces constant-amplitude cycling.
low. The analysis of selected settings is required to com-
Open loop
pensate for errors in measurement and adjustments. In
this respect, the procedures may be considered as an iter- 1001 _

ative approach. in _ Tf f iR (3)

Proportional-only mode Method: Cohen and Coon


The proportional-only controller flnds application in
processes that require a fast response and that, at the
same time, can tolerate a constant deviation from the set- (4)
point. The amount of this deviation is a function of pro-
portional band and bias. where pr = TD7/Ti.
The proportional-only controller has one adjustment If QAD is not desired, an increase in the proportional
for tuning. Therefore, QAD is an acceptable criterion. band will result in critical damping; a further increase
The recommended settings are: will produce overdamping. Decreasing the proportional
band from the QAD setting will create underdamping.
Method: Ziegler and Nichols, and Shinskey
Closed loop Proportional-plus-integral mode
The proportional-plus-integral (PI) controller is prob-
PB -2(PBt* (2) ably the one most often encountered. Its advantage is
fast response and zero deviation from the setpoint at
steady state. The tuning procedure for a PI controller is
somewhat more difficult to evaluate because two adjust-
ments exist, and many combinations of these will pro-
duce QAD. Therefore, other criteria are necessary to
evaluate the predicted controller settings.
Shinskey /jJ has shown that the damped period of a
properly tuned PI controller will be approximately 1.5To.
For processes in which the natural period, 7.a, is difficult
to determine, the value for 7-1 and 7-Dr can be determined
by the open-loop method; and the natural frequency, 7-a,
approximated from Fig. 6a.
The response of a PI controller having a period equal
to 1.57.a, and for QAD, is shown in Fig. 6b. By increasing
the integral action (Fig. 6c) in the controller, the damped
period will increase, and oscillation about the final value
will be longer. (The damped period of a proportional-
plus-integral controller will be referred to as Tpr, and
equals 1.5To.) Decreasing the integral action will produce
a response that will take longer to return to the setpoint.

8a
The recommended settings based on the measure- Open loop
ment of 7o, or 7.I and 7D7`, or both are:
100 0.9 0.9
- 0.363
Method: Ziegler and Nichols TEE~=fiF=T=-F5])(o:-61-)
Closed loop
PB -T75%
PB--2(PBt* / = 3.33 7-DT = 2.23 min
I -TJ1.2

-=---
Using the relationships of Eq. (7) yields:
where J is the reset time, min.
Open loop o.9(I+#) o.9(l+0.18,
100
- 0.428
loo 0.9 PB TDTRR 0.&] (3.])
PB TDTRR pB = 233cmo
I = 8.38 TDT

Method: Cohen and Coon I -3.3 TDT - 3.3(0.67) - 0.48 min

1.. (1+#)
(7a)
PB TDTRR Substituting into Eq. (8) produces:
PB = 2(PB)* = 2(140) = 280%
I -8.88 TD-I (7b) / = 0.43 7-o = 0.43(2.2) = 0.95 min

The predicted Cohen-and-Coon setting results in a


higher controller gain [where G = loo/PB], because their
Method: Shinskey
equations contain a factor to account for selfregulation.
PB -2(PB)* The Ziegler and Nichols methods make no provision for
/ - 0.43 7-a this characteristic. Settings predicted by the Shinskey
method result from a slightly different error-analysis
Using the example of the heat exchanger (Fig. 2) and approach, and are close to those of Ziegler and Nichols
the results from the open-loop and closed-loop tests, the for the closed-loop test. Errors in measurement between
settings will be determined by each of these methods for the open-loop and closed-loop tests contribute to slightly
a proportional-plus-integral controller where 7-D7` = 0.67
different predicted settings.
min, To = 2.2 min, 7-1 = 0.33 min, fizz = 3.7/min, and
(PB)* - ,40%. Proportional, integral and derivative modes
Using the Ziegler and Nichols relationships, Eq. (5) The three-mode (PID) controller cannot be used on a
and (6), the proportional band, PB, and reset time, J, for noisy measurement, or on one that changes stepwise,
closed-loop and open-loop responses are calculated as: because the derivative contribution is based on the meas-
Closed loop urement rate-of-change.
The PID controller is used on processes that are slow
PB = 2(PB)* = 2(140) = 280% to respond and have long periods. Temperature control
/ = 7-/I.2 = 1.83 min is a common application where the heat rate may have to

a3
change rapidly when the temperature measurement
begins to change. Derivative action shortens the response
period to an upset.
Due to the physical construction of most controllers,
an interaction occurs between the integral and derivative
modes. This interaction causes the effective values of the
modes to differ from their set values. The effective inte-
gral time, J,(,/T.F is actually the sum of two time constants:
It(ef f.) -It + D, (9)

The effective derivative time, D,(,,/i.), is:

Dt(elf.,-
11

I, + ir,
Two important points concerning Eq. (9) and (10) are:
1. The effective value for derivative time can never be
greater than one-fourth the effective integral time,
which occurs when D, = I,.
2. When D, is larger than J„ the contribution to each
control action is reversed. In other words, when setting
D, greater than J„ this changes the value for I, more than
Typical responses of a PID controller for several inte-
for Dt(eff.).
The rule-of-thumb is to never adjust a controller so gral and derivative times are shown in Fig. 7.
that derivative action is greater than integral action. Summary
The performance criteria for a PID controller can be Open-loop and closed.-loop methods can be used to
evaluated by measuring t,he damped period. Optimum measure deadtime and capacity, or the natural period, of
tuning generally results with a QAD-period that is ap- a control loop. In turn, these measurements can be used
proximately equal to the natural period. The damped to predict controller settings via the methods presented
period will be referred to as 7.p+D, and is equal to 7o. here, depending on the degree of selfregulation of the
Recommendations for response settings are: controller.
Method: Ziegler and Nichols In most cases, the desired response should be quarter-
Open loop amplitude damping, which is a practical measure of the
minimum absolute integrated-error. In the case of a PI
loo I.2 or PID controller, the shift in the period of the damped
ir _ TfrR response should be considered, as well as quarter-ampli-
I -2.0 TDT tude damping.
The next article in this CE REFRESHER will appear in
D -0.5 TDT
the Nov. 14 issue, and will cover advanced techniques for
Closed loop feedback control.
Steven Danatos, Editor
PB -+.66(PB)*
I -0.5 To
D - I(J8 References
I. Ziegler, I. G., and Nichols, N. 8., Opti'mum Settings for Automatic Controllers,
Method: Cohen and Coon rrc}us. ASME, Vol. 64, p. 759, American Soc. of Mechanical F.ngineers, New York,
(1942).
2. Cohen, G. H., and Coon, G. A., "Theoretical Considerations of Retarcled Con-
trol," Bull. No. TDS-10A102, Taylor Instrument Co., Rochester, N.Y.
3. Shinskey, F. G„ "Process Control Systems," 2nd ed., Mt`Graw-Hill, New York,
(13a) 1979.

The author
I -2.5 TDT Th'omas 8. Kinney is a systems applii`tltions
engineer at The Foxboro Co., Foxboro,
MA 02035. Currently, he is involved with
the design of control slrategies for process

D-H 0.37 7Dr


units-primarily in the chemical process
industries. the startup and commissioning
of these units, and the debugging of
existing process-control systems. His
experience includes design and
programming activities for large digital-
Method: Shinskey computer projects serving batch processes.
He has a B.S. in chemical engineering
PB--4.O(PB)* from the Unlversity of Maine, and is a
member of AIchE and the Instrument
I -0.5 To Soc. of America.

D -0.12 7.a

84
Feedbackmethodsfor
processcontrolsystems
Special feedback techniques provide stability and operability
to processes and their associated control loops whenever nonlinear
characteristics are present in the measured or sampled variables.

Thomas I. Myroi., ]r., The Fohao Co.

I Feedback control can be implemented via a number


of techniques. In Part 2 of this series,* a change in the TOT = i (TOT)»
measurement value, or an error signal, was shown as (2a)
„-I
being the basic input to the controller for processes hav-
TD7` = (TD7`)p + A! (2b)
ing reasonably Hnear characteristics. Here, we will ex-
plore some advanced techniques for feedback control. where 7o is the natural period, min; 7.DT is deadtime, min;
The input signal to the controller may be derived from
(TD7`)p is process deadtime, min; and AC is the sampling
sample data, ratio control, or the nonlinear characteris- Interval for the feedback measurement, min.
tics of the process system. A description of each tech- Substituting Eq. (2b) into Eq. (I) yields:
nique will indicate its applications:
I Sanpde deto--The process is dominated by dead- To = 4[(7.D7`)p + AC]
(3)
time; or a control variable is measured by a device that When a process has very little deadtime:
supplies an intermittent output to the control system,
e.g., a process chromatograph. (TDT)p < Lt
I fzafo-The process to be controlled is affected by
Then: To = 4 Aj
the ratio between one variable and (at least) one other
variable. The second variable can be either separately In this case, Eq. (4b) indicates that the natural period
controlled or what is termed "wild." of oscillation is dominated by the sampling time of the
I Ivnd.aetzr-The process has a highly nonlinear feedback measurement.
characteristic such that the process gain can significantly -An inline blending process is shown in Fig.I. Here, an
change, as a function of either load or setpoint. The pH additiveisblendedwithamainfluidinaliquid-fullpipe-
process exhibits such characteristics. line. Continuous control is performed by using a dis-
continuous measurement. The analyzer has a 5-min
Sample data control sampling time (Al = 5) and is connected to a proper-
The effect of using a sampled measurement having a tionalonly (P) controller. For this example, it will be as-
time interval, Aj, is to introduce another deadtime ele- sumed that the capacitive time constant is essentially
ment into the control lcop. Multiple deadtime elements Zei°f'fi::qveTy?ttifi:Cies(aTDPT|Pre<dAe`;££:r!Pp=rfessunderp-
in a loop are additive-i.e., five I-min elements are
equivalent to one 5-min element. From the viewpoint of only control, and the loop will oscillate, so that:
closed-loop behavior, an observer will not be able to dis-
7o = 2TDr (4c)
cem the individual characteristics of each deadtime ele-
ment-nly the additive effect. As was shown in Part 1 Since this process is dominated by the sampling inter-
(Cho. E"g., May 30, pp. 62-64), a typical process having val for the analyzer, Eq. (5a) can be written as:
hath deadtime and capacity will oscillate, such that:
To = 2 A! (4d)
To = 4TDr (1)
Fig. 2 illustrates the behavior of this loop with propor-
tional-band settings of 100, 200 and 400% for a change
Au*gt#pF7P9?¥rtT¥,i¥:us[.LiT9Tp:i::7CEREFRESHER:Partl,May3o,p66;part2,
in the setpoint. For example, if the setpoint were

8e
changed 10% (i.e., increased from 50 to 60%), the proc-
ess having the 100% proportional-band (PB) controller
(Fig. 2a) would have a loop gain of 1.0, and never settle
out. With the controller set at 200 or 400% PB (Fig. 2b
and 2c), the cont.roller loop would be stable, but the
measurement would settle out at 55% and 52.5%, respec-
tively. A high price, at least in terms of settling-time off-
set, has been incurred to achieve stable control.
Ideally, the measurement should come to setpoint in
one sample period (AC). Since measurement is not equal
to setpoint when P-only control is used (except for the
one load condition where the manual bias was set to
make the measurement equal to the setpoint), the addi-
tion of integral (I) control action is necessary to remove
the offset.
Fig. 3 illustrates the behavior of integral-only control
in the process of Fig. 1 to a setpoint change for different
integral-controller settings as related to the sampling
time. (Remember that a pure deadtime process under
a lonly control will oscillate, so that To = 4 7-D7`.)
1`E`60 C13cO50~Ia,E9I§=ag56oIaC1£5oEa'Ee!5I When I = A£, the ideal response is achieved without
the need for adding proportional action. However, the
dynamic characteristic of the process shown in Fig. I was
idealized by eliminating capacity and real process dead-
time. Should a process exhibit the assumed characteris-
Firal tics, the best control would be I-only, where J = AC.
The responses in these examples were initiated via a
twit-ed
setpoint disturbance. Had they been initiated by a posi-
tive (increasing) load disturbance, the results would have
been the same. A negative (decreasing) load disturbance
Sampling interva 1' Af, min would have caused the offset (if any) to appear on the
a. Proport ional band = 10 0% other side of the setpoint.
A generalized feedback-control example for a process
similar to that of Fig. I is shown in Fig. 4a. Here, a
+AJ steady-state gain, Kp, has been included, and the process
deadtime, (TD7-)p, is significant but less than the sampling
^aife =-5R---
interval, AC.
Fig. 4b illustrates the closed-loop responses of the
process. These are somewhat similar to those illustrated
-i- in Fig. 3. However, recovery in one Al (i.e., the sampling
time) is not realized for the conditions indicated. The
addition of process deadtime has changed the process
characteristics.
The question now arises whether a proportional plus
Sampling interval, Af, min integral (PI) controller can be used to permit the meas-
b. Proportional band = 200% urement to reach setpoint after one AC when additional
process deadtime, (7.D7`),, that is less than the sampling
time Al is present. When a setpoint change is introduced
to a process such as that shown in Fig. 4a, it is known that

z,y`
SEtpoint*1*-
-H t
"i
+i+-+++i+:+

ire a ¥£-*-
an error, a, results, such that:
Ac= Ar-Ac (5)

where r is the setpoint, and c is the controlled variable


IIi,
(measurement).
`\`\ i Starting at the controller output in Fig. 4a, it is known
thsureme+ that a change in controller output, Am, produces a
+

change in the measurement, Ac, such that:


A` --Kp ^m u6)
Sampling interval, Af, min
c. Proportional band = 400% The gain of the deadtime block is not included in
Eq. (6) because the steady-state gain of any pure dead-
oed-loop+respo etto changes in the+ ++ time element is unity.
setryeint fee ifiliae bFendfro9 pr`obkein, .A Fig,J2
For a two-mode, proportional + integral, controller, a

8e
XaeL1I,+ -0C|j* iLur-t #*yi
i
change in the controller output, Am, can be expressed in +£ri
terms of its proportional and integral settings as: I\ar

Am=(ig)Ae+(ig)(f)A£ (7)
S

Solving Eq. (6) for Arm,\ and setting the result equal to
Eq. (7), yields: S
+

2£+ Ia'E®LIt® 0=
ifl¥, r±---_peifife,r3[-`}
Ac=Kp[Jg+(ig)(f)AJ] (8) I ¥1I`1I
#£ J-i# >

>
}

Eq. (8) can be further modified when the following are


considered:
I. The c term in Eq. (8) is in reality Ac because all of
the error occurs when the setpoint change is made, i.e.,
e - he.
Ti in n
2. The Af term is the sample time of the analyzer.
However, the actual time is Af less the process deadtime, I * *

or.. ^1 - (TDT)P. f*r the i#II #€ end'ng H* IS FIg.3


Chancing Eq. (8) to reflect these adjustments yields:

Ac = Kp[(ig) Ac + (ig) Ae( A` -{TDT)¢)] (9)

Factoring out the Ae terms produces:

L±-Kp(i#)+
Ac

(¥)(ig) A, -(¥)(ig)(TOT,A (1o,


The objective is to have Ac = Ac in one sample period.
By letting I = (TDr)p, Eq. (10) becomes:

I-(#)(i#)A,
Solving for the proportional band, PB:
A'
PB--100Kp
(TDT)P
and: I - (TDT)a

ratio control
Ratio control is the simplest form of feed forward con-
trol in that a load variable (the wild flow) is used to calcu-
late the setpoint of another control loop. For the most
part, ratio control is primarily concerned with the ratio
of one flowing stre-am (gas, liquid or solid, or their com-
binations) with respect to another.
Ratio control can be applied in a manual-set mode
where the user fixes the ratio of one stream with respect
to another, or in a variable mode where the ratio is con-
tinuously adjusted-usually via a feedback loop.
Variable-ratio control will not be discussed in detail
here because it is better handled under feed forward con-
trol, where a know`ledge of instrument scaling is re-
quired. Variable-ratio control is applied when some
property of the process or process inputs is not constant.
In such cases, a manual setting will dive an inconsistent
ratio between the controlled and wild variables. For ex-
ample, the ratio of reboiler heat input to column feed
flow can be manually set. However, if the feed composi-
tion changes significantly (assuming no feed analyzer is
available) or fouling alters the heat-transfer characteris-
tics of the reboiler, the ratio must be increased or de-

87
With Eq. ( 18), a nonlinear division is replaced by a lin-
ear (constant gain) multiplication. Fig. 5b illustrates the
preferred control arrangement.
The "R" (ratio) and "FC" (flow controller) blocks of
Fig. 5b are normally contained in one piece of hardware
called a ratio flow controller. Typical ranges for the ra-
tios available in ratio controllers are: 0 to I.0, 0 to 3.0,
and 0.6 to 1.3. The range chosen is usually based upon
the application.
There is no limit to the ratio range that could be used.
In practice, it should be remembered that a ratio value is
essentially a "gain," and that the higher the ratio value,
the more sensitive the setpoint change becomes to
changes in the flow signal of the wild stream.
For a given application, the principal ratio factor is
handled at the transmitter level. If a ten-to-one ratio is
desired, the transmitters are selected so as to have a ten-
to-one difference in their flow ranges. This choice allows
the signal levels of each transmitter to be about equal, as
the actual flowrates vary from 0 to 10097o. The ratio set-
ting on the instrument faceplate is easily determined, as
illustrated by the following example.
Let: FA=Oto logpm= 10F'A (A)
FB = 0 to loo gpm= 100F'B (8)
R-Otoo.1 -0.1R' (C)
where F'A, F'B and R' are the percentage instrument-sig-
nal values, expressed as decimals.
Since A = FA/FB = 0.I, the equation to be solved is:
F'A -0.lF'B (D)

creased, depending upon, the changes. In this instance, a Substituting Eq. (A) and (8) into Eq. (D) gives the flow
temperature controller ih a stripping section of the col- relationships in terms of the instrument signals, or:
umn could be used to continuously adjust (trim) the F'A = 1.OF'B (E)
ratio.
In ratio control, the controlled variable is in reality the Since the coefficient of the F'B term is I.0, an instru-
ratio, fz, of one variable to another. For example: ment range would be selected to include the ratio value
of 1.0. Thus, any of the ratio ranges previously men-
R = FAIFB (\4) tioned could be used.
where FA is the flowrate of Material A, and FB is the The range 0 to I.0 might be considered if the user
flowrate of material 8. wanted to ensure that the flow, FA, never exceeded
0.1FB, but could go lower. Choosing 0 to 3.0 provides a
Nonlinear characteristics of ratio control wide range of ratios about the normal setting of I.0. The
Fig. 5a illustrates a ratio flow process where the ratio, 0.6 to 1.3 range would provide a high-resolution adjust-
fz, is the controlled variable. An examination of the proc- ment about the normal setting.
ess-gain characteristics (assuming FA is varied to maintain If the range of FA in Eq. (A) had been 0 to 15 gpm,
J3) shows:
Eq. (E) would become:

F'A = 0.67F'B (F)


R = FA(\lFB) (±5)
dRldFA = +/FB (+6) And, the instrument ratio setting for Eq. (F) would then
be 0.67.
Chanting the manipulated variable from FA to FB so
If a calculated ratio setting falls outside the ratio-ad-
that FA is the wild variable in Fig. 5a yields:
justment capabilities of the hardware (or software), the
dRldFB ---FjJ(FB)2 (17) user must reexamine the ranges of the transmitters, and
readjust them so as to bring the calculated ranges within
Eq. ( 16) and ( 17) illustrate the highly nonlinear nature
the span of the available adjustments.
of the ratio process when A is controlled directly. In this
arrangement, the loops would have to be tuned for the Nonlinear control
worst case (i.e., low flowrates). This would result in slug-
Some typical process-gain characteristics are shown in
gish control and higher loads. The problem is eliminated Fig. 6. In a linear process (Fig. 6a), the gain is constant no
by rearranging Eq. (14) to:
matter where the control point is set. Assuming no other
FA -RFB (18) nonlinear elements in the control loops, a controller

a8
tuned at one operating setpoint will remain stable over C, the result is a very low process gain, and a controller
the entire range of operating setpoints. having a very low gain will cause the process to hang at
Fig. 6b represents a moderately nonlinear process, Point A or C. The controller output eventually adds
e.g., the change in slope (gain) is equal to or less than 4 to enough reagent to cause the measurement to overshoot
I when the manipulated variable varies from 0 to 100%. Point 8. This sequence of events usually repeats itself
The process gain is a function of the operating point. A indefinitely; and if recorded on a chart, the area between
controller tuned at Point A would behave in a more slug- Points A and C rapidly f"s with many lines.
gish manner if the setpoint were moved to Point 8. Simi- A pH process described by Shinskey /I/ is shown in
larly, a controller tuned at Point 8 would be more re- Fig. 7. Here, a strong-acid/strong-base neutralization is
sponsive if the process operation were changed to Point being controlled to a setpoint for the neutral value of
A; and in the extreme could become marginally stable pH = 7.0. Shinskey calculates the controller gain, G„ for
(sustained oscillation) or even unstable. this process as 0.033; and the proportional band, PB =
Fig. 6b is typical of many thermal-type processes. The loo/(;„ as 3,030.*
simplest way to overcome moderate nonlinearity of such Since most industrial controllers do not have a propor-
processes is to include another element in the loop, hav- tional-band adjustment above about I,000, any attempt
ing characteristics opposite those of the process-gain to tune a loop similar to the one in this example will be
characteristic. Such an element is the equal-percentage ineffective. In fact, it might make one assume that the
valve whose characteristic is opposite to that of the pr()c- *For additional inl`ormation on gain, see Part I (C'he»t. I"g., May 30, pp. 65-66)
ess shown in Fig. 6b. The resulting combination of valve ai`d Shinskey /2/.

plus process has decreased the nonlinear characteristic


for the system. If the match between the valve and proc-
ess were perfect, the resultant characteristic would be
completely linear. Deadtime
If a final operator having a linear input/output rela- 7.o7- = 10 s
® Valve: linear characteristic;
tionship is used for a process such as that represented in flow is 0.05 gpm at Bange:
Fig. 6b, a signal characterizer (having the opposite char- + iooo/oof stroke PHTLJ -2 to 12 pHI-
acterization of the process) could be installed in the out- 1 0 a/o N a 0 H -,,
put of the controller. This would result in an overall lin-
ear characteristic.
Fig. 6c is typical of a highly nonlinear process, e.g.,
gain change is greater than 10 to 1. The pH process
(acid-base neutrali7' tion) is typical. Reflux-temperature
1 cO gpm
T- I,,,

loops also have a similar characteristic when the boiling- v=`8 2cOa', •^,

point difference between two components in a linear


mixture is very large, such as with methanol and ethyl-
ene glycol.
Invariably, in this type of process, the control point is
at Point 8 (e.g., for neutralization, pH = 7). The process +=`

gain is very high, which means that a very low controller


gain is required to maintain loop stability. However, if
load disturbance moves the measurement to Point A or

a9
For example, if (PB)D = 300%, and the slope = 0.1,

/ (PB)E` = 3,000%. Outside of the deadband, the propor-


tional band of the controller is (PB)a.

Breakpoint Gap action via a nonlinear controller


I

tt Adjustable
A nonlinear controller functions as a gap-action float-
ing controller when its output has a slope of zero. In
: '#:rd ;r;v;:,`:/f/ Fig. 9, flow to a process is sup|)lied from two pipelines; a
\``,
`\ \ ./ constant or base load is flowing through the large valve,
and a manipulated amount through the small valve.
\\` F/be b@ndJ As long as the output from the controller, XIC, is be-
tween 30 and 70%, the small valve is able to control the
\/ process. If the small valve tries to open more than 70% or
close less than 3097o, the base load needs readjustment.
'/^T#tn#r The deadband of the nonlinear controller, VPC, would
be set at ±30% with a slope of zero. The setpoint of VPC
could be any value within the deadband (say, 50%) and
its measurement is the output of XIC. The major short-
coming of gap-action control is that the measurement
Characteristics of a nonlinear controller Fig. 8
tends to hang at or near the edge of the deadband.
Other applications of gap-action control, using a non-
linear controller, include surge-tank level control, and
controller was inoperative once the pr()portional-band control of processes having measurement noise. For
adjustment had been set at its limit. surge-tank level control, the surge vessel will absorb the
In order to handle a process having characteristics inflow and not pass it to the downstream process as long
similar to those in Fig. 6 or the pH exainple, a contr()ller as the level is between 20 and 80% of measured height.
with a characteristic opposite that of the process is re- Here, a small amount of gain within the deadband might
quired. This is illustrated in Fig. 8, where the dashed be considered so as to slowly pass the surge to a down-
diagonal line represents a linedr control characteristic. stream process.
The two adjustments available to the user are the In some instances, measurement noise in the signals
deadband width and the slope of the line within the such as that caused by pulsations or mixing may be in
deadband, as indicated schematically in Fig. 8. The slope excess of say ±5%. It may not be desirable to have a final
varies from 0 to I.0. At zero slope, the line between the operator respond to this noise via the proportional action
breakpoints is horizontal. At a slope of I.0, the nonlinear of the controller. Here, a nonlinear controller with a
characteristic is completely removed, and the controller deadband equal to the noise band of the measurement
becomes a converitional linear controller, as shown by czin be used to steady the controller's output.
the dashed diagonal line in Fig. 8.
The effective proportional band, (PB)F, within the Coming Soon
deadband is: The next article in this CE REFRESHER will appear in
the issue of Feb. 6,1984, and will cover the direct synthe-
(P13)E = (PB)DIstope sis controller and adaptive control.
where (PB)D is the setting for the proportional band on Stevon Danatos, Edtw
the controller dial.

References
I.§r#:recyy'E.e:..i:,Pr¥,a,'Sj3P[°NC°ntrolinpTt>cessandwastestreams,"chapter

2. i3;n6:key, F. G„ "Prtx:ess Ct>ntrol Systelns," 2nd ed., MCGraw-Hill. New york,

The author
`pEac:ac=ka#ayg=mni±T.:'idm`3[n=?.gilroof
systems design a[ The Foxboro Co., 38
Neponset Ave., r`oxboro, MA 0`2035. Since
jao:i,£g,a:::s;,:n:+u::ee;r:%vo:;e.:;;:¥hne:nfo:n

;is:t&;c:a:|d,:y:s::sE;e::nec:he:nhrg!?:::g::rTnd
a::yT::-i'n?cnE;ns:i::it,:5:`:imhaRse:IS,Sde::erarken
graduate study in enchneering management
at Northeastern University.

30
Direct-synthesis
and adaptive controls
A model consisting of the steady-state gain, deadtime and lag for an
actual process is the basis for such control systems. Damping and speed
of response of the controlled variable provide the tuning adjustments.

Paul C. Badavas, The Foxboro Co.

I Direct-synthesis controllers (I)SC) and adaptive The flow loop usually responds much faster than the
direct-synthesis controllers (AI)SC) can be used in com- temperature loop. Hence, a standard PI controller is
position, temperature and vapor-pressure loops. They sufficient.
are also effective for plug-flow processes and solids- The outlet temperature of the exchanger responds
transportation loops that tend to be dominated by dead- slowly to changes in steam flow because it takes time to
time. And they can be readily used as feedback-trim overcome the lags associated with transfer of heat to the
controllers in feed forward control schemes.
product stream.
Direct-synthesis controllers provide a means for quan- Also, the rate of product flow affects the residence
tifying process information in a systematic and relatively time of the loop because at a given flow it takes a certain
simple way. amount of time to displace the product volume in the
Adaptive direct-synthesis controllers provide an op- tubes of the exchanger.
portunity to improve control further by adapting the Before designing a direct-synthesis controller for the
model parameters on the basis of the measured process outlet temperature, we must first obtain a process model
variables. This employs additional information to char- for the loop that quantitifies the variation in outlet
acterize process gain and dynamics. temperature with steam flow for a given product flow
We will discuss the basic design of a DSC by using a before specifying its desired response.
heat-exchange process. This will also be used to show
how an ADS(`, can be designed by adapting the parame- Process model
ters of the controller from the process variables. With the temperature controller on manual, or the
The direct-synthesis controller is designed by devel- flc`w controller on local set, a step change is made to the
oping a process model to achieve a desired response for flow controller, and the response of the outlet tempera-
a controlled variable. In most cases, the parameters for lure, r2, is observed. During the response time for this
the process model vary as a function of the measured experiment, it is assumed that the load variable remains
variables. The latter can then be used to continuously relatively constant.
adapt the model parameters to further improve the Fig.1b shows a typical response. The time it takes for
response of the controlled variable to load upsets. the temperature to respond after a step change is initiat-
ed is referred to as the deadtime, 7.D, , of the process.
Design of direct-synthesis controller The time it takes for the temperature to reach 63.2% of
In the heat-exchange process of Fig.1a, the objective its final value from its starting point, but excluding
deadtime, is defined as the lag of the process, 7.1. The
:sn:oaf::t,:Teetper#::'t:::,:Tr;:,:::,tnegm;te:a::::;:',5), steady-state gain of the loop, K, is obtained by dividing
manipulating steam flow, Ttr . A typical method is cas- the change in temperature by the change in steam flow,
cade control. Here., the temperature controller, TC, pro- Or:
vides the setpoint for the flow controller, FC. The flow
controller [a standard proportional-plus-integral (PI) de- K = AT2/AW\ (I)
vice] provides the demanded steam flow, regardless of where Ar2 is the steady-state change in outlet tempera-
steam-supply pressure and steam-valve characteristics. ture, and AW'\ is the change in steam flow, as shown in
Fig.1b.
pd:tit:,"t\'LL,;PBL`b,I:)`;I.;t'pth7L:)",;`c'H:'`3[[;I:p(tr,t?Ci|`;8`:'jTC;j?,;I,[p'd„\`+i"\:i,(:'`i:?i'983:33`
i).`3.I)
Steady-state gain, deadtime and lag constitute the

31
Undamped,'-±f//,a:cn,ii:tri:n I
Und8rdasetpoint1\/\/\mped \ \ \ \ I_\1
/
LNa)LI+COLaJC|

EaJ J+aJ
ii

iI./
/
\
\ \ \ \ \ \ •cri
£
I tIBed-J

+Ia
tic lly dam ed
overdaDampin

g ratio = A2/A1

Time-
Damping ratio affects the
close d-loop response Fig. 2

process model that is needed for designing the direct-


synthesis controller for the outlet temperature.

Closedlloop response
Desired outlet-temperature responses to setpoint
changes when the temperature is under closed-loop
control (i.e., temperature controller on automatic) are
shown in Fig. 1c. Since the deadtime of the process
cannot be speeded up or overcome, deadtime for the
desired closed-loop response is set equal to the deadtime
of the process.
The steady-state gain of the desired closed-loop re-
sponse must equal 1. This guarantees that the tempera-
ture is regulated at the desired setpoint, and that it
follows setpoint and, hence, load changes without offset.
In other words, the controlled variable returns to the
setpoint in the steady state, following load upsets or
setpoint changes.
The ratio of the desired closed-loop lag to process lag,
7-/7.I, is used to speed up the response. As shown in Fig.
Ic, the faster the desired response, the smaller is the
ratio 7./7.1 .
Possible responses for any control loop are shown in
Fig. 2. For an overdamped response, the temperature
slowly approaches but does not exceed the setpoint. A
critically damped response means that the temperature
approaches as quickly as possible but does not overshoot
the setpoint. Although an underdamped response exhib-
its cyclic behavior whose period is 7.o, the magnitude of
the periodic response decreases with time and, thus, the
loop remains stable. On the other hand, the under-
damped response has a uniform oscillation of constant
amplitude and period.
For the underdamped response in Fig. 1c, the ampli-
tude of the second peak divided by that for the first peak
is termed the damping ratio, €, or:

` -42/41 (2)
The smaller the damping ratio, the more damped the
response.
The model for the DSC is used in a complementary

38
Process has gain and pure deadtime

I Deadtime, 7-07-, ls the same for process and model


Stable for steady-state gain, K, of :

K>0.5Ka
I Gain, K, is the same for process and model
Stable for deadtlme, 7-07-, Of :

0.57ora<707-<1.257ora

Process has gain, deadtime and lag

(Deadtlme equal to 10 times lag)

I Deadtime, 7-07-, and lag, 7.1 , are the same for process and model

Stable for steady-state gain, K, of:

K>0.49Ka
I Gain, K, and lag, 7., , are the same for process and model
Stable for deadtlme, 7-07-, Of :

0.41TDTa<TDT<1.32TDTa

Note The subscript a indicates the actual process parameter.

way. For example, the controller gain, K'c, is the inverse


of the process gain, K (i.e., Kc = 1/K). Increasing or
decreasing the controller gain above the 1/K value pro-
duces the responses shown in Fig. 2. In essence, the
desired degree of damping is achieved by multiplying
the controller gain by a damping gain, KD. Controller
gain is then redefined as:
Kc-KD/K (&)
Damping gain is used to set the desired damping for
the response. For underdamped responses, it sets the
desired damping ratio. Of course, the damping gain can
be set low so as to obtain overdamped responses that
have no overshoot.
By varying the desired closed-loop lag 7-, the desired
period of oscillation, 7-o, is achieved. As the ratio of the
desired closed-loop lag to the process lag, 7-I, is made
smaller, the faster will be the response and the smaller
will be the period of oscillation for the response 7-o.

Example: deadtime plus lag process


A DSC was applied to a process having a I-min dead-
time,1-min lag, and unity gain (i.e., 7-DT = 1 min, 7-1 = I
min, and K = 1.0). Load upset to the process was added can be achieved by varying the damping gain and the
to the controller's output. The effects of the upset were 7-/7-I ratio. Of course, the process model for the simula-
recorded, and the responses are shown in Fig. 3. tion was exactly known. The following questions now
Speed of response as a function of the 7-/7-I ratio is arise: How accurate must the model be? and, What are.
shown in Fig. 3a. Process lag is I.0 min. As this ratio is the bounds of stability as far as the model parameters are
decreased from I.0 to 0.I, closed-loop response to the concerned?
load upset reaches the setpoint of 507o faster and faster.
In this case, the damping gain was set to KD = 1.0, and Parameter sensitivity
all responses were overdamped. The theoretical background for the DSC is given in Ref.
In order to obtain underdamped responses, damping I, as are the stability limits-a summary of which are
gain was increased to KD = 1.75. Speed of response and given in Table I. For a pure deadtime process (the most
period of oscillation, 7.o, as shown in Fig. 3b, decrease as difficult to control), model deadtime can be as low as
the ratio 7-/7-1 decreases. For 7-/7-1 = 1.0, 7-o = 3.I min, 507o below, and as high as 257o above, the actual
while for 7-/7-1 = 0.25, 7-o = 2.2 min.
process deadtime before uniform oscillation occurs. As
Responses to variable damping gain, KD , are shown in the ratio of the process deadtime to lag increases, these
Fig. 3c. As KD increases, the response becomes less stability bounds are relaxed even more.
damped, i.e., the damping ratio also increases. When the deadtime is known exactly, uniform oscilla-
This example shows that a family of desired responses tion occurs when the model gain is 507o that of the actual

33
Outlet-temperature,
r2 , setpoint

Direct-synthesis
controller, Dsc

Desired response:
• Damping gain, Ko
• Speed-of-response
ratio, 7./7.1

Process model :
• Steady-state gain, K
• Deacltime, 7.07-
• Lag, 71

a. Process b. Controller

Main elements of a process and model for the direct-synthesis controller Fig. 4

gain of the pure deadtime process. This implies that to be used, the tuning parameters for the DSC would still
reasonable estimates of the model parameters will pro- be desired damping and speed of response.
vide stable control. The equation for the controller [J] is:

Design summary M n --BM T,_\ + (I ~ B)M T,_I_\ + K\(e n- CLe „_\) (4)

The main elements of the direct-synthesis controller where K]=(I -6)/[K(1 -a], a=exp(-I/7.1),
for the outlet temperature of the heat exchanger of Fig. I A = exp(-I/7.), c = r -c, r = setpoint, c = controlled
are: (I) the process model, consisting of the steady-state variable, jw = manipulated variable, I = sampling time,
gain K, deadtime 7-D7`, and the lag 7-1; and (2) the desired K = gain of the process model, 7.1 = lag of the process
response, as chosen by the damping gain KD, and the model, 7-,„ = deadtime of the process model, 7. = lag
desired ratio for deadtime to process lag 7-/7.I. These for the desired closed-loop response, IV = nearest inte-
procedures are summarized in Fig. 4a. ger of 7.or/r, 7t = current sample number, "-IV-
Fig. 4b shows the direct-synthesis controller that con- 1 = sample number (IV + I) sample times ago.
tains the model of the process implicitly, i.e., the control- Microprocessor-based shared controllers have a rep-
ler complements the process. Fig. 4b also shows sche- ertoire of well-defined control algorithms (or ``blocks")
matically the two tuning adjustments-desired damping that can be selected and con figured by the user. These
and desired speed of response of the controlled variable. do not require special programming.
These are the only tuning adjustments, and there is no In many processes, steady-state gain, deadtime and lag
proportional, integral or derivative action to tune as is vary mainly as a function of variables that are measured.
true for a PID (proportional + integral + derivative) For such situations, the measured variables can be used
controller. Even if a more complex process model were to adapt the model parameters and, thus, provide further

Outlet-temperature
setpoint, r2

Adaptive direct-synthesis
control I er, ADsc
Desired respoilse :
Outlet • Damping gain, KD
tern peratu re, • Speed-of-response Steam flow,
r2 ratio, I/ri Ws

Process model :
• Steady-state gain, K
• Dead time, 7.or
Lag, rl

Characterization for K, rpr, Ti

Productflow,Wpl

Adaptive-direct-synthesis-controller
relationships Fig. 6

34
control improvement. This is consistent with the philos-
ophy of using as much known information about the
process as possible in order to improve its control.
Adaptive direct-synthesis control
In the heat-exchange example (Fig.1), the steady-state
gain varies inversely with product flow. For instance, if
the product flow is cut in half, the temperature will
change twice as much for the same change in steam flow.
Deadtime also varies inversely with product flow because
the volume of the exchanger tubes is constant. As prod-
uct flow increases, it takes less time to displace the liquid
in the exchanger. Heat-exchanger tests show that the lag
of the process also varies inversely with product flow [j].
A typical plot that characterizes the steady-state gain
as a function of product flow is shown in Fig. 5 (where
gain is expressed in multiples of the gain at full or 1007o
product flow). As the product flow approaches zero, the
gairl becomes infinite. To avoid division by zero (for zero
flow), the gain is kept at a constant value, say K]o, up to
some chosen value of flow-in this case,107o.
Characterizations of the process, deadtime and lag are
similar to the characterization for gain because they also
vary inversely with flow. Thus, the process parameters
are obtained either by using a characterization curve Feed forward-control systems are always based on
such as Fig. 5 or by computing them directly from the steady-state and dynamic models. They compensate for
inverse relationship. In either case, the product flow is measured load variables on the basis of feed forward
used to adapt the process-model gain, K, the deadtime, computations. These systems always have a feedback
7 -,,,, `, and lag, 7-1. trim controller to control the unmeasured load variables,
Moreover, speed of the desired response is also adapt- and to correct for inaccuracies in the feed forward model.
ed because the desired response lag is set in ratio to the The direct-synthesis controller can readily be used for
process lag. Since gain for the process model is adapted such feedback trim controllers because some under-
by the product flow, the damping gain (once set) pro- standing about the process model for the direct-synthe-
vides the desired damping regardless of product flow. sis controller already exists in the feed forward model.
The adaptive direct-synthesis controller is the same as An application of the direct-synthesis controller is for
the direct-synthesis controller, with the addition of the plug-flow processes such as are found in the pulp-and-
process-model characterization as a function of product paper and aluminum industries. Other applications may
flow, and is summarized in Fig. 6. be found in the minerals-processing industries.
The next article in this CE REFRESHER will appear in the
Where to use DSC and ADSC issue of Apr.16,1984.
Deadtime makes control more difficult. In particular, Steven Danatos, Editor
the ratio of deadtime to lag for a process is a measure of
the control difficulty. The higher the ratio, the tougher References
the control problem because no feedback information is
I Badavas, P C , Stabilit\ Ijimits and Parameter Sensitivit\ in Svnthe-
supplied to the controller during the deadtime portion si7.ed Digilal controllers, (,`r»i/7o/ I,`7/g , Sept l98l ` `
2 Bada\'as, P C , Mi(`ropro(`ess()r-Based, I)Ire(`t Svnthesis C()ntrollers for
of the response. Pulp and Paper I'ro(`esses, I'ro(`eedings of the i982joint Lsympt)slum ,
A typical PID controller will have to be detuned in Columbus, ()hio, Apr 13-15,1982.
3 SlvT:Z`,Srkac;I,:±`][{,;it,:::,rt)7:`,:`Sks, i:6);l(;rol S}'Stems," 2nd ed , pp 40 -4 1 , 8o,
order to remain stable for processes having a ratio of
deadtime to lag greater than, say, 0.5. The direct-synthe-
sis controller can be tuned more tightly because it uses
the model of the process implicitly.
Table 11 lists the properties of common control loops. The author
Paul C Bada\'as is a systems appli(`ation
Obviously, the standard PI controller is sufficient for (`onsultcint for The F()xboro Co , 38
liquid flow, level and pressure loops, and those for gas Neponsct A\c , Foxbt)ro, MA 02035
For o\'er 13 tears, he has been
pressure. These loops are essentially single capacity, or asso(`iated with the appli(`ation of
are very fast if multiple capacity. (`oiilrol to mduslritll processes
Currently, hc is responsiblc for the
However, the direct-synthesis controller is appropri- de\elopmcnt. dcsign and appli(ation of
ad\'an(`ed (ontr()I s\slems for the
ate for composition, temperature and vapor-pressure t`heniit`al prt>cess induLstries He
loops that are multicapacity and have appreciable dead- (ondu(`ts studics £`or pr()ccss plalils t()
imi)lemenl (`ontrol strategy and s\stem
time. Moreover, the adaplive direct-synthesis controller spet`ifi(`ation He has the B.S. and` M S.
can be used in these loops because process variables, in ele(`tr](`al engineering I rom the
Universit\ t>f Massat`huselts and the
such as feed flow, are usually available to adapt the Ph I) in i]le(`trical engineering fr(>m
model parameters. Purdue Uni\'cTsitv

33
H¥ffl" fffl use fieelELal¥ Tfflc]ms
flea lmeefl mlfenc:ass c:cDmcl:fl:cDris
Multiple feedback controllers combined with other hardware
components are a means for designing control systems to
handle the difficulties often met with in process systems.

Thomas C. Kurth, The Foxboro Co.

I Let us examine the principles of multiple feedback


controllers and other signal-processing elements in this
article, as a means of solving the more difficult control
problems. Signal-processing elements such as selectors,
arithmetic functions and switches, and their uses, will be
described. The discussion will include parallel, cascade,
ratio, constraint, multiple-output, and pacing control
schemes, along with the devices and circuits to imple-
ment such schemes.*

Parallel control
A simple example of parallel control is the use of
multiple controllers to maintain a single controlled vari-
able, such as level control (LC) in Fig.1. Although the
level in the filtrate tank would normally be maintained by
manipulating outlet (wash) flow, a second controller is
used to manipulate a makeup valve to ensure that liquid
is always present on the suction side of the discharge
Pump.
The setpoint of the second controller would be lower
than that of the first, and the proportional band would be
relatively narrow. (See Part 2, Chcm. E7tg., Aug. 8,1983
for information on the proportional band.) A low-level
condition would cause the makeup valve to open rapidly to provide the desired valve action for a given value of
to provide pump protection. A third, parallel controller the controlled variable.
(whose setpoint is fixed higher than that of the first
controller) and a dump valve could be added to prevent Cascade control
tank overflow. A feedback controller uses a single controlled variable
The general concept of one controlled variable with 7i (measurement) to set the manipulated variable. Fre-
controllers and " manipulated variables (provided 7} is quently, additional process information, particularly
greater than 1 but less than infinity) is shown in Fig. 2 load variables, can improve the performance of a control
(on the following page). The setpoints, of course, would loop.
be at different values, and the proportional bands adjust- For example, the application in Fig. 3a provides tem-
ed to cover different ranges of the controlled variable. perature control for the stream leaving a fired heater.
Care must be exercised in selecting the setpoints and Here, the manipulated variable is the fuel-gas valve
proportional-band settings for the respective controllers position. Changes in the fuel-gas header pressure (load
disturbances) would cause a change in fuel-gas flow, and
*Articles published thus far in this CE Refresher:
Part I, May 30,1983, p. 66; Part 2, Aug. 8,1983, p. 79; Part 3, Sept.19,1983, p. 67;
eventually, a temperature upset. Load disturbances can
Part 4, Nov.14,1983, p. 233; Part 5, Feb 6,1984, p. 99. be eliminated by introducing an additional measurement
The next Installment of this series is tentatively scheduled for the June 25th
Issue, and will cover the scaling of process variables for control for fuel-gas flow and a fuel-gas flow controller (Fig. 3b).

36
The temperature controller (TC) now manipulates the
setpoint of the fuel-gas flow controller (FC), which in
turn manipulates fuel-gas valve position. This is the
classic cascade-control structure for feedback controllers
that are nested within one another.
One aspect, however, begins to be dependent on the
control device. For example, suppose that the inner
flow-control loop were placed in manual. Because the
temperature-control loop could no longer manipulate
fuel-g`as flow, it would tend to cause a sustained error in
the temperature controller. This controller would satu-
rate at zero or full output because of its integral action.
In the example, integral saturation is prevented by feed-
ing back the fuel-gas flow measurement to the tempera-
a. Simple feedback ture controller.

b. Cascade control c. Proportional-plus-integral controller

37
To understand this, let us look at the construction of a
typical proportional-plus-integral colitroller (Fig. 3c).
Here an error signal is multiplied by a gain term (loo
divided by proportional band), and the product term
(error multiplied by gain) is integrated via a first-order
lag in a positive feedback loop.
Any nonzero error will cause the controller to eventu-
S,gnals
ally saturate unless the positive feedback loop is opened.
from
The source of the signal to the lag in the positive individuE
process-
feedback loop can be obtained from the controller out- Contro,
put or from an external feedback signal. Here, the use of loops

an appropriate external feedback signal, to open the


positive feedback loop, prevents integral saturation for
controllers that are used in cascade and other schemes.
External feedback is shown in Fig. 3b. The tempera-
ture controller (TC) is protected from integral saturation
by connecting the fuel-gas flow measurement (FT) to its Process
external feedback connection. When the fuel-gas con- demands
for air
troller is working normally, the setpoint and measure-
ment are equal. Therefore, the integral term in the
temperature controller operates normally. If the fuel-gas
Air
control loop is open (measurement not equal to the
compressor
setpoint), it has the effect of opening the integral loop of
the temperature controller and limiting its output. This
technique is useful for preventing integral saturation no
matter what causes the outer (temperature) loop from
reaching the setpoint. i

Valve-position control
An application in which the valve position is consid-
ered as a controlled variable is shown in Fig. 4. Here, an
air compressor discharges into a header having multiple
process demands as variables. Header pressure is main-
tained by the pressure-control loop, which manipulates
the compressor's inlet damper. Typically, the header-
pressure setpoint is maintained high enough so that
maximum demand for any process area will be met
without saturating the corresponding control valve.
While` this ensures proper control action, the scheme
does waste energy during those periods when process
demand is low and pressure drop across the control
valves is high.
What is desirable is to keep throttling losses to a
minimum when process demands are variable. To ac-
complish this, valve position can be used as a demand
index; i.e., at low demand, the control valve will tend to
close. Which valve has the largest opening can be deter-
mined by using a high (greater than) selector to choose
and output the largest of the input signals; a low (less
than) selector would output the smallest of the input
signals.
The output of a high selector can be fed to a valve-
position controller whose setpoint can be some value
that will allow process demands to be met in the respec-
tive process areas without saturating (i.e., reaching 907o)
the control valve. Output from the valve-position con-
troller will become the setpoint, in cascade, for the
header pressure.
If the valve having the largest opening is less than 907o
open, the position controller will reduce the header-
pressure setpoint. This will tend to open the other
control valves to minimize throttling losses. External
feedback provides protection against integral windup

38
(saturation) for the valve-position controller when the failure of a single field transmitter in a control loop
pressure controller is under manual operation. cannot be tolerated, because this would create an unsafe
condition or cause unacceptable downtime. Therefore,
Ratio control as many as three transmitters may be used to measure a
The blending of solid or liquid ingredients, or both, single variable. A means of processing these signals is
on a continuous basis often requires the maintenance of needed to give a reliable indication of the controlled
a ratio among them. The implementation is, of course, variable even if one of the transmitters fails.
dependent on the control device, -but the requirement is Fig. 6 represents a particular solution. Three high
to perform arithmetic computations within the normal selectors and one low selector are combined to generate
control scheme. This requirement introduces the ratio a median (middle number in a series) value. For the
device, which will take a given input signal, multiply it by configuration shown, if any one transmitter fails (either
some externally set ratio, and output the product. high or low), the resultant output will not be materially
Fig. 5 illustrates a blending application. Additives A affected. While the example is simple, it presents a
and 8 are to be blended with stock at fixed ratios, and the powerful concept-that of using selectors as ``switches"
total blend controlled to some total flowrate. To control to alter signal processing.
additive addition, the flow of stock is measured (flow
measurement here is linear) by a flow transmitter (FT). Constraint control
This signal is processed through respective ratio blocks In planning process-control schemes, it is not uncom-
for Additives A and 8 to generate the respective set- mon to encounter situations where process constraints
points for the additive-flow controllers (FC). Total require a more complex control strategy. The simplified
blended-product flow is controlled by measuring total reactor shown in Fig. 7 has reactant streams entering,
flow, and manipulating the control valve in the stock line. and a single-product stream exiting. Product composi-
Suppose that the availability of Additive 8 becomes tion is controlled by residence time and, therefore, exit-
limited to the extent that its flow is below the setpoint, stream flow. However, to prevent a critical overpressure
creating an off-specification product. It would be desir- condition, it may be necessary to override the product-
able for the control scheme to recognize this situation composition controller, as in the scheme in Fig. 7.
and take corrective action. For example, the scheme Under "normal" operating conditions, the composi-
could be modified so as to reduce total flow to the point tion controller manipulates the setpoint of the product-
where the flow of Additive 8 would equal its setpoint, flow controller. Pressure would be well below the
thus restoring the proper ratio. The blended product pressure-controller setpoint (pressure constraint on the
would now be on-specification, although at a reduced reactor). Under ``abnormal" conditions (upset in reac-
total flowrate. i tant composition) when the pressure starts rising above
setpoint, the pressure controller will override the com-
Signal selection position controller via the high selector. The pressure
A collection of techniques that incorporate logic to c.ontroller will now manipulate the setpoint of the prod-
provide the appropriate control action arises because of uct-flow controller so as not to exceed the pressure
an event such as an operator action or a process con- setpoint. As "normal" conditions return, the output of
straint. Simple examples will be covered first before we the pressure controller will decrease until the composi-
look at an enhanced ratio-control scheme. tion controller regains control of the product flow. Both
A simple example for using logic to solve a control controllers have increase/increase action; i.e., increasing
problem is signal selection. In critical applications, the measurement causes increasing output.

39
In this scheme, the ouput of the high selector is used
as the external feedback to both the composition and
pressure controllers. This provides protection against
integral windup to the controller that does not have
control of the product-stream flow, and a smooth transi-
tion from composition control to pressure control and
back. The power of this external-feedback concept
should become evident, and it will be used in the control
techniques that follow.

Multiple-output control
The control of a single variable with multiple similar
manipulated variables is frequently encountered. Let us
consider multiple compressors that discharge into a
common header, as shown in Fig. 8. The control objec-
tive is to maintain header pressure by manipulating the
compressor discharge valves. In designing this pressure-
control scheme, it would be highly desirable to make
sure that:
I Loop gain is constant, and independent of the num-
ber of compressors in service.
I Independent biasing of each compressor is allowed,
so that the proportion of the total load taken by a given
compressor can be adjusted.
I Manual control of each compressor discharge valve
is possible.
I Each compressor discharge valve is constrained in-
dependently, without upsetting overall pressure-loop
performance.
These objectives are attained by the scheme shown in
Fig. 8, the heart of which is shown inside the "envelope."
A very fast integral-only controller (MOC) adjusts the
signal levels to the respective valves simultaneously to
maintain the demand setpoint. This loop can be very fast
because it contains only hardware components that have
negligible dynamics. (The feedback in this loop is a
weighted average of signals to the respective discharge control objective of maintaining pressure, but will not
valves, and not actual valve positions.) necessarily provide safe process operation.
The summer (2/") computes the average valve-posi- Features that could be incorporated for safety include:
tion signal as a measurement to the integral-only con- 1. Letting air lead fuel on a load increase.
troller. The hand controllers (HC) on the discharge 2. Letting air lag fuel on a load decrease.
valves provide the ability both to bias one control-valve
signal relative to the others, and to manually set the
control-valve signal so as to baseload a given compres-
sor, or take it completely out of service. The low selec-
tors also allow a given compressor to be constrained
independently of the others without upsetting the pres-
sure-control loop.
This scheme will allow header pressure to be con-
trolled until all hand stations are in manual, or all
compressor discharge valves are constrained and/or sat-
urated. In both cases, the pressure controller is protect-
ed from windup via the external feedback of the average
valve signal.

Parallel metering
Controlling a single variable via manipulation of IV
unlike variables is a situation frequently encountered in
combustion control. A simple application is a single
boiler controlling header pressure with a single fuel and
single source of combustion air. In order to maintain
header pressure under conditions of variable load, fuel
and air must be adjusted in parallel. This will satisfy the

40
3. Requiring fuel flow to be decreased on a falling air-
flow condition.
4. Requiring air flow to be increased on a rising fuel-
flow condition.
Meeting these requirements will ensure that gas flow
does not exceed air flow. The parallel-metering scheme,
shown in Fig. 9, meets these design criteria. (For clarity,
only the control scheme is shown; boiler components are
omitted.) Both air and fuel flows are controlled via their
respective flow controllers. The output of the pressure
controller adjusts the setpoints of these flow controllers
in parallel. However, the setpoints can be constrained via
the high and low selectors.
In accordance with design requirements, output of the
pressure controller increases if steam load is increased.
However, the low selector will not let the fuel-flow Er....i.:....II
i;i;E':i:;i:;;!i:';;f;i;;!!!;;i!i;!!i:i;

.`..--.`..`..`,`..`..._.`_.`_.,.,_.+_.i++.r+..-..-..

setpoint increase until air flow increases. Thus, air leads :::':'' `'':i::: I

fuel flow on the load increases. On a load decrease, the


high selector ensures that air lags fuel flow.
Let us now suppose that air flow decreases at a given
load because of operator intervention, or some other
cause. The decreasing air-flow signal acting through the
lower selector will automatically decrease the fuel-flow : i:i+:L\
setpoint. Correspondingly, an increase in fuel flow at a
given load will automatically increase the air-flow set-
point through the high selector.
Control with constrained availability an index of availability. If either valve tries to open more
A combustion-control scheme for a natural-draft fired than 957o, limited availability of that additive will be
heater supplied with two fuels-one subject to limited assumed. In this case, the position of the stock-flow valve
availability-is shown in Fig. 10. will be overridden. Stock flow and, therefore, additive
The objective is to maintain the exit temperature of
the product leaving the fired heater. The manipulated
variables will be both waste-gas (subject to availability)
and natural-gas flowrates. Pressure will be used in the
waste-gas header as an index of availability; i.e., falling
pressure indicates a decreasing availability. In designing
the scheme, it would be desirable to make sure that:
I Waste gas and/or natural gas can be manipulated to
maintain temperature.
I Temperature can be maintained with no ``bump"
when waste-gas flow is constrained.
The flow of waste gas and natural gas is regulated via
respective flow controllers. The output of the tempera-
ture controller can be thought of as the manipulated
variable, or total fuel demand. The objective is to use
waste gas preferentially, subject to constraint. The out-
put of the waste-gas header-pressure controller can be
thought of as the constraint. Thus, the setpoint of the
waste-gas flow controller is the temperature-controller
output, limited by the constraint via the low selector.
Additional fuel to meet demand is determined by
subtracting the waste-gas flow from total demand. This
difference becomes the setpoint of the natural-gas flow
controller. Therefore, any decrease in waste-gas flow for
a given total demand results in a compensating increase
in natural-gas flow. Even if the operator placed the
waste-gas flow controller in manual, and closed the
waste-gas valve, temperature would be held.

Pacing control
The enhanced ratio-control scheme shown in Fig. 1 I
uses the positions of the valves for Additives A and 8 as

41
demand will be reduced-bringing the additive-flow
controller back in control. . . Hydrogenheader I
The high selector is used to determine which additive \!!

valve has the larger opening. The valve-position control-


ler (VPC) ensures that neither valve will open more than
957o, by overriding the total-product-flow controller via
the low selector. Both valve-position and total-product- +/val e•/.i;! .`:!j;,:!i .-
flow controllers are protected from windup by reset
feedback from the output of the low selector. Total flow
will be decreased, but the ratios of A and 8 flowrates to
stock flowrate will be maintained so that no off-specifica-
tion product will be made. Switch
I

I
I

Switching schemes I

To build logic capability into control schemes, ele-


ments other than the selector can be used as the primary
``switching" element. A simplified example involves
``Hy`drocracker
maintaining hydrocracker-bed inlet temperature by ma-
bed
nipulating the hydrogen quench-valve position (Fig.12).
Typically, a hydrocracker has multiple beds, most of
which have a corresponding hydrogen quench valve.
These valves are usually connected to a hydrogen head-
er. Therefore, the manipulation of one hydrogen valve I::::

tends to influence the flow of hydrogen through all the


other valves connected to the header. This interaction is I i]i.i.£i!jj!Ei.i!jiii!ji:i:i.:i;;;;gi!j. ''':

apt to propagate bed-temperature disturbances.


The tendency is to use cascade control to solve this
problem. For each quench valve, a flow-control loop
could be created, and the flow-controller setpoint could
be set from a corresponding bed-temperature control- In summation
ler. However, at certain processing conditions (causing A number of control schemes involving a combination
low quench-hydrogen demand), the output of the tem- of feedback loops and other elements such as selectors,
perature controller may fall considerably below the u-§e- arithmetic functions, and switches have been analyzed.
ful turndown of the hydrogen flow transmitter. Stated Note, however, that the examples in this article have
another way, the turndown ratio of the hydrogen valve is been simplified and should not be construed as imple-
significantly greater than that of the flow transmitter. mentation recommendat.ions. The intent here has been
The scheme in Fig.12 deals with this turndown prob- to present control-scheme concepts and the context in
lem. At low hydrogen flows, the cascade arrangement which they are frequently used.
switches, bypassing the flow controller, so that the tern- The next article in this CE Refresher will appear in the
perature controller now manipulates the hydrogen issue ofjune 25, and will deal with the scaling of process
quench valve directly (Switch Position 8). With normal variables for control.
hydrogen flows (Switch Position A), the temperature Steven Danatos, CorLinbuting Editor
controller would manipulate the flow-controller set-
point, the flow controller in turn setting the hydrogen References
valve. 1. Harriott, Peter, "Process control," MCGraw-Hill, New York,1964.
While this design has advantages, particularly in stabi- 2. S:i:S,kREWF+g;,k,„Fgn7e5gy conservation Through control,„ Academic
lizing the bed temperature, a number of implementation
3. g%£rnks,k]e87g: G.j "Process control systems," 2nd ed., MCGraw-Hill, New
factors have to be dealt with. It is obvious that, at the
time of switching, the output of the temperature control- 4. Shinskey, F. G., "ControllingMultivariable processes," Instrument soc.
of America, Research Triangle Park, N. C.,1981.
ler will not necessarily equal the quench-valve position.
Also, the switch requires a deadband so that the scheme
will not switch back and forth about the switch-flow The author
Thomas C. Kurth is a systems
setpoint. Whenever the flow controller is bypassed, it application consultant at The'uFoxboro
must track the valve position to prevent a "bump" when Co., 38 Neponset Ave., Foxboro, MA
02035. His work includes many studies,
being switched back into service. Also, the feedback path designs and implementations Qf analog
for the temperature controller must be switched, and and digital control systems for chemical
process plants including distillation
temperature-controller tuning may also be changed, for columns, chemical-recovery boilers,
incinerators, waste-heat boilers, and
the different modes of operation. The schematic does lime kilns. He has a B.S. in chemical
not cover these and other specific hardware-dependent engineering from Clarkson College, and
has completed course work for an M.S.
implementation details. The important concept is that in chemical engineering at Case
control-scheme structures can be switched in response Institute of Technology. He is a
member of the Instrument Soc. of`
to some process event, so as to achieve better America.
performance.

4a CHEMICAL ENGINEERING APRIL 30, 1984


PROCESS AUTOMATIONf7

Scaling converts process


signals to instrument ones
Mathematical equations for the relationships among process variables
cannot be directly acted upon by control systems. Instead, a scaling
procedure transforms such equations into ones having normalized
variables, which are compatible with instrument and control systems.

Lewis M. Cordon, The Foxboro Co.*

I Implementing advanced control schemes often re-


quires the use of signals from process transmitters and
regulatory controllers to perform engineering calcula-
tions. The equations that define these calculations must
then be "scaled." This procedure transforms the calcula-
tions into a form that is consistent with the way the
variables are represented and with particular hardware/
software being used.

Why scaling is required


The need for scaling arises from the distinction
between "engineering" and "normalized" variables.
The former are physical variables in the plant such as
pressure, flow and temperature; they are expressed in
engineering units of psi, gpm, and °F.
The latter (e.g., output signals from process transmit-
ters) are representations of engineering variables com- transmitter calibrated to a 100-to-500°F range is shown
pressed into a standard signal range. Most control sys- in Fig. I. At 100°F, the transmitter generates a 3-psig
tems, including many digital systems, operate by using output. As the temperature approaches 500°F, the out-
normalized variables. For example, pneumatic transmit- put approaches 15 psig. Outside the calibrated range,
ters generate a standard 3-to-15 psig signal to represent the transmitter cannot respond, and therefore does not
various engineering variables for user-defined ranges. represent the measured engineering variable.
Other control systems operate on different standard- The straight-line relation within the calibrated range
signal ranges. Most electronic two-wire systems use a 4- (Fig. I) identifies this as a linear response for which a
to-20 rnA d.c. signal. "Split-architecture" systems that midscale temperature of 300°F will be represented by a
separate the display and computing hardware often use midscale pneumatic signal of 9 psig. Depending on the
voltage signals-usually 0 to 10 V d.c. Internally, digital application, the lags created by the pneumatic tubing
systems often use a standard range to represent field make transmission of the signal impractical for more
signals. Usually, a 12-bit field (to store values in the than approximately 200 ft. Instead, the signal may be
range of 0 to 4,095 "counts'') is assigned to represent converted to a 12-rnA d.c. value, then to a 5-V d.c. within
the complete variation in the field signal. a split-architecture system, and finally stored in a digital
Where information is passed from one type of hard- system as 2,000 counts.
ware to another, the interface equipment transforms the For simple control systems, a user does not need to be
signal in the input range into a comparable value in its aware of the specific standard signal involved. The value
own range. The output of a pneumatic temperature of the normalized signal can be expressed as a decimal
fraction (0 to 1.0) of its standard range. Variation in the
normalized variable from 0 to 1.0 is observable evidence
3;iF#::b;tg:g¥sg§8:4;:6:+r5;:g::erg:n;t§;TA¢p§r,gR];#§:yghp:e:;7:,agE;ta;:,t,¥£X6o3vo,][4:8]3§g3,6p6,2Pg§: of variation in the engineering variable through the

CHEMICAL ENGINEERINGjuNE 25,1984 43


range to which the transmitter has been calibrated. The approximated by simpler functions. This will normally
signal range of the transmitter is then the analog of the be adequate if the variation in the inputs is not too great.
engineering range. Occasionally, signal characterizers may be used to devel-
Wherever humans interface with the s\stem, the nor- op an unusual function. For example, the flow of a liquid
malized .variable can be con\'erted back to`its engineering o\'er a rectangular weir in an open channel is related to
value. The con\'ersion mat. be as simple as an indicator the difference in height, A, between the surface of the
scale that reads the same as the range to which the liquid and the edge of the weir:
transmitter was calibrated, or as complex as the software
routines that generate displays for digital systems. In any Q.= kl VFT
case, the use of normalized variables is generally not where Q=liquid flow in volumetric units, /=length of
evident to the user, and allows standardization in the the weir edge, and 4 is a proportionality constant.
design of hardware and software. Thus, flow is proportional to A3/2. The relation be-
However, calculations with normalized variables are tween a variable and its value to the 3/2 power can be
not so direct. First, a given value of a normalized variable plotted, and then created with a signal characterizer. The
is meaningful only in relation to its transmitter range. A output of this unit can be used in a flow calculation.
flow of 50 gpm in a range of 0 to 100 gpm, and a However, the application of signal characterizers is sub-
temperature of 500°F in a range of 0 to I,000°F, would ject to constraints on things such as the variation in slope
each be represented by the same normalized value, i.e., and the number of straight-line segments in the repre-
0.5 or 507o signal value. sentation. For these reasons, the use of these character-
Also, the result of a calculation with normalized vari- izers is usually limited.
ables is itself a normalized variable, and is meaningful The instrument engineer must also be careful to in-
only in relation to an engineering range defined for the clude the constants necessary to make the equation
output. Sometimes, this will be dictated by how the dimensionally correct. For example, the calculation that
output will be used, but in many cases, it may be chosen computes mass flow from measurements of volumetric
by the instrument engineer to improve the overall utility flow and density might be expressed as:
of the calculation.
For both of these reasons, an engineering equation
M-pF mH

that expresses a relation among engineering variables where IW=flowrate in mass units, p=density, and F =
cannot be rewritten using normalized variables via sim- flowrate in volume units.
ple substitution. To be meaningful, the equation in Let us assume that the mass flow is to be measured in
engineering variables must be transformed into an tons/h, while the densit`. is measured in lb/ft3. Volumet-
equivalent equation in normalized variables by "scaling" ric flow is measured in gpm. Adding three conversion
the engineering equation. This is done in a way that takes factors, [o make Eq. (I) dimensionally correct, yields:
the ranges of the inputs and outputs into account.

Scaling procedure ¥ = H X Exmlno i34 # x


The procedure for implementing an engineering cal- 6o=x
h `
lton
2,000lb
culation in the domain of normalized variables (scaling)
is a rigorous and straightforward four-step process: Thus. for the a\-ailable measurements, the correct
S/pP J-Develop an equation using engineering `'ari- engineering equation for mass flow is:
ables to express the desired calculation.
S/cP 2-Express the engineering variables (inputs and
-\, - 0.004pf (2)

outputs) mathematically, in terms of the normalized


\'ariables, Step 2-Relate the variables
St¢ i-Substitute these relations into the engineering If the relation between the engineering variables and
equation, and simplify to obtain the normalized their normalized representation is linear (as in Fig. 1),
equation. the two variables are related b}' a simple linear equation:
Sfcp 4-Rearrange the normalized equation in a form
Engineering value = (Span)(\.ormalized value) +Zero
consistent with the capabilities of the hardware or algo-
rithm that will be used. Usually, a prime mark (') is used to indicate the
normalized representation of an engineering variable.
Step 1-Develop engineering equation For example, for a temperature transmitter whose repre-
The most important step in developing a normalized sentation is shown in Fig. I , the range of 100 to 500°F has
calculation is the derivation of the engineering equation. a span of 400°F, and a zero a[ 100°F. Then:
All relevant terms must be included, and their mathe-
matical relationships precisely defined. Generally speak-
' -400''+ loo
ing, instrumention is limited to the basic mathematics where !=temperature, °F; and £'=the normalized repre-
of addition, subtraction, multiplication and division, sentation of temperature. Thus, when !'=0.5:
together with square root. Powers, odd roots, trigono-
metric and logarithmic functions are usually not
' -400(0.5) + loo -300OF
available. Normalized calculations are usually restricted to first-
If an engineering equation calls for an unavailable order relations because of the limited functions available
function, the equation must usually be linearized and/or in hardware. The shape of the relation between the

44 CHEMICAL F.NGINEERINGjuNE 25,1984


engineering variable and its normalized representation this equation, the density term goes to I.0 as the output,
usually depends on the primary sensor. For example, an
p', from the density transmitter approaches I.0.
orifice differential-pressure cell combination (the most
common technique for measuring volumetric flow) gen- Step 4-Match to the hardware
Crates a signal that varies with flow squared. The scaled equation must be matched to the equation
Where the nonlinearity of the sensor is not too great, available in the hardware, without changing its mathe-
some transmitters include the capability to provide a matical identity, e.g., the equation performed by a Fox-
linearized output. In other cases, normalized signals boro M556-8 pneumatic multiplier is expressed as:
must be linearized before they can be used in calcula- P = f B(z+sG) (4)
tions. For example, the output from a differential-pres-
where 8 and C are normalized inputs, P is a normalized
sure cell measuring flow must pass through a square-
output, / and 5 are adjustable gain factors, and z is an
root unit before it can be used in the calculation
adjustable bias factor.
described by Eq. (2).
This unit can perform the calculation described by Eq.
Typical instrument ranges and their applicable equa-
(3). From a comparison of Eq. (3) and (4), the linear,
tions are:
normalized flow signal must be connected to input 8,
Variable Ran ge E quatio n while the density signal must be connected to input C.
Flow,/ 0 to 500 gpm /= 500/' Finally, the unit must be calibrated to make: /= I.14,
Pressure,P -10toloin.H20 P=2ap'-10 z=0,5, and 5=0.5.
Specificgravity,g I to I.25 g = 0.25g'+ 1.0 Matching the available hardware [o the scaled equa-
In each case, a normalized variable of 0.0 corresponds lion is a highly specific step. The forms and capabilities
to the zero of the range. A value of 1.0 corresponds to of hardware and software equations vary widely, not only
the upper value of the range. among various manufacturers, but also from one equip-
Sometimes, inputs are received directly from an ment line to the next. There are many unique limitations
operator through manual stations in hardware or soft- such as the number of inputs and the number and ranges
ware. In this case, the zero and span may be assigned by of adjustable constants. For example, the equation per-
the instrument engineer. formed by a Foxboro 2AX + MUL electronic multiplier is
expressed as:
Step 3-Substitute and simplify D±c! = (±G<4±o)(GBB±b)/(GcC ±c) (5)
Once the relations between the engineering and nor- where ,4, 8, and C are normalized inputs; D is normalized
malized variables have been defined, they may be substi- output; Gi4. GB, and Gc are adjustable gains, and cz, a, c
tuted directly into the engineering equation. Simplifying and cZ are adjustable biases.
the resulting equation produces the scaled equation. The variation in available equations is too great to be
The algebraic path taken in simplifying this equation can
considered in this article except by way of example. Each
be important. For example, for the mass-flow calculation line of hardware or software must be considered individ-
described by Eq. (2), let us assume that iw=O to 175
ually. Occasionally, the scaled equation cannot be
tons/h, p=50 to 100 lb/ft3, and F=O to 500 gpm. Then: matched to any single equation available. Hence, it may
M--175iivI' be necessary to perform the scaled equation in sections,
and then combine the results to obtain the final answer.
p -50+50p'
Several examples will illustrate the application of the
F - 500F,
scaling procedure.
By substitution into Eq . (2):

175IW` = 0.004(50p'+ 50) (500F)


Example: flow computation
The differential pressure created by a gas flowing
M,-- (0.004) (500)
(50 p' + 50) (F')
through an orifice plate may be measured to determine
175 its volumetric flow at standard conditions. But, the varia-
Next, we factor a constant out of the density term: tion pressure and temperature must be included in this
computation. A typical engineering equation* might be:
(0.004) (500) (50)
M'- (p' + I) (F') F = T]8 vrirrfe
175

In this version, the density term will always exceed I.0. where A=differential pressure, P=absolute pressure,
Because of hardware limitations, the equation should be r=absolute temperature, and F=volumetric flow.
factored so that no term that contains a normalized Let us assume that the ranges of the transmitters are:
variable can exceed I.0. The correct technique is to fr = 0 to 100 in. H20
factor out a new term equal to the upper-range value.
P = 14.7 to 64.7 psia (or, 0 to 50 psig)
This may be done as: 7T = 500 to 600°R
Maximum flow, F,w,4j¥, will correspond to the maxi-
(0.004) (500) (loo)
M|- (0.5p' + 0.5) (F') mum differential pressure and absolute pressure and
175
minimum absolute temperature, or:
Simplifying this last equation yields:
M' = I.14(0.5p'+0.5)F' (3) F+4zix = 278 V(100) (64.7)/506 = 1,000
*Miller, RW„ "Flow Measurement Lngineering Handbook," Chapter 9,
Eq. (3) is the correct form of the scaled equation. In MCGraw-Hlll, Ne\`r York,1983.

CHEMICAL ENGINEERINGjuNE 25.1984


4e
The range of F can be taken as 0 to I,000 standard
ft3/h. Then:
h - look,
A -14.7+5ap,
I -500 + 100T,
F -1,000F,
By substituting into the equation for F,w,4.¥, we get:

(100^') (14.7 + 5ap')


I,000F' -278
(500 + 100r')
Or, after simplification and reduction, we find:

0.83A'(0.23 + 0.77¢')
F,-
(0.83 + 0.17r')

Note that the pressure and temperature terms have been


Dividing the equation by 50, we obtain the scaled
factored so that they cannot exceed I.0.
equation:
Since the 2AX+MUL card, as defined by Eq. (5), can
both multiply and divide, the expression under the F'F --F'T-0.5F'R (1)
radical can be performed and used as the input to a
The gain factor of o.5 allows for the difference in span
square-root function. Fig. 2 shows the arrangement of
between FR and the other two variables. Since its span is
the instruments. Comparison with Eq. (5) gives the
only half as wide, a given increment in F'R represents half
constants:
as many gpm as the same increment in F'r or F'F.
G4 = 0.83 cz=0.0 Let us assume that Eq. (7) is to be implemented in
GB=O.77 b=O.23 pneumatic hardware. The Foxboro M136 pneumatic
Gc=0.17 c=0.83 summer performs the following equation:
d = 0.0 P -B+G(K\-C)
Example: Fuel-flow calculation Where: P is normalized output; 8 and C are normalized
inputs; G is adjustable gain; and K] is an adjustable
Fig. 3 shows a generalized representation of a furnace
constant.
and its fuel supply. Let us assume that is necessary to
Eq. (7) can be performed by the summer with G=0.5,
indicate the flow of fuel to the furnace. However, it is
and Ki =0. The total flow signal would be connected to
impossible to measure this flow directly. Instead, mea-
input 8, while the return flow signal would be connected
surements of the total flow and the return flow are
to input C. The output, P, would be the input to the
available for calculating the furnace-fuel flow.
indicator whose scale would have to read 0 to 50 gpm,
Let us assume that both flow measurements are in
linear.
gpm, and that the furnace flow is to be indicated in gpm.
The calculation is described by the engineering Scaling a feedforward equation
equation:
Fig. 4 shows a feed forward-control scheme for a heat
FF=FT-FR ® exchanger that uses steam to heat a water stream. The
scheme is designed to compensate for changes in inlet
where FF=furnace flow, gpm; Fr=total flow, gpm; and flowrate and inlet temperature before they upset the
FR=return flow, gpm. output temperature from the exchanger.
The flow ranges of the two transmitters are: The engineering equation performed by this scheme
Fr - 0 to 50 gpm ls:

FR = 0 to 25 gpm Fs* = kFw (To* -TIM


The next step is to define the range of FF. Since the
where: Fs*=steam flow setpoint, Fw=water flowrate,
signal is only going to an indicator, the instrument
To* = desired outlet-water temperature, r/= inlet-water
engineer can assign any desired range for this signal. To
temperature, and fa = units reconciliation factor.
be safe, it should be chosen to cover all possible values of
Rather than try to determine the proper value of A
Fr-FR. The maximum value of FF occurs when Fr is
analytically, an alternative technique is to compute its
maximum and Fj3 is zero. Its minimum value is zero.
value from a set of normal conditions.
Then: FR=O to 50 gpm.
For this exchanger, let us assume that typical values
The equations relating the engineering and normal-
are: Fs=12,000 lb/h, Fw=75 gpm, ro=120°F, and
ized variables are:
I/=60°F. For this system, the value of fa is calculated
FT - 50F'T
according to the following:
FR = 25F'F
By substituting into Eq. (6), we get: fa = 12,ooo/(75)(i2o-6o) = 2.67

50F'F = 50F'T-25F'R The ranges for the steam-flow and outlet-temperature

4e CHEMICAL ENGINEERING juNE 25,1984


setpoints may be taken from their respective transmit-
ters. For this exchanger:
Modifying transmitter ranges
Fig. 5 shows a system for calculating the mass of liquid
Fs = 0 to 20,000 lb/h in a tank from measurements of its level and density.
Fw -0 to 120 gpm Level is measured by a displacer. Let us assume that the
ro = loo to i5ooF variation in level along the displacer's length represents
rJ -50 to loooF a variation in tank volume of 800 to I,000 gal. Density is
Then:
measured by a differential-pressure cell transmitter,
20,000Fs*' = (2.67)(120Fw')[(loo + 50ro*') -
both taps of which are always submerged. Since the
(50 + 50r,')] vertical distance between the taps is fixed, the output
or, ` Fs*' = 2.40Fw'(0.33ro*'-0.33rf'+0.33) varies with the density of the liquid. Let us also assume
that this cell is calibrated to measure a density variation
The output range of 7.5 to 8.5 lb/gal. The maximum mass, M, in the tank
Defining the range of the normalized output is the could be: I,000 X 8.5=8,500 lb. The engineering equa-
most difficult part of scaling. As the examples for flow tion is: M=PA. By direct substitution into this equation,
compensation and fuel-flow calculation show, the range we find the scaled equation to be:
may sometimes be computed directly from the ranges of
the inputs. However, in the feed forward equation, the 8,5OOM' = (7.5 + I .Op')(800 + 2oofr')
output serves as the remote setpoint to a controller. = (0.88+0.12p')(0.8+0.2h') (8)
Therefore, its assigned range must be the same as the
range of the measurement to that controller. Occasionally, the designer may find it necessary to
Occasionally, the designer has the opportunity to readjust the transmitter ranges to allow the implementa-
choose the range in a way that improves the calculation tion of a scaled equation. For example, following the
functions given in the discussion of step 4, Eq. (8) can be
performance. Assuming that in the flow-compensation
example, the designer knows that differential pressure implemented in a Type 2AX+MUL unit but not in a
never exceeds 75 in. water column, and that pressure single M556 multiplier. Both a summing unit and a
never exceeds 40 psig, then flow will never exceed: multiplier would be required. However, if the differen-
tial-pressure cell were recalibrated to indicate 0 to 8.5
(75) (54.7) = 796.3 lb/gal, the sca,led equation becomes: '
FMAx = T]8
8,500M' = (8.5p')(800 + 200h')
Any upper-range values exceeding this one may be
assigned. For this case, the resolution of the calculation or, M' = P'(0.8+0.2A')
can be improved by assigning a range for F from 0 to 800 This equation can be implemented in a single pneumatic
standard ft3/min. By substitution, we find: multiplier.
Similarly, the scaling procedure, on occasion, will
800F' = 278 (100^') (14.7 + 5ap')
(500 + 1007") produce a scaled equation containing constants having
values outside the allowable range of the constants in the
hardware. By recalibrating one, or more, of the input
I.30^'(0.23 + 0.77¢')
F,- transmitters, or by redefining the output range, the
(0.83-+ 0.17r`') scaled equation can be changed into an acceptable form.
However, the designer must be certain that these This technique can often avoid the costs associated with
limits will never be exceeded. Should differential pres- breaking up a calculation into smaller segments.
sure and/or pressure exceed the expected conditions,
the calculation can saturate. S\imilarly, if differential Scaling in computers
pressure always exceeds some low value, a larger zero Scaling has traditionally been associated with analog
may be assigned for the output range. instrumentation. However, the responsibility for calcula-

CHEMICAL ENGINEERINGjuNE 25,1984 47


measurement of differential pressure is one input to an
Normalized variables input multiplexer, which periodically "reads" the signal
h' = Tank levelW'=Massintank from the transmitter and converts it to a raw digital
I- p' = DensityA,-.;;::i.::.,::''!'..,:. . „,
value. The software of a "scan" block, in turn, reads and
Displacer ` £.'.....

pe:# processes thls value, performing functions such as lin-


earization, noise rejection, and alarming. The final nor-
malized value is stored as the output of the scan block
and is available to other blocks in the software.
ffi For totalization, the total number of gallons may be
determined from this scanned value, using the approxi-
mation that the flow remains constant between readings.
a, - The engineering equation is:

Q.T -_60fi# + f4¥


60
+ . . . + IRE
60
rmffi!iffiEN

where/!=actual flowrate when scanned, gpm; A£=scan


interval, s; QT=total gallons from f=O to f=n.
This equation may also be expressed as:

QT--QTLl+#fi
Let us assume that the range of the flow transmitter is
0 to loo gpm, and that the scan period, A£, is 2 s. Flow is
to be totalized for 8 h. Finally, let us assume that even
though the transmitter is calibrated to read up to loo
gpm, the maximum flow will never be greater than 80
gpm. Thus, the maximum possible totalized flow is:

QM4x = 80mlnh
£SL X 60 Eil X 8 h = 38,400 gal

Then, the normalized variable that represents the


integrated total must represent the range of o to 38,400
gal. The relational equations are:
QT = 38,400QT'
f t - loot t'
By direct substitution, we get:
tions is increasingly shifting to digital computers. Actual-
ly, computational applications are even more prevalent
in digital systems because local control loops are usually 38,400Qr' = 38t400Qr-l' + qu
implemented in analog equipment to prevent loss of
control if the computer goes down. The question is: How 1

does the scaling process relate to digital systems? or, Qr' = Qr-l' + iTiin/`'
Scaling is a general technique, applicable whenever
calculations are performed with normalized variables. Let us assume that the equation performed by the
Whether or not scaling applies to a particular digital totalization routine in the digital system is given as:
system depends on .how the variables are stored.
Process variables are stored in one of two ways: S' - S'_I +
Smaller systems designed for process control usually
use a normalized range with their memory to pass values where S!= normalized variable representing the totalized
among control ``blocks" that are analogous to analog variable, and S!= the sample, normalized input. Then, by
instruments. comparisbn, a totalizer can be implemented on this flow
Larger systems, primarily designed for data analysis by assigning 41=1 and -,42=11,520 in the hardware
and process management, often store engineering val- unit.
ues. For such systems, engineering values are normal- In general, the block functions available for computa-
ized far transmission but reconverted to engineering tion using normalized variables in digital systems win
values for storage in the computer memory. have their unique equations-the same as the analog
Systems storing engineering values can perform engi- computational instruments that have been described.
neering equations directly. For systems storing values in The procedure and concepts for applying normalized
a normalized range, the scaling procedures presented in variables in digital systems are identical to those for
this article are entirely applicable. analog computation.
Let us consider the problem of using a digital system The next article in this CE Refresher, on advanced
to totalize a liquid flow over a period of 8 h. Fig. 6 shows types of control, will appear in the Aug. 20, 1984 issue.
the essential elements for such an application. An analog Steven Danatos, Editor

48 CHEMICAL ENGINEERINGjuNE 25,1984


IALrdaiceffi cofflffiol rliefELocls
Feed forward control, decoupling methods, relative gain,
and process modifications are techniques for resolving
the difflculties that arise in some process-control loops.

n4. nys, The Foxboro Company*

I Though simple feedback loops are dominant in con-


trolling a typical process plant, about 10 to 20% of the
control loops are more complicated. Typical examples
include: cascade, ratio, auto-selector and override con-
trols. The most difficult of these (about 597o) will require
the use of advanced techniques.
In this article, we will discuss the advanced techniques
of feed forward control and decoupling. Succeeding arti-
cles will cover additional procedures. Also, the problem
of interaction in multivariable processes will be illus-
trated by using the concept of rela-tive gain, and by com-
paring techniques for decoupling.
Better results can often be obtained by modifying the
process design. This should not be overlooked as a
means for achieving improved control. In this article,
process modification is considered the most important
method and will be discussed first. ;;;;;;;:.`:;:;;;i;
I...,:;:::•:.:.:~.,,-,

. ,. ,, ,,.,.,.':::::,`,`
i;:;;:;i;::;;:::;;:;::::;:;

::;:;;;;:i:;;:.:;:::;;;:;;:;;;;;;;;;;;..;;;;:;;;:::::;;;;;::;;;;:::;:;:;;:;:.;;;;:;;:;;;:;:;;;:;;::.:;;;;;:::;;;;;;:.

Modifying the process design -:;::;:::;;;;:;:;:;:;;;:;;;;:;;:::;::;;;:;:;:;.:;:;:;:;::::;;::;;::;:::::;;:;::::::

::i;;;;:;;:;ii'::::;;;::::::;;:i:;::::;::;:;;::::;;:
:.::;::;

: :'.: .,., .,•,., .,.•:.:.,. .;: :;~`,.,


` ,` , ,. ,. 1. ,` ,` 11 1. ,. ,` ,, -, _, -, -, -, -- -X. +-.- _.- _.- r r .;:;:;:i::i;;;;:;;..::::;::::i:;:::;;;:;:;.:;:;:;:;:;;;;;:.

•:: ;;;i:::i;:::::;:;:;:;;..:;;::;;;:;;:;;;::;;::;;;

The classic example for achieving advanced control ;:i ;:.:.:.:., ,., .,.,:; : :

through design modifications is the pH process. Here, .,i: ; :; i: ;: ;: ;:.,;: !;i:


:::;;; ;;;;;;;;;:;;;;i:;:;;.I:::;::;i::;;;;;:i:::
;;:::::i;;:;;;;;;i;;
. .` .\ -:::::,,,.,:.:.:. ;;;::;;;:;;;::i-;;:;. ;;;;: ::;::;:.. ::;: `:;;;;:;;:

•;;::;;:::;i;i;:;:;:: ..i::;:::;:::;;:::::::: ;:;,..i::,-:i;;:::;:;:;;;;::.;;;::;i;;;:;i::::;;:;:;;::;::::.,.

attention must be paid to details such as sizing mixers #


.,: : :;:., •,:. ~` ,` ,`•:.:.:.:.:. `.`\:;: ;•:.:.

with sufficient horsepower, and optimizing impeller - ,. ,. ,. ,I ., .` .1 ., ., ., -, .~ + -.


;E;::;;;:;:;i;

__
:;:;:::;:;;;:;:;;;:: ;i;;;;;;:::;:;;i;::;;; ;;;;:;::::;;;;;;;;;;:;;;;;;;;;::;:::;:;;:;:;:;::;:::;':;;:;;;;;;;

speed to minimize deadtime. In addition, the pH elec- • ;;::; :;;;:;:;::: ;:;;:;;:;::;;;;;;:;;:: ;:;:::;;;;;:;:;:;;;: :..ii!:;;:::

;:;:;:;;;:;:;;:;:::;;:::;:;;:.;;:;;;;;::::.:;:;;;.:;:;;::;::;;;;;::;::i;;;:;:;;; :;;:;:;;;;;::;;;'.;:;;::;;:;::;;;:;:'.;;:;;:;;:;;:; :;!!!:;;;::;:;:;:;:;;;;!¥'';::;:::;::


•.....:::...:,.I,...

:;:::;:;:;::;;:::;:::;:;:;;;:;;;;;;;;;;::;;;;;:::::;;::::;:;;;:;;::;::;;::;;;;;;;:;:;;::;;;;;;;::;;:;:;:;;:::;:

trode must be placed to allow fast response without ex- ffi


::::,.,...:.:.

;;;;;;;::;::;:;::;;;:;::::::;;;:::::;::;;:;;:;;::;:;;:;:

cessive signal noise, and the control valve must be located


•::;:;;;:;;;;;.:;;;;::;;;:;;;;;i;;;;:;.:;:::::;;;::;;;;;.

;;;::;:;;i;;:i;;;i:':::;;:;:;:;;;:;;::;:::;::;;;;:;;::i::;::: :::;;:.:`!i:;.'.i:::;:':i!.i!ii::i:i;:ii:'`:;;;;;:;;:;:::i;
::;:::::;:;::;:;:;;:;:i;:::;; ::::;i;:::.;;:;;;;i;;;;;:;::::;;.:;;;:;;::;i;;;;:;;;; :;;;:::::;:i::;;;::;:.::;::;;;;:;:;:::::::;:,::;;:;;.

`_''_|'-.'_.':1|J::rr:i:i+il:++|[zlrz±Izl:i

at the point of reagent addition to avoid downstream


.:.:.:.•:.:.:.

v ffli88888888gr88888888S#jREHREffi
*

piping and unwanted draining and dripping after the


valve is shut.
In single-vessel pH processes, even all this attention to
detail will not be enough to meet the accuracy require-
ments, which sometimes exceed one part in a million,
when simultaneous upsets occur. Hence, a design solu- The attenuation tank acts 5s a capacitor does in filtering
tion must be considered. For example, a second tank a noisy electrical signal, and will significantly reduce the
coul.d be placed downstream from the first for attenua- amplitude of the pH oscillations from the control tank,
tion, and a pH recording measurement could then be producing an effluent that is within the required range.
made on the effluent from the attenuation tank (Fig. 1). The attenuation tank is a low-pass filter-attenuating
Let us suppose that the pH neutralization tank (Fig.1) high frequencies and passing low ones. (An electrical
is oscillating at high frequency (30 s, peak-to-peak) with capacitor attenuates oscillations in voltage. Similarly, an
an amplitude that exceeds the required control range. attenuation tank smooths out oscillations of ions in solu-
tion.) It is necessary that the deadtime in the pH-control
* Articles published thus far in this CE REFRESHER: Part 1, May 30,1983, p. 66; Part
2, Aug. 8,1983, p. 79; Part 3, Sept.19,1983, p. 67; Part 4, Nov.14,1983, p. 233; Part
tank be minimized in order to keep the frequency high,
5, Feb. 6,1984, p. 99; Part 6, Apr. 30,1984, p. 77; Part 7, June 25,1984, p.141. and the attenuation-tank volume and cost to a minimum.

49
A tendency exists to blame the controller for poor con- (up to 1 part in 100,000) simulation and control capabili-
trol. Small improvements can be made by modifying the ties. This ability to solve equations accurately and contin-
uously was ideal for feed forward control. Today, analog
pH controller; e.g., a properly tuned nonlinear control-
ler will help to keep the amplitude of pH oscillations to a computers in the control industry have been separated
minimum and the frequency as high as the process will into modules to perform specific functions such as addi-
allow. In the final analysis, modifying the controller will tion, multiplication, division, integration, filtering (lag),
have only a minor effect on performance, while modify- differentiation (lead or derivative), characterization,
ing the process will have a major one. ramping, and signal selection.
The pure time-delay (deadtime) algorithm needed for
Feed forward control techniques realistic simulation and for dynamic compensation of
In principle, if an upset can be measured en route to a feed forward control systems cannot be implemented
controlled variable, feed forward techniques could be practically in analog controls. However, this algorithm
applied in a manner that would allow corrective action. can be easily implemented in digital processors, along
This exactly cancels the upset and maintains the con- with all the functions previously available in analog hard-
trolled variable constant. In practice, perfect correction is ware. In addition, digital processors can easily handle
seldom achieved with feed forward control because accu- complicated systems of equations, gate logic, sequencing,
rate feed forward compensation can be very complex on and iterative calculations. With present-day analog and
all but the simplest systems. digital computers, application of feed forward controls is
Combining feed forward control with feedback trim limited only by the availability of suitable load measure-
can be an effective way of achieving improved regulatory ments or sufficiently accurate process models.
control. High accuracy in the steady state can be obtained
with a feedback controller (especially if it has an integral Ratio control
or reset mode), and good response to upsets can be Ratio control is an effective form of feed forward con-
achieved by using feed forward techniques. (See Part 2, trol. Fig. 2 shows a commonly used ratio-control system,
Cfacm. E7ig., Aug. 8, 1983, p. 79.) which continuously adds 20% NaoH to a varying flow of
Clearly, conditions causing upsets should be elimi- water to produce 5% NaoH. If the water flowrate were
mated or reduced when this is more cost-effective than to change, the setpoint to the caustic flowlcontroller
adding feed forward control. When feedback controls would be increased or decreased proportionately, main-
cannot be arranged to respond fast enough to catch the taining a constant ratio of caustic flow to water flow. In
upset, feed forward controls can be an effective answer. this way, the upset has been compensated for before the
composition (density) has been affected.
Hardware and software constraints Feed forward action substantially reduces the amount
Feed forward techniques were first applied to control of feedback correction required for upsets in the water
boiler-drum level (three-element control), but it was not flowrate. The multiplier is scaled for twice the product of
until the pneumatic multiplier/divider was introduced the A and 8 function to obtain a feedback controller out-
that they began to gain acceptance for other processes. put of 0.5 (i.e., midscale). This allows the feedback trim
In the late 1960s, electronic analog computers (based to adjust the ratio equally well up or down from the nor-
on the operational amplifier) provided highly accurate mal value.

30
The "2AB" rule of thumb is acceptable when the flow- output limits of the controller can be used to restrict the
meter sizing for both water and caustic flows is consistent adjustable range of K by setting them for minimum and
with respect to orifice overranging. In other words, both maximum ratios.
flow measurements are normally of the same fraction of
the full-scale range. Feedback trim can be introduced Feed forward reactor control
with a summer, adding to or subtracting from the feed- Fig. 3 shows a feed forward control system for a refin-
forward calculation. The choice of using a summer or a ery reformer. This is a reactor manufacturing hydrogen
multiplier for feedback trim is mostly a matter of mini- for a downstream hydrocracker. The hydrogen pressure
mizing feedback corrections. in the hydrocracker system is an indication of hydrogen
Preferably, both flow measurements have been linear- inventory.
ized (i.e., square-root extractor for differential-pressure If conversion at the hydrocracker is increased, more
transmitter). Ratio control can work if both signals are hydrogen is consumed and the pressure will fall. To
"flow squared." Without square-root extraction, the ratio
maintain constant pressure, the controller increases the
of caustic flow "squared" to water flow "squared" will be
maintained. This is an accurate and acceptable imple-
mentation of a mathematical model. Mixing squared and
linear signals does not fit the mathematical model for
blending, and would not produce accurate results.
Feed forward controls are based on a model of the
process. In the ratio-control example, the model seems
intuitive. Actually, the model for the blending is based on
two simultaneous equations: the overall and the caustic-
materials balances, or:

Ft--Fw+ Fc (1)
F,xt -F( (I)
Solving for the desired value of the manipulated varia-
ble, Fc, the required caustic-flow setpoint can be calcu-
lated from the measured value of water flow, Fz„ and the
desired dilution concentration, x„ or:

Fzu
Fc- -KFu, (&)
[(X/X,) - 1]
For constant concentrations, the flow of caustic is di-
rectly proportional to the flow of water. Unlike open-
loop ratio control, the ratio value, K, is not calculated but
is determined by the feedback controller output. The

el
„`-'^.J- ,Moles of hydrogen produced per moles of feed +``1~^ the previously identified production factors will yield a
scaled equation for the summer:
F'i = 8|,F'L -82F'2 (7)
Component, H2 produced,
Component moles moles where the prime mark (') indicates scaled values, and gf
H2 Xa Xa and g2 are gain terms determined by scaling.
lnerts Xb 0 If the pressure controller is in automatic mode and is
CH4 Xc 4xc controlling well, its output could be determined by:
C2H6 Xd 7xc/
C3H8 Xe 10xe FL-%+ 8% (8,
CxH(2x + 2) X 3x+1

I -100 The calculation for Fi in Eq. (8) should be used by the


pressure controller for its external integral (reset) feed-
back connection to help prevent windup if the natural-
gas flow controller cannot follow its setpoint. The
natural gas to the reformer in order to produce more process operator greatly appreciates this feature-
hydrogen, which restores the inventory of hydrogen at particularly when the pressure controller is put in track
the hydrocracker. mode, as it should be if the flow controller were to be put
Hydrogen consumption at the hydrocracker changes on local setpoint or on manual.
slowly, and pressure control would be good except for an When the track bit (0 = no tracking, 1 = tracking) of
uncontrolled feed to the reformer. This uncontrolled the pressure controller is set, the controller output is
feed (called "wild gas") is a mixture of hydrogen, inerts, immediately kept equal to the integral feedback value.
and light hydrocarbons. When the wild-gas flow changes Thus, the output of the pressure controller is back-calcu-
suddenly, the pressure in the hydrocracker system will lated to precisely the correct value to keep the external
be upset. The feed forward system should reduce natu- setpoint of the flow controller matched to its measure-
ral-gas flow whenever wild-gas flow is increased by an ment. To put the system on control, the operator merely
amount that will not upset the hydrogen pressure in the puts the flow controller in automatic with remote set-
hydrocracker. point. Since the operator is relieved of adjusting the
Since wild gas has a composition that is different from pressure-controller output to line up the external set-
natural gas, it does not produce the same volume of hy- point and measurement of the flow controller, the possi-
drogen. The overall reaction in the reformer and the bility of bumping the process is reduced.
shift converter for methane is:
Two-variable feedforward control
CH4 + 2H20+ C02 + 4H2 (4)
Feed forward techniques can be used to compensate
Notice that four moles of hydrogen are produced for for simultaneous upsets. Let us consider the control sys-
each mole of methane. The overall reaction for an arbi- tem in Fig. 4. The feed composition and flowrate to a
trary alkane, CxH2x+2, is: single continuous distillation column are variable. This
can cause excessive impurities to appear in the bottoms
CrH2x+2 + 2XH20 + XC02 + (3X + I)H2 (5)
product. The temperature controller cannot be tuned
For light hydrocarbons (where x = 2) fed to the re- fast enough to catch these upsets.
former, moles of hydrogen per mole of feed is more Feed forward control is used in order to keep the bot-
than four. The hydrogen-producing power of the wild- toms temperature steady. The temperature of the two-
gas and natural-gas feeds must be compared on a volu- component mixture in the tower-bottoms sump (column
metric basis because they are metered as gas (using an pressure is controlled) has a direct relationship to the
orifice-plate flowmeter). The accompanying table lists impurity concentration. Even with multicomponent mix-
the moles of hydrogen produced for each component in tures, a sufficient correlation often exists. If the feed, F,
the feed. is a liquid at its bubblepoint when its flow is increased, the
The natural gas contains some ethane along with level on the feed tray will rise, spilling over the down-
methane, and will produce 4.2 volumes of hydrogen per comer weirs onto the tray below.
volume of feed. The wild gas contains some hydrogen The feed is analyzed for the fraction of lights, z. The
and inerts and will produce only 3.0 volumes of hydro- first multiplier simply calculates Fz, which is equal to the
gen for each volume fed to the reformer. In addition to flow of the lighter component entering the tower. The
the ideal gas law, 10097o reaction conversion is assumed flowrate of steam should be nearly proportional to Fz
and losses of hydrogen are neglected. because an increased flow of lights must be vaporized
The scaling (see Part 7, Cfac7". E"g., June 25, 1984, p. and removed as distillate. This proportionality, however,
141) of the summer can be obtained from a material bal- will not exist for all columns because variations in feed
ance of hydrogen in the hydrocracker system. If good composition may affect reflux ratio, or a substantial
pressure control is maintained, the overall hydrogen bal- amount of steam may be required merely because the
ance should be: 0 = Accumulation = Inflow -Outflow. feed is subcooled.
Specifically: The second multiplier simply allows the temperature
controller to adjust the ratio. The feedback trim from the
0--(F\+ F2) -FL (6)
temperature controller compensates for possible errors
The identiflcation of transmitter ranges and the use of in the measurement, model, and calculation.

8a
Fz-``. EI- "
Jy:'J+'6(;8 j;FBKisres tfe dback`:i+;i
4'\*` x * og`*a*#xxgrF°rfeS `

L'quldfeedf S-Z § F€
'`\`!`'
DTI M is deadtime gr83iif:;;i3`}t,1::c:3;';1r?xO;
BTIM - T ooaX o 'J'®*-a, ,

•E z I I.FTr_ -~FBK `,I


// Fs ,//
fz ' f 7Z *®
•ro
RTife°r=as\
s\` TC 1Ilo;,,Q.,a

a o' `o0000

O=pegracTTRT& `OOOOOO r¢`y

FTTTS
€:(„
` ;:, =if _-.
-,`a
team ¢(``
$5:J -- a lip;- J` i a- A(r -^=^1

: tl. r

± B%td°u_Tf ;o8:
- ri(:i;^

^i e€d
f
ormiard cantrnd §yslem handles tvur* upsets simurfarfeSusl¥
Lo a.`,a.``^
i,`v ^` r ,s-.,={i.:':J.f.'S.

111111111111_ ;:8!jg§,,`;€:(r„y5/:.,a::.,,,`:`^i;.?t^{;+?`as[£?:.^5:os`:

The most difficult part of this system involves the dy- of steam-typically by adjusting firing rate at a boiler. If
namic compensation. If the flow of feed suddenly in- this pressure controller is fast compared with those of the
creases, the effect at the bottom of the column is delayed other users or suppliers, the pressure in the header can
by the length of time it takes for the increased liquid to be maintained, thus minimizing interaction.
cascade down the column. Such interactions can be decoupled explicitly or im-
For most columns, it will take roughly 10 s/tray before
plicitly. Explicit decouplers use a process model (often,
the upset begins to affect the bottoms temperature. The including dynamics) in order for each controller to in flu-
deadtime and lag settings for the feed forward dynamic ence other interacting controllers in such a way that any
compensation should be based on actual column testing. changes in output reduce or eliminate the propagation
The deadtime is primarily a function of mixing on the of upsets to the other interacting controllers. Implic`it
tray, and transport delay in the downcomer liquid level. decoupling involves rearranging and/or tuning the con-
The lag time of each tray is approximately equal to the trollers in ways that make individual loops inherently less
actual volume of liquid on the tray divided by the liquid interactive. The major problem with interaction in multi-
flowrate. variable processes is the lack of identification of the ex-
The easiest and fastest way to get an estimate for dead- tent and mechanism of interaction.
time and lag time is to increase the feed flow about 10%
with the steam flow held constant, while at the same time Relative gain
recording feed and bottoms flowrates. This procedure To analyze loops for interaction, Shinskey [2] employs
has been successfully used to accurately set the dynamic- a technique developed by Bristol [4] called "relative
compensation time constants, which change with flow-
gain." It has been increasingly used to guide control-sys-
rate, composition and tray level. These may require peri- tem arrangement for distillation columns, and is applica-
odic readjustment. ble to a wide range of interaction problems. It is success-
ful because it quantifies the speciflc amount of
Multivariable control techniques interaction and can be used for any control loop. Relative
Process interactions arise from interconnected net- gain is defined as:
works of mechanical, fluid or electrical components. In
some cases, the interactions are intentional; in others,
they arise as an unavoidable consequence of the process ^3,-
EEI
design. For example, if a large steam user suddenly starts acJ (9)

up, it will decrease the pressure in the steam header- 6m,1 c-k
possibly causing upsets to other steam users on the same where C; = Controller c., ^Gj = relative gain for Controller
header. This interaction occurs as a result of the piping i with Valvej.
network. The relative gain, which is the ratio of two gains, can
When a user or a supplier moves a control valve, all be used to determine if Controller 8. should be connected
other users and suppliers are affected. Usually, the to Valvej. The numerator in Eq. (9) is the change in the
header pressure is controlled by manipulating the source controller measurement with a change in the valve posi-

85
CV-2 C;N -2

a. Blending process with implicit decoupling b. Blending process with explicit decoupling

(NOT) (oFl)
_-----Track
(1 = Track mode)

x?8iaer§+Oi,oy `aog _y ) +. ,+'

EM/ 1 = Automatic/

?StreamA x ---FBK FBK--,"SKesri;X`?r%3

pr
V1

+gf%mein¢%

unllllNI CV-1

y *:J3:+8Zl` tot?Sg$3
-;S
§:u¥„9RE£®gr?:

EM
/ilgES€.`o*:-a./
tic
?-Autom-a,Streams
I

~-V2 EM= Panel or cathode-ray tube (operator interface)

External manual switch (0 = Manual)


FBK= Beset feedback
AMS= Automatic/manual station
A/M= Automatic/manual
0'1= Binary logic

c. Explicit decoupling with improved operator interface

tion with w = A (that is, all other valve positions in the relative gain is equal to 1.0, and Valve 2 will control
plant are fixed). The denominator is the change in the Measurement I without interaction from any other con-
controller measurement with a change in the valve posi- troller in the plant.
tion with C = fa (that is, all other controllers in the plant Let us consider the blending process in Fig. 5a. To
moving their valves as needed in order to maintain their determine which valve should be used to control the total
measurements at setpoint). flow, calculate the relative gain for the flow that is con-
The relative gain can be obtained by field testing. In trolled by valve CV-I. If Valve I is opened and Valve 2 is
practice, it has usually been calculated from simple mate- fixed, the flow increases, indicating that the numerator
rial-balance equations. Ideally, moving the valve should of Eq. (9) is a positive number. If the composition con-
affect the measurements to be controlled equally, troller is in automatic when Valve 1 is opened, the flow
whether the other controllers are in automatic or man- will increase. However, the composition controller will
ual. In this ideal case, numerator equals denominator, open Valve 2, increasing the flow further in order to

e4
keep the composition at setpoint. Thus, the denominator gain of 0.8, the opposite pairing would give a gain of 0.2.
is greater than the numerator and the relative gain is The 0.8 combination would be the preferred controller
somewhere between zero and one. arrangement, but some interaction would still exist.
To determine the specific relative gain, calculate the For two valves and two controllers, the relative-gain
numerator first. The analysis is slightly easier if valve array is a 2 X 2 matrix. Here, the sum of the columns
position is proportional to the individual flows, i.e., a lin- and the sum of the rows is equal to I.0. For a few proc-
ear installed-valve characteristic. The material balance is: esses, the relative gain is actually negative. The general
matrix is:
FT= FA+ FB (10)

The controlled variable, C„ is the total flow, F7i and M+M2

the manipulated variable, M,, is the flow FA through All I-All


I-All All (14)
Valve 1. Therefore, with Valve 2 held fixed, calculate the
numerator as: This property of Eq. (14) helps to simplify calculations
because only one relative gain, A I I, in Eq. ( 14) needs to be
@F7, 6(FA + FB)
-I.0 (11) evaluated in order to determine the interaction of two
in-h 6FA FB=k 6FA
loops.
To calculate the denominator, the flow FB is no longer
constant but must be adjusted to allow the composition Implicit decoupling
controller to hold its setpoint. Clearly whenever FA is It would seem logical that a decoupler be required to
changed, ft`B must also change if composition is to be held handle two interacting loops with relative gains of 0.5.
constant. If perfect composition control is to be realized, However, its use may not bejustified even in the case of
then the ratio of the two flows must be constant, i.e.: maximum interaction. If the blending process does not
have upsets or setpoint changes, both the flow and com-
FB/FA= k (\Z)
position controllers would eventually settle out and an
Keeping in mind that FB is now a function of FA: extra decoupling device would not be needed to achieve
stability. Also, the consequences of an occasional upset or
8(FA + FB) setpoint change might not be sufficient to justify a
-1
decoupler.
;.' Even if the interaction mechanism is well understood,
(13)
the vast majority of interacting loops do not have
6(F A + hF A) decouplers because a large number of less-costly and
-1+4
c-k 6FA easier alternatives often exist.
When controllers have been assigned to particular
The relative gain of the flow controller connected to valves, tuning can have a dramatic effect on the amount
Valve 1 is the ratio of the numerator and denominator
of interaction. The blending example of Fig. 5a shows
calculated for Eq. (13), or: implicit decoupling; controller tuning can have a decisive
^z' -1'(I + fe) impact on its effectiveness.
Typically, it is more important to keep composition
Here, it is apparent that the relative gain is not a con- under control. If the composition measurement is fast-
stant but is a variable, dependent on the value of 4. responding (density, conductivity or infrared analysis), it
If FB is equal to FA, then fa is equal to I.0. The flows will would be possible to tune the composition controller
be equal at only one particular setpoint for the composi- tightly. This would be done using techniques discussed
tion controller if feed compositions are constant. For this earlier in the series except that the flow controller should
particular setpoint, the relative gain is 0.5, which is the be in manual while the composition controller is tuned.
worst interaction possible for the blending example. A (See Part 3, Cfoc7ra. E7tg., Sept. 19,1983, p. 67.) The flow
relative gain of 1.0, which occurs when the numerator is controller should be tuned sluggishly (wider or larger
equal to the denominator, is ideal: without interaction.
proportional band), and the integral (reset) time should
This ideal case will occur when none of the other control- be longer.
lers in the plant has any effect on the prospective control Tuned in this manner, the flow controller takes the
loop. The flow controller in our example will have a rela- brunt of an upset, while composition control remains
tive gain of I.0 when the flow through Valve I is much nearly as good as could be achieved in a noninteracting
larger than flow through Valve 2.
process.
In this analysis, it has been assumed that no other If the composition measurement is slow (e.g., a process
valves in the plant affect the total flow. Other loops can chromatograph) or a fast measurement is used having a
interact if they cause variation in feed composition or in long time delay associated with the sampling system, the
valve pressure-drops. These effects must be considered composition controller will be slow with a long period. In
in evaluating the denominator for the relative gain. If this case, the flow controller can be tuned tightly. The
they are neglected, the relative gain will not reflect these composition control will be slow, but it would be nearly as
potentially troublesome interactions. slow without interaction.
If the relative gain for a controller connected to either Generally, if two interacting loops oscillate at different
one of two valves is equal to 0.5, interaction is maximum. frequencies, their controllers should be tuned to further
The controller connected to either valve will work separate these frequencies-thus minimizing interac-
equally well or equally bad. If one pairing gives a relative tion. Two interacting loops oscillating at the same fre-

Be
sentative of the valve position, and a separate in
quency can always be separated somewhat by tuning.
Furthermore, frequency separation can also be accom- display would have to be used.
If direct manual control of the valves is required
plished by changing the deadtimes or lags in the process.
Sometimes deadtime can be reduced inexpensively by automatic/manual (auto/manual) device could be
moving the measurement location, changing the sam- serted between the control valves and the decou]
The operator could use the auto/manual station in r
pling system for the measurement, or moving the control
valve. If surge tanks are used to smooth flows or compo- ual mode to stroke the valves to any desired posii
sitions, then the associated lag can be used to stabilize one However, upon return to automatic mode, a complic
or both of the interacting loops. balancing procedure would be needed to avoid bum
Another way to help reduce interaction is to prevent the valve position. If this trial-and-error balancing pl
upsets from reaching the process. Feed forward control dure must be done often, the operator would prot
could be used to catch these upsets. Also, some upsets find the decoupler difficult to use.
can be scheduled or can be reduced by changing operat- Fig. 5c shows the decoupler system with an impr
ing procedures at the origin of the upset. operator interface and automatic balancing. If an
In the home, feed forward scheduling and changing manual station is put in manual, a logic signal swi
operating procedures are used to reduce interaction. For the other auto/manual station into manual, and
example, the shower water becomes scalding hot when- both of the feedback controllers into track. This a
ever cold water is used at the kitchen sink. To overcome the operator to stroke one valve with the other
such an interaction, the temperature loop can be tuned flxed.
faster, the use of the sink can be delayed until after the Upon return to automatic mode, bumpless tra
shower, the piping to the sink can be restricted, the ther- can be achieved because the controller reset feedb;
mostat on the water heater can be turned down, or a back-calculated. With both controllers in the track n
temperature-controlling shower valve installed. All of a positive feedback loop is established and the poss
these techniques, except the last, are examples of implicit of an unstable feedback calculation exists. A first-(
decoupling. lag with a setting of perhaps 0.1 min could be ins
downstream of either decoupler (see Fig. 5c) to
Explicit decoupling stability.
Implicit-decoupling techniques such as previously de- The cost and complexity of the explicit decoup
scribed are not always practical or effective. Occasionally, often increased in order to provide an acceptable o
explicit decouplers are the most effective way of remov- tor interface. The operator need not know the deta
ing troublesome interaction. The temperature-control- the decoupler design, but must know how to start, I
ling shower valve is an effective decoupler with inde- ate, and shut it off. If the operator is comfortable
the decoupler interface and if the decoupler works,
pendent adjustment of flow and temperature. Many
building codes require them for new construction. The the operator will view it as a way to achieve impi
decoupling shower valve is effective because it (1) works, operation.
The next article in this CE REFRESHER will app
(2) is economical to build and install, and (3) is compre-
hensible to the operator. These three requirements are the Oct. 15, 1984 issue, and will deal with feedfo
control.
prerequisites for a successful explicit decoupler.
Stwen Dam
The explicit decoupler shown in Fig. 5b is designed to
allow the flow controller to adjust the total flow, y, and
the composition controller to adjust the ratio of flows, X.
Linear installed-valve characteristics are assumed, and
References
I. Buck]ey, P. S., "Techniques of Prow:ess Control," Wiley, New York,1964
the dynamics can be neglected if both valves have the 2.§iHn[sFkfyd3y:FE:£rao:'hMnino£`[an:eaB,g±¥#oPos#}sfenfsT£:t3::,Pketcot
same time constant and are very close to the blending
Canada, jul}' 1977.
junction. The factor K in Fig. 5b has been added to 3. Ryskamp, C. J., Explicit `.ersus Implicit Decoupling in Distillation Contr
equalize valve size. n6ering'Foun-datioh Conference. Sea Island, Ga., lan. 1981

The flow-controller output y is proportional to the 4. Bristol, E. H., On a New Measure of Interaction for Multivariable Pros
trol, JEEP rra7as., Vol. AC-I I, Jam. 1966.
total flow, Vi + V2:

- '-I r-
5. Rinard,I. H„ A Roadmap to Control System Design, Cfeow £`7ig., Nov
pp. 46-58.
Xy y(I +X) -y
vl+v2-#+ X+1 X+1
(15)

The author
The composition-controller output X is proportional =-_----

to the ratio of flowrates, V2/Vi:

Xy
V2 _ X+ 1 -X
distillation columns. He has a B.S. in
Vly
.i.i :¥e#::::cehnu9entets:rE8]fsr:mm:h#rn:Vfe
Instrument Sou. of America, and is a
registered professional engineer in
This decoupling scheme may meet the first two re- Massachusetts.

quirements of an effective decoupler, but the operator


interface may be confusing. If a standard controller dis-
play is used, the controller output would not be repre-

ee
Feedforward methods for
processcontrolsystems
The design of feed forward-control systems requires the use of steady-
state, inferential and empirical process models, dynamic compensation,
and feedback correction. Practical examples using these elements show
how to apply the procedures for achieving good control.

Paul C. Badavas, The Fondoro Co.*

I In all process-control loops, the points where the load If all the load variables for a particular process were
variable enters the process and where the controlled var- measured without errors in the primary measuring,
iable is measured are not the same. The longer the proc- transmitting and computing devices, and if the relation-
ess deadtime, the more difficult it becomes to maintain ship between manipulated variables were exactly known,
the controlled variable at the desired setpoint. This is then perfect control could be achieved. Of course, this is
particularly true when load variables change frequently not practical nor, even necessary.
in relation to the process deadtime and when the change The berinning step is to decide which are the major
is large. load variables (based on how often they change in rela-
A more effective approach becomes necessary when tion to the process residence time) and how large is the
the process deadtime is large, when load upsets are fre- change. The major load variables are the ones that need
quent and large, and when high purity is important or to be measured. All others (termed minor load variables)
the monetary value of the product is great. A relation-
provide load upsets that are relatively small or are usu-
ship for the controlled variable must be found among ally controlled by individual feedback controllers having
load variables, manipulated variables, and desired set- relatively fast dynamic responses.
pointssothatmanipulatedvariablesarechangeddirectly The next step is to find the steady-state relationship
as load variables change. Typically, this relationship is
(be it material and/or energy balance, inferential or em-
based on material and energy balances and on inferential
pirical) that relates the manipulated variable to the meas-
and empirical relationships (to be discussed later in this ured load variables and the setpoint of the controlled
article). Designing control systems in this fashion has
been termed "feed forward" control.
The three elements necessary for implementing feed-
forward control are shown in Fig. I. These are the
steady-state model for the process, dynamic compensa-
tion, and feedback correction.

Steady-state model
The steady-state model is developed by using material
and energy balances, and inferential and empirical prac-
ess relationships. The manipulated variable is then com-
puted as a function of the measured load variables and
the desired setpoint. Thus, control correction is made
directly as the defined load variable changes.
fro mcct the author, see Chem. Erg., Feb. 6, 1984, p. log.

gi?;,cf:g;f:I,§¥f!#9:,|i5f':,.£iprtc3,:9,,:i:8:3#e;77,,Pja¥;,,J¥i!e;:,;il39?828,,gpi3i;

e?
variable. This provides the steady-state model of the trolled variable from its setpoint. To eliminate this offset,
Process.
a feedback controller is added to the control system, as
shown in Fig. I, to force the controlled variable to its
Dynanic compensation setpoint. In essence, the feedback controller makes up
Rare is the case in which load and the manipulated for all the inadequacies and inaccuracies of the feed-
variables occur at the same point in the process, and dy- forward model over a period of time.
namic compensation is not needed. Typically, load and Feed forward control was introduced because the feed-
manipulated variables enter the process at different loca- back controller alone could not maintain the controlled
tions. This causes a dynamic imbalance because the re- variable close to the desired setpoint for large and fre-
sponse of the controlled variable to changes in the ma- quent changes of major load variables. The feed-
nipulated variable is different from its response to forward-control computation does most of the control
changes in the load variable. In this situation, dynamic work by responding to changes in measured load varia-
compensation in the form of lags, lead/lags and/or dead- bles directly. The feedback controller \t,hen provides less
time (see Parts 1 and 5) is required so as to minimize the corrective action than the feed forward part. That is why
effect of the dynamic imbalance and thus improve the feedback correction is referred to as feedback "trim."
performance of the feed forward-control system. Integral control action is usually sufficient for feedback-
trim controllers. However, proportional and derivative
Feedback correction action can help to control some of the unmeasured
The feed forward system can provide perfect control if minor load variables that are not in the feed forward
the process can be modeled exactly, with completely ac- computation.
curate measurements and computations. However, this Typically, feedback trim is introduced as the replace-
is not possible because of errors that can result from: ment for the setpoint of the feed forward model. In some
1 Inndequale steady-state models for the process-:Mzrterizh cases, however, it provides adjustment to one of the coef-
and energy balances usually contain mass- and heat- ficients of the model because variations of process char-
transfer coefficients whose values vary as functions of acteristics, corresponding to that coefficient, provide the
time and the operating point of the process. For exam- major upset to the controlled variable.
ple, the heat-transfer coefficient of a heat exchanger var-
ies as the heat-transfer area fouls over time. For complex Heat-exchanger control
processes, the models are complex and nonlinear. Line- Fig. 2a shows a heat-exchanger process in which feed-
arization about operating points of the process and sim-
forward control has been successfully implemented by
plification of the models usually results in less-accurate using an energy-balance computation, along with dy-
representation of the process phenomena.
I Inaccuraci,es in load-variable rneasure'rnents-Such inzLc-
namic compensation, and feedback-trim controllers. The
objective is to heat the product of liquid flow Wp and inlet
curacies in the measurements occur in many places, e.g„
temperature 7`i to a temperature T2 by manipulating
the primary measuring devices, orifice plates, vortex
steam flow Wf .
meters, thermocouples, resistance bulbs, and analyzers.
A typical system is cascade control, whereby the tern-
Another source of error is in transmitters and other con-
verters that bring the signal to the control and comput- perature controller (TC) provides the setpoint of the flow
controller (FC). The flow controller [here, standard pro-
ing devices, and back to the valves and other actuators.
In digital systems, the analog-to-dicttal (A/D) and ditl- portional-plus-integral (PI)] provides the demanded
steam flow, assuming that the steam supply from the
tal-to-analog (D/A) convert.ers also add to the error. Of
boiler or header does not vary. The flow loop usually
course, the more bits (for the digital representation) used
responds much faster than the temperature loop; there-
in the conversion, the better the resolution. For practical
fore, a standard PI controller is sufficient.
purposes, these errors are essentially negligible for A/Ds The outlet temperature of the exchanger responds
and D/As in comparison with errors during transmission
more slowly to changes in steam flow, because it takes
and in primary sensors.
I Measureri'ients Of major load ved]les only-The set of time to overcome the lags associated with transferring
heat to the product flow. Also, the rate of product flow
minor load variables is not measured, and therefore, is
influences the residence time of the loop, because at a
not included in the computation. Hence, the feed-
forward system does not correct for changes in minor given flow it takes a certain amount of time to displace
the product volume in the tubes of the exchanger.
load variables. (Flow through a heat exchanger is an ex-
The major load variables that affect the outlet temper-
ample of a major load variable, while a small variation in
ature are the product flow and the inlet temperature.
the steam-supply pressure is an example of a minor load
Outlet temperature control is substantially improved by
variable.)
I Coy"Pt4tocio?'i¢/ c7Torj-The accuracy of the comput- using feed forward control. The steady-state heat-bal-
ance model is:
ing devices is important in analog systems, where pneu-
matic or electronic devices such as summers, multiplier/ Q." -W.ZJs (I)
dividers and leadAag units are used. For practical pur-
gout = WpCp(T2 -Ti) (2)
poses, digital systems (including those that are
microprocessor based) are capable of providing the com- where Ws = mass flow of steam, kg/h; Wp = mass flow of
putation, with negligible error. feed, kg/h; QS„ = heat-transfer rate input by steam, kcal/
The cumulative effect of errors in the feed forward- h; Qot„ = heat-transfer rate to the heated product, kcal/
control computation does result in an offset of the con- h; Hj = latent heat of steam, kcal/kg; Cp = heat capacity

88
of the liquid product, kcal/(kg)(°C); Ti = inlet tempera-
ture, °C; and r2 = outlet temperature, °C.
The heat supplied by the steam is equal to the heat
absorbed by the liquid, or:

Qj„ -Q.ut (3)


w5Hs = wpcp(r2 -T]) (4)
Solving Eq. (4) for steam flow (the manipulated varia-
ble) yields:

Ws = KWp(T2 -Ti) (5)


where K --Cp/Hs (a;)
Letting r¥ indicate the desired outlet temperature,
and substituting it for r2 in Eq. (5) yields:
ws = Kw4(T¥-r]) (7)
In contrast, Fig. 2b shows the feed forward control
computation as implemented by using Eq. (7). The re-
quired steam flow is computed to compensate directly
for changes in the major load variables of product flow
and for changes in inlet temperature.
Fig. 2c shows a typical feed forward-control-system re-
sponse with the temperature controller on manual, thus
providing a fixed desired setpoint r¥. Let us consider
the case (Fig. 2c) without dynamic compensation. As the
product flow increases, the steam flow also proportion-
ately increases. However, the change in product flow is
faster than the change in heat transfer to the product in
order to maintain the outlet temperature at the desired
value. This dynamic imbalance between load and manip-
ulated variables causes the outlet temperature to drop
before eventually returning close to its setpoint.
The effect of the dynamic imbalance is minimized by
using a lead/lag dynamic compensator with predominant
lead action, as shown in Fig. 2c. As soon as the product
flow is raised, steam is increased much more than the
proportional amount needed when steady state is
reached. The steam flow then decreases, at a rate set in
the lead/lag compensator, to the steady-state value
necessary to maintain the outlet temperature close to
its setpoint.
In essence, the dynamic compensator provides extra
heat during the transient part of the response at the ap-
propriate rate. This minimizes the amount that the out-
let temperature varies from its setpoint because of the
change in product flow. Dynamic compensation is usu-
ally not needed for the inlet temperature because it does
not vary as fast or as much as the flow.
The outlet temperature response in Fig. 2c shows the
existence of an offset; i.e., the actual temperature does
not reach the desired setpoint. One reason for this effect
is that the coefficient K in Eq. (6) is only approximately
known. It changes as the heat-transfer area fouls and
causes changes in the heat-transfer coefficient. The coef-
ficient A also changes with variations in steam enthalpy
because of temperature and pressure. Another reason
for the effect is that minor load variables such as heat
losses are not taken into account in the energy-balance
equation. Transmitter errors and other calibration and/
or computation errors also contribute to the temperature
offset.
The temperature controller, as shown in Fig. 2b, is


used to produce the feedback trim by providing the set- For an evaporator having 7& effects, Eq. (14) becomes:
point T¥ in the feed forward computation. The tempera-
ture controller produces the necessary feedback correc- wo[l + #] -€flv3 ,15,
tion continuously, and forces the temperature to reach
its setpoint without offset in the steady state. Moreover, where #„ = weight-fraction of solids coming out of the
this controller can be tuned to minimize the upsets re- 7t-th effect, and where V2 = mass flow of vapors in the
sulting from minor load variables not taken into account 8.-th effect.
in the feed forward model. The energy balance relates the amount of total vapors
Evaporator control removed to the amount of steam used in the first effect,
Feed forward control has been implemented in evapo- and is described by:
rators using a material and energy balance, along with
dynamic compensation, and feedback-trim controllers. £v3-EWs (16)
To improve heat efficiency, the vapors generated in €-I
one stage (or effect) are used as the energy source for the
where Ws=mass flow of steam, kgth; and E=
next. This is called multiple-effect evaporation and is
evaporator economy defined as the ratio of vapor pro-
accomplished by lowering the absolute pressure-and
duced to steam used, kg vapor/kg steam.
hence the boiling point of the liquid as it goes from the
For example, if the efficiency of each effect were
first effect to the second, and so on.
100%, the economy, E, would equal the number of ef-
From two to six effects may be used, depending on the
fects. However, heat losses and latent-heat differences
nature of the feed and product. The absolute pressure
cause the economy to decrease as the number of effects
decreases from the first effect to the condenser. The low
increases.
absolute pressure in the last effect is maintained by a con-
The feed to the evaporator is normally measured volu-
denser exhausting into an air ejector. In most applica-
metrically, e.g., using magnetic flowmeters. Therefore,
tions, water is the evaporated solvent. Typically, it is re-
the mass flow is related to the volumetric flow by:
moved by a direct-contact condenser. The steam-jet
ejector removes air dissolved in the feed and also pro- Wo -poFo (17)
vides vacuum for startup.
where Fo = volumetric feed rate, IJh; and po = feed
Fig. 3 shows a double-effect evaporator with the com-
density, kg/L.
plete feed forward-control system. The objective is to Substituting Eq. (16) and ( 17) into Eq. (15), and solving
maintain a constant concentration in the product. The
for the steam flow (i.e., the manipulated variable) results
two major load variables are the solids concentration in
ln:
the feed, and feed flow. The steady-state model for the
feed forward-control computation is obtained by using
both material and energy balances.
ws-iFopo (1 -#) (18,
The amount of solids entering each effect must equal
Here, A* indicates the desired weight fraction of the
the amount of solids leaving, because the vapors of each
are composed only of solvent. Moreover, the total mate- product.
Fig. 3 shows the instrumentation for implementing Eq.
rial entering each effect must also equal the material
leaving. Thus, for the first effect: (18). The feed density, po, and the feed solids weight-
fraction, #o, are related. Therefore, the signal characte-
Wo#o -WIXI rizer,/too), is used to obtain the calibration curve of #o vs.
Wo - Wl + Vl Po.
The feed-flow signal, Fo, is dynamically compensated
For the second effect: with the lead/lag function. It is set to a predominant lead
Wixi = W2X2 function because the product density does not respond
with equal speed to changes in the feedrate and steam
Wl - W2 + V2
flow. Changes in steam flow produce a slower response
where Wo, Wi, W2 = mass flow of feed, and of first and because of the thermal time lags associated with the heat-
second effects, respectively, kg/h; #o, #i, x2= exchange surfaces. Hence, simultaneous increases in
weight-fraction of solids in feed, first and second effects, feed and steam flowrates produce a temporary reduction
respectively, kg of solute/kg of solution; and Vi, V2 = in density.
mass flow of vapors in first and second effects, respec- Since steam is the manipulated variable, lead action is
tively' kg/h. necessary to minimize the dynamic deviation of the
Combining Eq. (8) and (10), and solving for W2, yields:
product density from the desired setpoint. The dynamic
W2 = (Xo/X2)Wo (12) compensator in this case is similar to that for the heat
exchanger shown in Fig. 2c. Feed-density dynamic com-
Combining Eq. (9) and ( 11), and solving for Wo, yields:
pensation is not included, because the density normally
Wo=W2+Vi +V2 (13) does not vary as fast as the feed flow, and is also fairly
slow-varying in relationship to the residence time of the
Substituting Eq. ( 12) for W2 in Eq. ( 13), and combining evaporator.
Wo terms, results in: Feedback trim is provided by the density controller
that manipulates the desired setpoint, ##, in the feed-
wo[l +¥] -vl+v2 (14, forward model.

®0
Steam, WS

Moisture control for adiabatic dryers Initially, the wet feed goes through a period of con-
stant-rate drying, during which the surface of the solid
Feed forward control has been achieved for adiabatic material is totally covered with liquid. The mass of liquid
dryers by using an inferential steady-state model, along transferred from the surface to the hot air is replaced
with dynamic compensation. Fig. 4 shows a rotary-type continuously by liquid from below the surface of the rest
adiabatic dryer, in which the fuel is burned in the com- of the material. After a certain period of time in the
bustion chamber to heat the ambient air. The hot air dryer, surface moisture is evaporated because the rate of
enters the dryer, where it comes in contact with wet feed. mass transfer of liquid to the surface begins to be less
The hot gases and the feed move parallel to each other. than the mass-transfer rate of liquid from the surface.
Heat is transferred from the hot gases to the wet material Thus, the material enters what is called the "falling rate"
to remove moisture. Moist air is exhausted to the atmos- of drying.
phere,andthedriedpred.uctisremovedatthedischarge Shinskey [J2] provides a derivation of the moisture in
end of the dryer. the falling-rate recton by combining heat- and mass-

®1
Eq. (20). ete erature differe
between ds. The rate which water lea
the surface of the solids and enters the air is proporti
to this temperature difference.
Now consider what happens as the evaporative
(Fig. 4) increases with a rise in moisture of the feed, a
the outlet temperature decreases. To restore the ou
temperature to its setpoint value, the outlet-temperat`
controller raises the setpoint of the inlet-temperat`
controller, thereby increasing the inlet temperature.
The increased inlet temperature is now used to
crease the setpoint of the outlet-temperature control
through the lag and the ratio-plus-bias station. This I
in outlet-temperature setpoint heightens the drivi
force (i.e., increases the rate of water removal) in ort
to meet the higher evaporative-load requirement. Th
the material exits the dryer at the desired value of prt
uct moisture. In this manner, the inlet temperature
the dryer is used to infer load changes.
Increases in feedrate also result in a boost in the ev
transfer relationships in the dryer. The result is: orative load and are controlled in the same fashion as
T-Tw rise in feed moisture.
(19) In describing the action of the control system thus
a positive feedback feature is evident in that elevating
where x = produc( moisture, kg liquid/kg wet solid; inlet temperature to the dryer causes the outlet-temp
#c = critical moisture at berinning of falling-rate zone, kg ture setpoint to rise, and causes a further increase to
liquid/kg wet solid; G = air flow, kgth; C = heat capacity inlet-temperature setpoint. This can cause long-
of air, kcal/(kg)(°C); H = latent heat of id, kcavkg; shon-term instability. I.ong-term instability will occur
y = mass-transfer coefficient, kg/(h)(ELH; (OC); A- in Eq. (20) is set at a greater value than the steady-si
surface area, m2; T! = inlet air dry.-bhlb temperature, °C;
gain of the process. (The steady-state gain is the cha`
To = outlet air dry-bulb temperature, °C; and r„ = air in outlet temperature divided by the change in inlet
wet-bulb temperature, °C.
perature.) Thus, for long-term stability, fa must be set :
The terms before the logarithm are generally constant value less than the steady-state gain of the process.
for a given dryer and product. Air dry-bulb tempera- Short-term instability is caused when the setpoint
tures, I,. and To, are easily measured. The wet-bulb ten- the outlet-temperature controller is changed faster tl
perature, however, cannot be measured readily inside the actual inlet temperature. The lag unit in the positi
the dryer, because the surface of the bulb is easily cov- feedback loop provides the adjustment to retard t
ered with solids. change in setpoint long enough for the actual outlet t€
The inferential-control model then has to maintain the
perature to respond to the change in that of the act
relationship between inlet- and outlet-temperature con- inlet. In this way, dynamic or short-term stability
stants, so that the logarithmic term in Eq. ( 19) is also con- achieved.
stant. This provides for constant dryness, and hence, The moisture controller is used to provide long-te
constant product moisture. correction by var)ring 4 in Eq. (20), to maintain the acti
For a reasonable range of operation, the following moisture at the desired setpoint. It is implemented in t
straight-line model [J,2] with adjustable gain (slope) and manner because the inferential model is based on
bias is used: measurement, and air moves through the dryer fas
To= kT.+ b (20) during transient load changes, and therefore provit
more effective control.
where A = gain, °Crc; and a = bias, °C. Normally, the mass flow and moisture of the feed,
Variations in ambient-air humidity affect this compu- well as the air flow, are no`t economically feasible to m€
tation, particularly at inlet-air temperatures in the neigh- ure so as to provide more effective control. The in fen
borhood of 180°C, and below. In this case, dewpoint tial-control strategy presented here is a practical way
correction (based on the psychrometric chart) is imple- improving control by using the readily available air-tc
mented by varying the bias term, a, in Eq. (20) as a func-
tion of the dewpoint-temperature measurement. Typi- perature measurements.
The next article in this CE Refresher win appear i
cally, direct-fired dryers operate at inlet temperatures Dec. 10, 1984 issue, and will cover other advanced
sufficiently above I 80°C so that the effect of variations in trol techniques.
ambient humidity is negligivle.
S~ Dantws,
The objective of the control system is to control prod-
uct moisture when major load variations such as feed- References
rate and feed moisture occur. The control system accom- 1. Shinskcy, F. G., "Praccss Control Systems," 2nd ed., MCGraw-Him, New
1979.
plishes this by varying the driving force for drying as 2. Shinskey, F. G., "Energy Conservation Through Control," Acadelrfe Pliess,
these load variables occur, by using the relationship in York, 1978

ea
Advancedcontroltechniques
for distillation columns
Predicting and applying the control strategies for distillation columns
will depend on the interactions among the process variables and the
economics, and on equipment and process limitations.

Rj,clrard A. Rys, The Fouboro Co.*

I The distillation process is a perfect example for illus- Doubling the feed will double reboiler heat require-
trating advanced control techniques because: ments, assuming other variables constant. Increasing the
I Distillation-column operation involves high energy number of trays will reduce energy use, as will increasing
consumption, and a corresponding large increase in the tray efficiency.
product value compared with feed value. Many columns The tray efficiency, E, is the number of theoretical
in current operation were engineered and constructed trays divided by the number of actual trays, 71. It is a
during a period of low energy cost, and most are measure of the extent that equilibrium is reached. The
equipped with only basic measurements and controls. number of theoretical trays is what counts for energy
I Distillation columns are difficult to control because consumption.
of significant interaction between the controlled varia- The relative volatility, c¥, for ideal solutions (Raoult's
bles and upsets. Certain of these variables should be con- law) is the ratio of vapor pressures of the two liquids
trolled at optimum values. However, variable constraints being separated. For example, ordinary water (boiling
often shift the optimum operating point. point 100°C) and heavy water (boiling point 101.4°C)
The techniques to be discussed in this article are based have a relative volatility of about 1.05 at 100°C. Relative
on practical field-proven approaches that produce ac- volatilities near 1.0 produce little separation per tray and
ceptable returns on investment. require a large number of trays to achieve a high purity.
Conversely, methane (boiling point -164°C) and
Basic concepts for distillation heavy oil (boiling point 200°C) have a relative volatility of
A distillation column separates two, or more, liquids about 20,000 at 100°C, and can be separated to very high
on the basis of differences in boiling points. For simplic- purities in a single tray or flash tank with only sufflcient
ity, our discussion will be confined to a column handling energy to vaporize the methane. Increasing the relative
a single feed and two products. The lighter, low-boiling volatility can usually be accomplished by lowering pres-
product is referred to as distillate, D; while the heavier, sure, which will reduce heat input requirements.
high-boiling product is bottoms, 8. The separation factor, S, is defined as:
The energy required to separate two liquids is essen-
tially related to the feed rate, F, number of theoretical y(1 -#)
(1)
trays, 7tE, relative volatility, cr, and the particular separa- #(I -,)
tion' S.
where x is mole fraction of "lights" in the bottoms, and )
Let us briefly look at how each variable affects heat
is mole fraction of "lights" in the overhead.
consumption. Although most of this consumption (usu-
The heat input to most columns is nearly proportional
ally steam) occurs primarily at the reboiler, some heat
to the logarithm of the separation. Reducing separation
energy may be required for a feed preheater. Heat sup-
will reduce energy requirements when other variables
plied to the reboiler boils some of the feed, F, and reflux, are constant. When S is equal to 1.0, no separation oc-
i, to produce a vapor flow, V, in the column. Heat-en-
curs. A separation of 9,800 will allow both distillate and
ergy consumption is essentially proportional to feed rate.
bottoms to be 99% pure.
*To meet the author, see Cho. E7®g., Aug. 20, 1984, p. 158.
Economic benefits
In an ordinary two-product binary or multicompo-
i;§e:;;C|§:[i:9:8:3[,:Sp|e#hE:as:t:i;inepi:1;lo:,:199¥4;,EPS;7;,?E:r%r;,J[jitu¥V:y!?:;i:;;P:i:;::::
8, Aug. 20,1984, p.151; Part 9, Oct.15,1984, p.103. nent distillation column, economic benefits resulting

e3
the pressure controller should be set at a safety margin
slightly higher than the low-pressure constraints to ab-
sorb short-term pressure upsets. The condenser con-
straint is variable, usually because of temperature or flow
variations of the (`ooling fluid. Coolant temperature
often varies from night to day, and with the seasons.
Fig. 1 shows a control system to "float" pressure in a
depropanizer column. A valve-position controller [J] is
used to implement this system, which should be set high-
perhaps to 9097o of valve signal. To keep pressure low,
the cooling-water valve should be kept open. This will
allow good pressure control that floats at a constant
safety margin (10% of valve travel) away from the con-
denser constraint.
2. Pulse duration Of feed rate-1£ the feed rate to a cch-
umn is very low and if feed and product storage tanks
are very large, it may be economical to operate the col-
umn at a high rate part of the time, and to shut it down
until feed-tank inventory is refilled. Ambient insulation
losses are reduced, and the column would likely operate
from advanced control can be achieved in three ways: with much higher tray efficiency at the higher rates.
I. Maintaining a constant separation between distillate On the negative side, the utility cost to maintain total
and bottoms composition, but reducing energy costs. reflux operation during several hours of startup can be
2. Maintaining constant separation and energy costs, expensive, especially if the run time is just a few days.
while varying the column material balance to increase Also, inventor\' flexibility is reduced because of the step-
the total value of distillate~plus-bottoms products. wise operation'. For instance, a sudden influx of feed into
3. Changing separation, with a corresponding change a feed tank that is nearly full could present problems.
in energy costs, to increase the value of the distillate and Similar problems `\'ould exist in product surge or in-
the bottoms-minus-energy costs. ventor}. tankage. This t}'pe of pulsed operation can be
For our discussion, varying the feed rate and feed accomplished profitabl}' if the right circumstances exist
composition will not be considered, and comparisons will and if scheduling is carefull}' planned. The feed rate to
be made per unit of feed. In most cases, the feed rate to the feed tank is still a dependent variable in pulsed oper-
a column is a dependent variable (limited by upstream or zition. The feed rate to the column itself would resemble
downstream processes or inventories). The feed rate a low.-frequenc}' square wave. The duration of each
may be set independently at an optimum or maximum square-``'a\.e pulse `\'ould vary in proportion to feed rate.
rate. Here, feed rate is considered a dependent variable 3. Eliiiiina[ion Of b[eirding-Another way to reduce utll-
that is not adjustable by the column-control strategy. it}. costs and still produce distillate and bottoms product
of specified purity. demonstrates the need for good regu-
Reducing energy costs lator\- control. To understand the significance of this
It might appear that a distillation column could be method. it is necessar\. to understand how real processes
operated only one way (i.e., with a fixed energy cost) to actuall}. produce on-specification product.
obtain products of a particular purity from a given feed. Let us suppose a column is to process loo volumes of
This is not true. Three alternative methods can be used: feed, \\'hich contains 507c light component A and 50%
I. Floating Pressure control-Go\umn pressure c2\n heav}' component 8. The distillate is to be 95% A and 597o
often be varied. Since operating at lower pressure almost 8, while the bottoms is to be 95% 8 and 591o A. A simple
always improves relative volatility, a lower column pres- material balance sho``'s that loo volumes of feed will pro-
sure with a corresponding increase in relative volatility duce 50 \'olumes of distillate and 50 volumes of bottoms.
increases separation on every tray. And, the improved Let us also suppose that the distillate and bottoms are
separation is free. sampled and anal}'zed e\'er}' time 10 volumes of feed
To prevent products from becoming too pure, the re- leave the column. Table I gi\7es an example of how this
flux and boilup must be reduced to return the separation column could produce 50 volumes of distillate and 50
to its original value. In actual practice, achieving savings volumes of bottoms that meet the specified purity. The
is not as easy as it seems. It is seldom possible to merely variation of the individual samples from the specified
lower setpoints of the pressure controller. Pressure af- compositions is an indication of how well the regulatory
fects tray hydraulics, temperature measurements, vapor composition controls are controlling. Reboiler energy
velocity, and condensing and boiling temperatures. If consumption will be reduced whenever the distillate and
pressure is to be changed, it must be done slowly, and bottoms composition can be held more steady.
compensations for temperature changes must be made. Operation can be closely approximated by assuming
As pressure is reduced, condenser limitations aLre likely that the compositions were steady at the analyzed value
to appear, which will determine how low the pressure for the duration of the sample period. In this way, the
can be reduced. It is seldom acceptable to operate a col- energy consumed during each of the 10 sample periods
umn without pressure control, because upsets to con- can be calculated separately.
densers could easil}' affect column compositions. Instead, The total energy consumed to produce 50 volumes of

e4
effect to the reactor. Solvent values
depend mostly on how the reactor
performance changes with the con-
centration of solvent.
The value of finished product
often drops sharply if specifications
are not met. Often, there is little or no
incentive for exceeding specifications.
When the value drops sharply, eco-
nomical operation can be achieved by
making the product no purer than
specifications. Operators often refer
to this as "crowding the spec," mean-
ing that overpurification should be
minimized. In this case, the value
added would change abruptly with
material balance.
Most specifications are composi-
tion-related, and the safety margin
for overpurification of finished prod-
ucts is based on the ability to control
compositions. In order to ``crowd the
spec" without going off-specification
(which may contaminate product-
storage tanks and require expensive
rerun), regulatory control of compo-
sitions must be accurate. The over-
purification safety factor is based on
experience. The safety margin will
from a bent flapper sticking to the nozzle on a locally be narrowed if the operator is convinced that off-specifi-
mounted pneumatic bottoms-level controller. The sticky cation compositions will be prevented by the control
flapper held the bottoms flow from the de-ethanizer System.
wide open for a period, then fully closed until repairs Intermediate surge tanks placed between the column
were made. The upset was handled with all controls in and product storage can be used to attenuate composi-
automatic. The bottoms compositions was tangibly af- tion variations. However, inefficiency will exist. Compo-
fected but responded to the upset with nearly quarter- sition variations should be prevented by better control
amplitude damping. and not merely attenuated.

Optimizing material balances Optimizing entire column


Shifting the column's material balance at constant sep- The remaining procedure for achieving economical
aration is another way to achieve economical operation. operation is the most general. For instance, a given feed
Value added per volume of feed must be determined for is to be separated into a distillate and a bottoms product.
various material balances to find the optimum. No particular product specifications are to be met. In-
For a particular separation, the fraction of the feed, F, stead, separation should be optimized, and distillate and
that is distillate product, D, can be varied from nearly 0 to bottoms should be controlled at optimum purities.
100%. Fig. 3a shows how the value added (value of prod- The material balance may be optimized for a particu-
ucts minus value of feed) may vary for a particular sepa- lar separation. As separation is increased to the limits of
ration as the D/F (distillate to feed) ratio varies. For one the column, the optimum balance will likely vary (Fig.
3a). The relationship between value added and separa-
particular separation, the maximum value added will be
achieved at the optimum D/F material balance. If more tion (Fig. 3b) must be determined from the locus of opti-
separation is provided (by increasing boilup or improv- mum material balances. Just increasing the purity of one
ing efficiency), the optimum D/F value may change. product while keeping the other constant will not allow
The economic data for this type of analysis are seldom for operation at maximum profit.
adequate. What is required here is the value of both dis- Separation is directly related to energy cost (Fig. 3c).
tillate and bottoms products as a function of purity. Val- Utility cost per unit of feed approaches infinity as separa-
ues of intermediate products in a process plant are sel- tion approaches the maximum achievable by a particular
don accurately assigned. Economic data to determine column. Maximum separation occurs at total reflux. As
it is approached, capacity to handle feed nears zero.
product values are dependent on upstream and/or
downstream processes. Fig. 3d shows how profit (value added minus utility
A distillation column (recycling solvent to a reactor) costs) increases to a maximum as separation is varied.
can return the solvent over a wide range of purity by The optimum separation will define the most profitable
varying the material balance at constant separation. The composition-controller setpoints.
value of this solvent varies with purity, depending on its CoPczc¢try /3.7„iJczl3.o7as-The previous discussion has as-

®e
sumed a constant feed rate and adequate column capac- Since temperature measurement responds more slowly
ity. However, when equipment limitations occur in an than does pressure measurement, compensation may be
unlimited market, optimization becomes more complex. in error during transients. Also, the pressure measure-
The instantaneous profit (S/h) must be maximized. Max- ment may not be made at the temperature-measurement
imizing profit per unit of feed is not sufficient.
point, introducing an error. In addition, the slope of the
Since utility costs and product values may vary with vapor-pressure curve (needed for calibration) will vary
feed rate, the profit function can no longer be described with concentration and temperature.
in a two-dimensional graph. A three-dimensional surface
relating profit, separation and feed flow is needed to de-
termine the most profitable operation. Since feed-tray
Other methods for composition
location and preheat are both optimizable, they must also Differential vapor-pressure transmitters can often be
be factored into the analysis. used to provide an inherently pressure-compensated
Most operating columris have relatively poor regula- measurement of temperature. The instrument can be
tory control. Yet, it is a prerequisite for optimization and filled with a fluid having the same composition as that at
is needed for economical operation. the temperature-measurement location. The instrument
outputs a signal that is very sensitive to concentration
The optimization of distillation columns can take
changes, and is virtually unaffected by pressure.
many forms, and can be successful if two requirements
are met. First, a set of regulatory controls must be pro-
In many columns, the effect of third components
vided to continuously control at the optimum-composi-
makes temperature-based measurements unsuitable for
composition control. Variations in concentrations of
tion setpoints within the constraints of the process. Sec-
third components can significantly change temperatures,
ond, the optimizer must be able to accurately determine
causing upsets to the concentrations being controlled. In
the real optimum of the process.
some columns, dozens of components can be involved.
Off-line optimization is generally adequate. However,
online optimization has been attempted. In either case, Columns separating close-boiling components may not
be able to use temperature-based measurements because
the real optimum is difficult to determine because the
of limitations in instrument accuracy. In such cases, ana-
comprehensive economic data that are needed depend
lyzers are needed to verify compositions.
on a large number of unknown variables.
Online analyzers are commonly used in closed-loop
Composition measurement control when upsets are common and monetary incen-
In an ordinary distillation column, five independent tives are large. Online analysis has the advantage of pro-
variables must be controlled: pressure, bottoms level,
viding a fast, accurate measurement for prompt feed-
accumulator level, bottoms composition, and distillate back correction. Long deadtimes in the sample transport
for both online and off-line analyzers, or deadtime aris-
composition. For the most part, the measurement and
ing from shared analyzers, will degrade control. Feed-
control of pressure and level are reasonably straightfor-
back-controller error increases as the square of the total
ward. Composition measurements are more difficult.
loop deadtime. Deadtime should be minimized for both
Temperature vs. composition online and off-line measurements.
Most columns use temperature measurements to indi-
cate composition. Every column has a temperature pro- Regulatory control
file that is hottest at the bottom and coolest at the top. Distillation columns are subject to many upsets includ-
Generally, a mixture of two components in equilibrium ing, but not limited to, variable feed rates and composi-
between liquid and vapor at a particular pressure will tions, variations in product-purity specifications (setpoint
have only one degree of freedom (Gibbs phase rule). changes), sudden rainstorms dumping cool water on an
This means that the temperature and the composition air-cooled condenser, tumbleweeds getting caught in
are directly related, and a particular temperature can condenser fans, lightning setting relief valves on fire,
occur only at a corresponding particular composition. and water in the feed forming ice (hydrates) that com-
Liquid or vapor temperature in a distillation column pletely blocks off the top tray with slush and prevents
will be a good indication of composition if the: reflux from reaching lower trays. Some column upsets
1. Pressure at the point of temperature measurement are extraordinary and the column controls cannot ade-
remains relatively constant. quately respond to them.
2. Temperature is not significantly influenced by the Many upsets can be reduced or eliminated by control-
presence of third components. ling interrelated processes. Upsets in feed flow, feed
3. Temperature sensor is located in a position where it composition, steam pressure, or coolant temperature can
is in temperature equilibrium with the stream whose often be reduced most economically by controls else-
composition is desired. where in the plant. Upsets can be handled by an infinite
4. Temperature-measurement system has sufficient variety of feed forward and feedback techniques (see
accuracy to provide adequate resolution. Parts 4, 6 and 8 of this series).
If these are adequately met, the temperature scale It is not uncommon for distillation-column controls to
could be renumbered in units of concentration. exhibit sustained oscillations, possibly triggered by an
The most severe problems occur with interferences upset. Interaction between two composition (or tempera-
due to pressure and third components. Temperature ture) controllers can cause this instability. Since distilla-
measurements can be pressure-compensated by using a tion columns are multivariable processes, interaction will
simple summer, but some difficulties must be overcome. always be a problem.

e7
Potential pairs for relative gains Table ll
--I--
BD
FF
Distillate-composition variables, my "E,(1 -J)

I0+iE*!£`!'3,ain 0 i S(L/D\ S(O/V) I S(i/V)


2(%")ky-x,
0(8)
0.32 0.5
DI
(5)
V 0.75 3.44 1.52 v_I
L/B 0.85 1.57 1.21 i+1
V/B 0.99 1.02 1.01
where D/F is the distillate-to-feed ratio, B/F is the bot-
toms-to-feed ratio, D/V is the distillate-to-vapor ratio, and
€ is a column-design parameter.
By substituting the numerical values for the example
into Eq. (2) through (5), we obtain D/F = 0.5, B/F = 0.5,
Increasing reflux with constant boilup will influerice
c = 0.35, and D/V = 0.29.
both distillate and bottoms composition. Increasing
To quantify the relative gains, A, for various control
boilup with constant reflux also affects both composi- strategies, we must apply the following relationships:
tions. Columns with dual-composition control, columns
with multiple sidedraws, and heat-integrated columns
are likely to experience significant interaction because #lD--I--1-6 (6,
boiling and reflux are manipulated simultaneously in
dyl J(l -J) -1=o (7)
one or more column sections.
a#ls #(I-#)
Simple strippers or absorbers are unlikely to experi-
ence severe interaction problems, because only one com-
position is controlled. Typically, boilup is manipulated to
control the liquid composition in a stripper, as is reflux to
#lL-= - 2.08 (8)

dy I a--6€[1 + (D/i)] -2.94 (9)


control the vapor composition in an absorber.
axiv 1 -€[l + (D/I)]
Relative gains
11 a--6€(F/B) -5.67 (10)
Interaction between two composition controllers in a 6xllJB 1 -€(FIB)
distillation column can be alleviated only partially by con-
troller tuning. Shinskey [2] has shown that various com- dy I Cr -6E[1 + (D/I)](F/B) -94 (11)
binations of variables can be manipulated with consider- a* lve 1 -€[l + (D/i)](F/a)
ably less interaction than can be done with conventional The relative gain, A, for various pairs of manipulated
reflux and boilup controls. variables is calculated by using:
Material-balance and reflux-ratio controls have often,
but not always, outperformed conventional reflux/boilup
controls. Shinskey has applied the relative-gain approach
to distillation columns to quantify the interaction that will
Am'm*

1-# (12)

exist in a large number of different control strategies (see The results for the example are summarized in Table
Part 8 of this series). 11, where the variables that may be manipulated for dis-
The benefits for predicting the degree of interaction tillate-composition control are listed across the top, and
of a control strategy applied to a particular distillation those available for bottoms-composition control are listed
column are obvious. An organized method for determin- down the side.
ing the basic structure of the control system helps to en- Those pairs of variables that will produce unaccepta-
sure that the selected control strategy will be effective. ble values or that have unfavorable dynamic properties
Changing control strategies on an operating column, have been discarded. Each intersection identifies a po-
with or without a shutdown, is sure to introduce upsets tential pairing. The relative gain that quantifies its inter-
and will increase the cost of the control system. Few plant action can be calculated via Eq. ( 12). Pairings having rela-
managers would allow experimentation on an operating tive gains closest to I.0 are the least interactive.
column in order to adequately compare the perform- The calculations of these relative gains is made by
ance of a dozen or so potential control strategies. using the simplified model developed by Douglas, ]af-
To show how relative gains are determined, let us con- frey and MCAvoy [j]. A slight improvement in accuracy
sider a distillation column having theoretical plates, can be obtained with tray-to-tray calculations. However,
"I = 30; a reflux-to-distillate ratio, i/D = 2.5; mole-frac- a simple model has ample accuracy for this purpose and
tion of "lights" in the bottoms, # = 0.05; mole-fraction of is fast and easy to use with hand-held calculators.
"lights`. in the overhead, } = 0.95; and the mole-fraction
Variations can be added to each of the basic control
of lights in the feed, z = 0.50. strategies. For instance, column-pressure control can
The following equations apply for this distillation: have many different arrangements, or constraint over-
D z.-xi
rides can be added.
Fig. 4b shows the control strategy that will have the
(2)
F yt-x. least interaction as calculated for the column in the ex-

ee
techniques such as I.eadback checks on output compo-
nents impro\'e the abilit}' of the system to confine the
effects of a fault to the smallest possible area. The ability
of the system to continue functioning acceptably in the
Category Name Description
event of a fault is increased by fault~tolerance measures
Control PID Proportional, integral, derivative, including
such as redundancy.
absolute, deviation and output alarms,
If plant process control is viewed as a multilayered
adaptive tuning.
entity (basic PID feedback loop, advanced control strate-
DGAP On-off control,with tri-state (dead band),or
gies, unit and plant control strategies), the application of
microprocessor controllers now allows the user to imple- pulse-duration control.

ment levels of control that heretofore were thought to be AMB Automatic/manual switch with bias. Bate
available or implementable only with host computers. can be applied to measurement, track
capabil ity .
Application to distillation INT Integral only, control, skip feature allows
The power of the microprocessor-based controller t,o per- use as sample data control.
form advanced control strategies with enhanced loop se-
DOUT Collects upto8 single-bit logicsignals,and
curity, as well as single-loop control, is illustrated in transfers them to an output component.
Fig. 4. Here, a typical depropanizer in a refinery is de-
Input and AIN Signal conditioning and alarming, will filter,
signed to separate propane from a multicomponent gaso-
conversion zero, span.
line feedstream. The bottoms stream is fed to a
deisobutanizer column where there is a "hard" propane NONL Operatesasan extenderforastandard PID
block for control loops having nonlinear
specification on the isobutane product.
gain.
Plant economics dictate that the depropanizer over-
head-distillate product contains a minimum amount of CHAP Linear segment x-y calculator,
isobutane because it has a higher value than propane. 10 segment.

Therefore, both top- and bottoms-composition control PCIN Pulse accumulation with batch feature
systems are required. logic bit out if accumulation exceeds batch.
Column reboil heat is supplied by a steam reboiler;
D|N Brings in 8-bit word from hardware, and
overhead vapors are condensed in a partially flooded, compares with Lip to 16 user patterns,
water-cooled condenser; and the overhead receiver is alarm on match.
operated liquid-full. The heat-removal rate is controlled
Computation BTIO Multiplier with input/output scaling, bias
by manipulating the flow of liquid from the receiver,
and alarms.
which determines the amount of liquid backed up in the
condenser tubes, and therefore, the heat-transfer area SSEL Of 8 inputs, select the highest, lowest,
median or average value.
available for condensation.
In this example, it is economically desirable for the BAMP Dual adjustmentramp (upordown) with
depropanizer (operated in parallel with another column) reset, follow and hold features.
to take as much of the total feed as possible. It is known TIMB Sets output bit and leaves it set for speci-
that the overhead condenser is the process constraint or fled time period, has two-stage and repeat
bottleneck. Frequently, during the summer months, the features.
column pressure increases because of the condenser con- SUM 4-input adder with input/output bias and
straint, and the operator is forced to reduce the column gain controls.
feedrate.
ACuM Variable-scale rate integrator/totalizer has
The operator's clue that the condenser is overloaded is output limit alarm for batch.
the appearance of a liquid level in the overhead accumu-
CALC Multiple-input-35-step computation, 3
lator-i.e. , the pressure rises even though the pressure
memories for store-retrieve operations,
controller has increased the accumulator flow until all of 3 separate outputs.
the heat-transfer area is exposed. It would be desirable
Logic GATE 16 logic inputs, 8 logic operators speci-
to automatically operate at the constrained feedrate
tied for 8 (2-input) logic gates.
without upsetting the column.
SEQ Produces 8 numbered 8-bit word patterns
Feed forward and composjlion control in a stepped numerical sequence.

An appropriate control strategy for this column might SWCH Double-pole, double-throw toggle switch.
include a feed forward system (Fig. 4) to set heat input,
ALBM Absolute, deviation, rate alarms;absolute,
with feed flow as the measured load variable. Because rate output limiting.
the feed is primarily liquid, feedrate changes are pro-
Dynamic LLAG Causes gradual (lag) or rapid (lead) output
jected toward the bottom of the column. compensation change in response to input.
To maintain a dynamic balance, changes in heat input
must be'delayed until the load change reaches the column DTIM Adjustable-length tapped delay line; 25

base. Delays are due to the hydraulics of the trays below taps, user-selectable delay.

the feed tray. In effect, a feed forward system is a model


of the process because it consists of both steady-state and
dynamic elements.

7a CHEMICAlu ENGINEERINGffEBRUARY 4,1985


v..,;..-'. -y_ -.- . ..-..-.-fa` ,--. .`..-i.- `*---

'Communi cation bus

\ RE
\1

idllllllll
\

Local
communication
device
I ~----Bus

ae
11111------iliii
-
ae;
-i-i---:.f#

^7jy
yl[=FHrTl \ TT= /I


•.
iiiiiiiiiiiiiiiiiiiiiiiiiiiiiiR-
= • •..
.
.
..-. .

•...
.

. .
. .

..
---.

.-¥#;
.. •.. . .. -¥REI

•y `' . .' ` . \ ........ ....`.``,£:`.:`#¥,rs:`%%.,.`ifees

The feed-flow signal, F, is dynami-


cally compensated lfrtJ], with a dead-
time and a first-order lag function.
This signal is multiplied by a gain
factor, (Q/F), to generate the reboil-
er's steam flow setpoint, Q.
For precise bottoms-composition
control, the gain factor is trimmed by
the bottoms-composition controller.
An onstream chromatograph is used
because the variable, multicompo-
nent nature of the bottoms stream
prohibits use of an inferential tech-
nique such as temperature measure-
ment to determine composition. The
chromatograph ensures that the dis-
tillate product in the downstream
deisobutanizer column will contain a
constant percentage of propane.
Derivative action is desirable in
the composition controller, AC, in or-
der to compensate for the exponen-
tial lags in the column. However, the
chromatograph analysis is performed
at discrete intervals of time (2 min),
and therefore produces a concomi-
tant series of step changes in the AC-
controller measurement signal.
This situation would normally pre-
clude the use of derivative action in a
continuous PID controller. To over-
come this, a second-order lag func-
tion, /rtJ, is used here to digitally
smooth the chromatograph signal to
allow derivative action in the AC
controller.

73
Rellux ratio and temperature conlrol
Close pressure control is mandatory for good composition
control. The column overhead pressure controller (PC) Description
manipulates both reflux and distillate flows (I+D) out of
Automatic/manual (can be latched).
the overhead condenser for a fast-responding control
Remote/local s,etpoint (can be latched).
loop.
Output tracks integral-feedback signal.
Variations in distillate composition are minimized by a Local setpoint tracks measurement
reflux-ratio control system developed by Ryskamp [J]. in REM.
The pressure-controller output (I+D) is connected to Error-squared action.
both a multiplier and a summer. The other signal into the Nonlinear proportional gain.
multiplier is the fraction of total overhead [D/rL+DJ] Sampled-input operation.
that is to be withdrawn as distillate product. The output Preselected system-flunk options.

of the multiplier is the desired distillate flow, D. The ac-


tual distillate-flow signal goes to the summer where it is Block input bad (out of signal range).
subtracted from the desired overhead flow (I+D) to ob- Absolute alarm on input (high/low).
tain the desired reflux flow, I. Deviation alarm setpoint vs. measured
With this system, if the column feedrate increases, the variable (high/low).
Controller output is clamped (high/low).
boiling rate and column pressure will increase, and both
Open loop (block is in track or manual).
reflux and distillate will increase in the proportion fixed
A setpoint increase will have no effect
by the signal, D/rL+DJ. This removes additional propane
on valve.
from the column. Because the reflux ratio is maintained A setpoint decrease will have no effect
constant, the distillate composition will tend to remain on valve.
constant. Process input rate-of-change alarm.
A tray-temperature measurement in the upper portion
of the column should be sensitive to C3 /tc4 composition
changes, because tray composition at this point is essen-
tially a binary mixture. Therefore, overhead composition
is controlled by a tray-temperature controller (TC), • Would the operations department prefer to have the
which sets D/rL+DJ. operator decide when to place the AC controller in autor
With this system, if the tray temperature increases matie after the chromatograph is repaired?
(high 97otc4), the controller will decrease the D/(I+DJ Security questions to be answered for general control
factor, which will decrease distillate rate and increase are:
reflux rate, to drive the {C4 back down the column. De- • Should the AC controller be protected from integral
coupling the pressure controller and top-composition con- saturation (i.e., reset windup)?
troller is explicit in the design. • What will be the startup procedure for the bottoms
cascade-control system so as to ensure bumpless transfer
Conslrainl control to the primary AC controller?
An overhead-receiver level controller (LC) is used to de- The answers to these and other questions about con-
tect and operate at the constraint for condenser capacity. trol security and operator interface determine the oper-
The LC output signal is compared with an operator-en- ating philosophy. For our example, the philosophy will
tered column feedrate. The lowest signal is selected as be to use operator-entered high and low absolute limits
the setpoint of the feed-flow controller (FC). on the chromatograph signals as a validity test for con-
During normal operation, the receiver level will be trol. An alarm (audible and visual) will be received at the
above the LC setpoint (here, 9597o), and the LC output operator command center if one of the chromatograph
signal will not be selected as column feedrate. When the signals is determined to be bad.
condenser becomes limited, a level will appear in the re- The AC controller will automatically go to a manual
ceiver; the LC output signal will decrease until column state, and freeze at the last good output (Q/F) signal. The
feedrate is selected as the manipulated variable for the feed forward loop will continue to manipulate the steam-
LC constraint controller. flow controller's setpoint for changes in feed-flow load.
Based on periodic laboratory samples, adjustments to
Operating philosophy the Q/F factor will be made manually through the AC
Before our function-control strategy (see Fig. 4) can be output signal.
implemented, an operating philosophy must be estab- After the system has determined that the chromato-
lished. The operator interface with the strategy must be graph signals are validated for control, the operator win
defined, during both normal and degraded operation. For be notified. The AC controller will remain in manual
example, if the bottoms chromatograph should fail, the (1atched) until the operator places it in the automatic
following questions must be addressed: mode.
• How will the failure be detected and the operator Primary controllers, with integral action, are subject
notified? to integral saturation (reset windup) in cascade-control
• How will the AC controller prevent itself from dis- loops. The secondary controller may be unable to re-
rupting the process? spond to the directions of the primary controller for rea-
• Should the feed forward loop remain on control, and sons such as a sticky control valve, or a secondary con-
if so, how will the operator adjust the Q/F factor? troller in either manual or local setpoint mode. The

74
?3
Feedfoiward and composition control
Fig. 5a is a microprocessor-based-controller block dia-
t, F, ' gram of the column's feed forward and bottoms-composi-
tion control system. Process inputs to the controller are
F,
shown on the left, and the one process output (steam
valve) is on the right. Each block is labeled with the algo-
rithm type and block number. Block logic signals are
BK Setpoint
shown as dotted lines.
valve
- • ffiaeeeffl
The feed forward system is depicted in Blocks 2, 3, 10,
and 11. Dynamic compensation on the column feed-flow
signal is accomplished in first-order lag Block 2 and dead-
time Block 3. Steam-flow setpoint (Q)' for PID Block 11
is determined by multiplying the compensated feed flow,
F, by the desired (Q/F)' in calculation Block 10.
Bottoms-composition feedback trim control is accom-
plished in Blocks 1, 4, 5, 6, 7, 8 and 9. Chromatograph
i--i- :i== inTRE signals %C3 and %¢C4 are received by AIN Blocks 4 and
5. The AIN Blocks allow the operator to set high and low
absolute-alarm limits for each signal.
The controlled variable (C3/¢C4) is calculated in CALC
Block 6 (1), and then passed through a second-order filter
(Blocks 7 and 8). 'two first-order lags in series produce a
second-order lag. This signal is the measurement to PID
controller Block 9, the output of which is the desired (Q/
primary controller will then continue to change its output F)' factor in CALC Block 10. Integral feedback (FBK)
(windup) as long as there is a difference between its set- for the AC controller is the actual Q/F, which is calcu-
point and measurement. lated in CALC Block 6 (2). (The CALC Block can have rip
The external setpoint of the secondary controller will to three outputs.)
be driven farther and farther from its measurement. Loop security is handled in logic GATE Block 1. The
When the secondary control loop suddenly becomes re- absolute-alarm status for each chromatograph input sig-
sponsive to its external setpoint, the process will be nal is input to OR GATE Block 1 (1). The output of the
upset until the primary controller can drive its output OR gate is connected to the MAN parameter of PID
back to the proper steady-state value. Block 9. As a result, if either chromatograph signal is in
Integral saturation can be prevented by incorporating an alarm state, Block 9 will be latched in the manual
the actual process variables, represented by the output mode.
of the primary controller, into the integral calculation of Bumpless transfer of the steam-flow controller (PID
the primary controller (external feedback). In a simple Block 11) to remote setpoint is accomplished in GATE
cascade loop, the secondary measurement variable would Block 1 (2). In Block 11, two status bits, open loop (OLP),
be used as feedback to the primary controller. The bot- and local setpoint (LOC), are inputs to the OR GATE of
toms-composition controller output, however, represents Block 1 (2). The output of the OR gate is connected to the
the desired Q/F factor for the feed forward loop. TRACK parameter of PID Block 9. If either status bit is
The proper integral-feedback variable for this control- set, Block 9 will go to the TRACK mode.
ler is the actual steam/feed flow ratio. With this refine- The startup procedure would be to first establish
ment, the AC controller will not saturate as it would if steam-flow control in automatic and local setpoint. Then,
the integral action were connected directly to its own the AC controller would be placed in automatic. Because
Output. of the tracking logic, the steam-flow external setpoint
Table 11 shows a partial list of the operation and status would follow the steam-flow measurement. The final step
modes associated with the PID algorithm for the control- would be to place the steam-flow controller in remote set-
ler used in this example. The TRACK mode is useful on a
point to close all loops.
primary controller to prevent a bump from occurring in
the secondary controller's setpoint when the cascade loop keflux ratio and lemperalure conlrol
is closed. In the TRACK mode, the primary controller's Fig. 5b is a block diagram for the column's overhead-
output will follow its feedback signal when the secondary pressure, reflux-ratio, and temperature control systems.
controller is in either local setpoint (LOC) or open loop Column-pressure control is implemented in Blocks 14 (1),
(OLP). 16, 17, 18 and 19. The pressure-controller output (PID
For example, the AC controller in TRACK will always Block 16) represents the desired total reflux and distill-
calculate an output signal, which when multiplied by the ate flowrate (I +D)'. This signal is multiplied by the de-
column feedrate signal will produce a steam-flow setpoint sired proportion of distillate flow [D/(I+D]' in CALC
equal to steam flow. When the steam FC is placed in Block 17 (1). The resultant product is the desired distill-
remote setpoint (REM), there will be no upset. ate flow, D'.
The combination of available algorithm types, opera- Desired reflux flow (L') is calculated in CALC Block 17
tions modes, and block-status information can be chosen (2) by subtracting the actual distillate flow from the de-
to meet other operating philosophies. sired total flow. Distillate and reflux flow controllers are

?e
PID Blocks 18 and 19, respectively. Integral saturation overhead-accumulator level transmitter. When the col-
is prevented in the pressure PID block by calculating an umn heat-removal demand has exceeded the condenser
FBK, which is the actual I,+D flow in CALC Block 14 capacity, a level will appear in the accumulator. The LC
(1). controller will reduce its output signal (F)' until it is less
Column-tray temperature control is achieved in Blocks than the operator-entered value and is selected as the
14 (2),15 and 17 (1). 'Thay temperature (inferred composi- feed-flow setpoint.
tion) is the measurement to PID Block 15, the output of At this point, the LC is the primary controller in a
which is the desired [D/(I + D)]' factor in CALC Block 17 cascade loop. Feed flowrate will reach a steady-state
(1). Feedback for the temperature controller (TC) is the value equal to the available heat-removal capacity. This
actual D/(I + D) ratio, which is calculated in CALC Block feedl.ate is the highest the column can tolerate and still
14 (2). maintain top- and bottoms-composition control.
This strategy presents a subtle control problem. 'I\vo When the heat-removal capacity increases (due to
primary controllers (TC and PC) are setting two second- colder cooling water or venting of inerts), the LC control-
ary-flow controllers (distillate and reflux). If for any rea- ler will increase until the local setpoint is selected for
son either flow cannot respond to its external setpoint, feed flow. The LC controller will not saturate because
the two primary controllers could be in contention for the actual feed-flow signal is used as feedback.
control of the remaining flow controller. For example, if Bumpless transfer to remote setpoint is assured by the
the reflux-flow controller is placed in manual, both the tracking logic of GATE Block 20. If the transfer is made
PC and TC would attempt to control their respective con- in the unconstrained condition, th`e external setpoint will
trol variable by manipulating distillate flow through gradually increase at the LC integral rate until the local
CALC Block 17 (1). setpoint is selected.
This control contention can be prevented by first de- The column base-level controller is PID Block 24,
ciding which primary controller is the most critical. It which outputs directly to the bottoms control valve.
was decided that column-pressure control was more criti- This distillation example required 24 blocks out of the
cal than tray-temperature control. Logic GATE Blocks 60 available to the control designer. 'I\vo such control
12 and 13 solve the contention problem in favor of the PC designs could be easily accommodated with this type of
controller when either flow controller is unable to re- shared-control equipment.
spond to its external setpoint. Microprocessor-based control equipment offers the
Two block-status bits (LOC and OLP) from each flow system designer many choices. The line between pure
controller are inputs to OR GATE Blocks 12 (1) and 12 regulatory and advanced control has become blurred. No
(2). The output of each OR GATE block represents the longer is a large digital computer required to implement
contl.ol condition of its respective flow controller. Each advanced control strategies. In a single ``box," the user
condition signal is sent to both OR GATE Block 13 (1) can implement a range of control strategies from single
and AND GATE Block 13 (2). The output of the OR loop to advanced feed forward, without the need to spec-
GATE is connected to the TRACK parameter of TC ify exactly all the necessary functions months before the
Block 15, and the output of the AND gate is connected to control equipment is delivered.
the TRACK parameter of PC Block 16. In addition, process changes are no longer a cause for
If either flow controller is unavailable for direction, the major concern. The restructuring or addition of loops is
TC will go to the TRACK mode and the column pressure easily accomplished without the need for a large pro-
will stay under control with the remaining FC. If both gramming effort, or for shutting down a process unit to
flow controllers are unavailable, neither the TC nor the install control hardware.
PC can affect the process. Hence, both are placed in The next article in this CE Refresher will appear in the
TRACK mode by the logic gates. The PID tracking logic Apr. 29 issue.
will maintain both FC external setpoints at their meas- Steven Danatos, Editor
urement values.
The startup procedure would be to first establish re-
flux and distillate flow control with local setpoints. Then, References
1.g*1aeEptoqLi.ris?uHa;£,Oo?oprobs;%OFr::ens:sr.?I;Ef:e:tg!g6.EvenonThermally
the PC and TC primary controllers would be placed in
automatic. Both primary controllers would immediately 2.#£yB:e°#ri:P£:;:SLsgeso?P:%t;°e¥en:igst:¥ne£::#°cne:j°t]h:°p%e:t„eddppoufB7a7n„q
go into the TRACK mode. One of the FCs would then be
placed in remote setpoint. The PC controller would go
out of TRACK, and into control. When the next FC is
placed in remote setpoint, the TC controller will go out of The author
TRACK, and into control.
%°n=[uiTin¥:fieh;SFa:xg%F;!C6;i.?n3S8
Constraint control
Fig. 5c is a block diagram of the overhead-condenser con-
straint-control system and the base-level controller. Col-
:E::pi:inis!:::h::::;ic;:::o;?:iiii:;¥gr:tre;::a3s5a
umn feed-flow setpoint (PID Block 23) is determined by
signal selector (SSEL) Block 22. It selects the lowest sig- %5lil:F:fsttha:elnus?#emr:l#,s::.doi;a
America.
nal between an operator-entered value and the output
from constraint LC controller (PID Block 21).
A condenser-constraint condition is measured by the

CHEMICAL ENGINEERINGITEBRUARY 4, 1985 77


SEQUENTIAL CoNTnoL
Electromechanical relays, programmable controllers,
hybrid controllers, general-purpose computers, and distributed
systems are devices for implementing process sequences
to control and manage process operations.
Ernest M. Cohen and Willi Fehervari, The Foxboro Co. 5. Opens fuel valve on selected-side burner.
6. Pulses fuel atomizer on selected-side burner, and starts
volution' in input/output devices, bus structures, com- flame.
munication gateways, and local cathode-ray tubes 7. Selects opposite-side burner after a fixed time period.

F (CRTs) has resulted in a more orderly approach to


sequential controls. Today, communications
between various modules in the control network are
8. Returns to Step 1.

being facilitated. Work is also being done to ease


the implementation of complicated logic, because
- „
the time in workhours needed to apply this logic to
complex batch applications is still great. „
Eleclromechanical relays
FCr-
Limit
Electromechanical relays use series- and/or paral- switch, Ci
lel-input contact closures to energize a coil that, in
turn, opens or closes one or more associated output
contacts. In its simplest form, the output contact of Left fuel
the relay converts a current higher than that of the valve

input contact. Suitable serial and parallel intercon-

EILIIIJI--I
Set-
nections of several input contacts to a coil imple- Pol n
ment their Boolean-logic relationships.
Fig. 1 shows a system for maintaining the tem- „
I

perature of molten glass, implemented by using IITL TTR


analog controllers and relay logic. This system
controls the fuel-to-air ratio for a flame that heats a
pool of molten glass, regeneratively.
The flue gases are routed through a ceramic
matrix that stores heat from them. Then, dampers --tsan..€sis...lil

are set to preheat fresh air by routing it through


the heated matrix. Fig. 2 shows the relay logic that Leftburner iiziiiiREiiErmunREwmus#nREREERERE!rmi

Molten glass
(at the proper time) accomplishes the following:
1. Closes both fuel valves for a short time, allow- Makeupairi
ing makeup air to purge the heating chamber.
2. Sets the damper to route fresh air through the
selected ceramic matrix, and to route flue gas
FC set-
through the other one.
3. Switches temperature measurement to the point
selected--side temperature sensor.
4. Switches all controllers to balance and hold, ~
during switching of measurement and outputs. ]`

Articles published thus far in this CE Refresher: Part 1, May 30,1983.


p. 66; Part 2, Aug. 8,1983, p. 79; Part 3, Sept.19,1983, p. 67; Part 4, Nov.
14, 1983, p. 233; Part 5, Feb. 6, 1984, p. 99; Part 6, Apr. 30, 1984, p. 77;
Part 7, June 25,1984, p.141; Part 8, Aug. 20,1984, p.151; Part 9, Oct.15,Figure 1 -Temperature control for a regenerative furnace
1984, p.103; Part 10, Dec.10,1984, p. 75; Part 11, Feb. 4,1985, p. 67.

?8 CHEMICAL ENGINEERING/APRIL 29, 1985


Because complex Boolean relationships are required, the
design and implementation of relay logic rapidly becomes
I start
pushbutton unwieldy and expensive. When solid-state switching devices
T were made available, the stage was set for the programma-
o Automatic/
ble controller (PC).

Cyclestart
Progammable controllers
:: i:Ttucah'R] Pet:Sye Solid-state programmable controllers permit arbitrary inter-
connections between a number of elementary devices that
R, „ Latchlng relay R Side select,anddamperpositionSide-select include switches, relays, timers, counters, registers, and
'12
register arrays. The output switches may be connected to an
R2 „ `^` Right arbitrary set of input/ouput (I/0) devices.
„,U\ When connected with I/0 devices having the proper volt-
F\2 I I, `,\, Left indicatorsHoldandbalanceforal control ersAtomizing-airtimerCycletimer,left
I,u\ age and power outputs, PCs can switch machinery on and
off, as well as input and output analog values to and from
R1 " T, T their internal registers. The only limitation becomes the
I Hmomentarycontact, 1 amount of I/0 and solid-state memory. More importantly,
program changes can be implemented by reprogramming
instead of rewiring the machinery. This improvement has
8,, resulted in a huge reduction in the cost of implementing,
documenting and maintaining controls that depend on Bool-
T2L;I T2-T2 ean logic.

Ci , , Left fuel valve open


Logic functions
11T3 The ladder line is the basic PC programming element. It
provides for a limited number of serial or parallel logic
C2,,Pightfuelvalveopen T Cycle timer,right elements that drive a coil or output element, In turn, the coil
'14
drives a logic switch that may appear in other ladder lines.
Ci,, C2lF T5 Several ladder lines constitute a ladder diagram, as shown in
Air-purgetimerFuelvalve'leftFuelvalve,rightisopenmallyopen
Fig. 3a.
T5oL±L IL A ladder diagram often has an equivalent logic circuit (Fig.

c,,+ _, Enable
3b), which is sometimes used to clarify more-complicated
logic relationships, and a programming-language equivalent
(Fig. 3c). This becomes important when a higher-level lan-
T5 -" SW,,t;h guage such as BASIC is used to program the controller.
Special-purpose high-level languages (called batch lan-
guages) become important when general-purpose computers
c2l+ " ITN0menclature are used to implement complex logic.
The power in a given PC lies first in the generality of
definition of elements. Such elements might consist of logic
switches, counters, timers, comparators, registers, and reg-
ister arrays. Newer PCs provide for programming by using
-- Pushbutton logic circuits and a PC version of BASIC. The PCs also
RFtLaTC, provide for direct operator interface through a CRT [J].
L5:,eanytahcetj:::,:dR:,y,srej:¥:::,r,gpz,;dTt.,:.,reh:¥ The logic diagram has immediacy because the coils of its
ladder lines may switch various forms of power. These may
_:I+ Belay activated by relay coil, 8,. This relay is closed
when the coil, F`h is not energized, i.e., nor
mally closed include 110 or 220 V a.c.; 5, 24 or 48 V d.c.; BCD (binary coded
ulse decimal); or ASCII over RS-232 or RS-422 lines.
rgized, the
R.
I Coilofpulserelay,
contacts change theirBi.Whenthecoilisene
state for a brief, limit In addition, PC registers may input and output voltages
ed interval
ching
and currents representing continuous measurement signals.
The continuous signals may then be displayed on a CRT, and
Rj Coil of latching relay, R,, when coil is momentarily
energized may be used in comparator logic lines to switch a logic switch
- Switch closed on time-out of timer, Ti when continuous signals have crossed a threshold value.
They may also be used as measurement, setpoint and output
i, Is closed
¢:?ij#::::r,'nT:.rvAaica°frtreers¥?lsd:nnge:gvy!tecdh'T for a PID (proportional + integral + derivative) calculation.

+ vL:#:t,§#f?'o%16hs closed when Its corresponding Example: Equipment-shutdown circuits


In the process of refining raw materials into a product,
Lamp
\ plants route a stream of these materials through a large
number of machines. Failure of any one of these machines
Figure 2 -Belay logic for regenerative-firing temperature control requires shutdown of the upstream machine in order to
prevent subsidiary failures.

CHEMICAL ENGINEERING/APRIL 29, 1985 70


The PC is frequently used to start and stop each machine
Voltage,+V in a predefined order. An equipment shutdown (ESD) logic
Ground
circuit monitors the functioning of a given machine. If a
A CDE
IE malfunction occurs, the logic circuit sets its output (called an
81 late//,` K ESD bit) to 1. Other logic circuits, on I.eceipt of the ESD bit
value of 1, shut down the machine [9]. An ESD circuit may
I
I "il
''
have inputs from a number of sensors on the machine, and
inputs from downstream machines. This ESD bit is I.eset to 0
only when the machine is restarted.
H Fig. 4a shows an intermediate product being fed from a
II
bin with a low-level alarm into a rotary valve. The valve feeds
(a) Ladder diagram into a variable-speed screw conveyor that conducts product
into a pneumatic conveyor. Appropriate ESD logic must shut
) I-E down the rotary valve if certain failures occur.
Fig. 4b shows an ESD circuit to protect the rotary valve, as
programmed in a PC. This particular circuit sets the rotary-
valve ESD to 1, if any one of the following occurs:

Ln i-
-i-I(b)Logiccircuit
K 1. Rotary valve fails to start after a specified time.
2. Pneumatic conveyor is not running.
3. ESD of the variable-speed screw conveyor is set to 1.
4. Low-level alarm from bin has been set to 1.
Restarting the area sets the rotary-valve ESD to 0.

Hybrid controllers
E (A . OB . B) . AND . C. AN D . D
Initially, contl.ollers were responsible for continuous feed-
K ( E . AN D . F ) . O R . G . OF{ . H
back control by using PID action. So many logic functions
(c) Batch language
have now been added to these continuous control functions
Figure 3 -Logic functions as realized for programmable controllers that the controllers will be referred to here as hybrid control-
lers. These modules control a number of process variables to
operator-entered setpoints by using control algorithms hav-
ing discrete or continuous outputs. Discrete algorithm out-
puts may have more than two possible values in addition to
the two-valued outputs used in PCs [fl.
Model.n hybrid controllers may use both continuous and
logic algorithms. A typical set includes PID, differential gap
(continuous control with discrete ouputs), a bias that pro-
vides an offset or bias between input and output, integral-
only action, digital input, digital output, ratio, low- or high-
signal select, ramp, timer, generalized sum, accumulator or
totalizer, deadtime, lead-lag, general nonlinear algebraic
calculation, a characterizer composed of piecewise linear
segments, a gate function capable of calculating general
Boolean functions, a switch for analog signals, and a se-
quencing block for outputting series of bit patterns.
In addition, the internal modes of continuous algorithms
may be switched by using logic inputs calculated from other
algorithms in the system. This greatly facilitates interaction
between the logical and continuous algorithms required for
sequential control. Most hybrid controllers have no onboard
operator or programmer interface, and require additional
modules to program and operate them.
Free-standing hybrid controllers having up to eight analog
and/or digital outputs and a limited onboard operator display
appeared soon after single-chip microprocessors became
available. Programmable with a simple keypad, they may be
readily used for small applications, with or without logic, as a
popular replacement for relay devices [4].

Example: Batch PID controller


A temperature controller for batch processes that would not
Figure 4 -Programmable controller for monitoring shutdown overshoot is an example of sequential control. Fig. 5a shows
such a control loop implemented in a hybrid controller. The

eo CHEMICAL ENGINEERING/APRIL 29, 1985


output of the algorithm is split-ranged so that 50 to 07o of and reporting capabilities than are usually found in PCs.
controller output corresponds to 0 to 100C/o of cooling-water When the number of discrete inputs and outputs exceeds
flow. Also, 50 to 1007o controller output corresponds to 0 to 1,000 or so, and elaborate logic is required, the increase in
|oo°/a steam flow. system complexity often forces the use of a general-purpose
The PID algorithm in this controller has several features. digital computer and a special-purpose batch language for
First, an external feedback signal (FBK) may be inputted to sequential processing [c)', 6', 7].
the algorithm. When a track logic input is set (TRK= 1), the An application of the general-purpose computer to sequen-
normal PID calculation is suspended, and the controller tial control is in a chemical plant where a single batch reactor
output is set equal to the external feedback signal. When is used for a large number of products. This application
TRK is set to 0, the PID calculation resumes. Logic outputs I.equires many raw-material and final-product tanks. Such
are provided that give the status of the automatic/manual tanks are often piped to a common header, requiring many
switch (A/M), the low deviation alarm (LDA), and the low switching valves. Shared equipment introduces scheduling
absolute alarm (LAA). considerations and switching requirements that increase the
The hybrid controller also includes: GATE algorithm to complexity of the system logic.
provide a number of logic gates, TIM-
ER algorithms, and SWITCH algo-
rithms. As seen in Fig. 5a:
1. Switching the A/M switch to auto-
matic starts the control.
2. If the temperature is below the
absolute low-alarm limit (LAA = 1), the
output of Switch A is set to 100%. Since
the low-temperature deviation is also
exceeded (LDA = 1), the output of
Switch 8 is set to 1007o. This applies a
value of 1009`ci to the reset feedback
input (FBK) to the controller. Since the
track input is also set to 1, the control-
ler outputs 1007o to the process, which
applies full steam (Fig. 5b).
3. When the temperature rises above
LAA = 0, Switch A outputs 07o, which
is passed to the controller output of oC/..
This applies 100%\ of the available cool-
ing-water flow.
4. When the temperature rises above
the setpoint minus the low deviation
limit (LDA = 0), Switch 8 outputs 45%,t,
which is passed to the controller output
to apply 10% of the available cooling-
water flow. At the same time, the TIM-
ER algol.ithm begins a delay, T1.
5. When delay Tl is exhausted, the
track command is removed (TRK = 0),
which replaces the reset feedback sig-
nal with the controller output, and be-
gins continuous control using the PID
calculation. Temperature attains its
setpoint without overshoot.

®enera I-purpose compulers


Although general-purpose computers
can perform arbitrary arithmetic and
Boolean calculations on both continu-
ous and discrete inputs and outputs,
they generally do not use ladder-dia-
gram notation for logic. Instead, they
use general-purpose programming lan-
guages with an added process-inter-
rupt capability, called batch languages.
In addition to continuous control, these Figure 5 -Hybrid controller implements batch-process temperature control
systems have more elaborate display

CHEMICAL ENGINEERING/APRIL 29, 1985 81


Other examples arise in a food-processing plant where describe complex relationships between logical val`iables.
food is cooked in a large number of kettles, and a pulp mill However, these scientific languages did not provide for
with a number of batch digesters involved in "cooking" wood interrupt-driven computational processes or computational
pulp. processes that did not proceed in a predefined order. A set
of languages was developed to satisfy the needs of batch
Ultimate logic controller applications.
The general-purpose computer with a stored program is built An important feature of large batch systems is their
on Boolean logic functions. Because its languages (for ability to keep a record of events and to report this record on
instance, FORTRAN, BASIC and PASCAL) were initially the appropriate display device at the correct time. The re-
created for scientific computation, this computer contained ports may include:
many arithmetic and logic functions. In such languages, a •Contact alarm states, time of alarm, appropriate
single logical switch became a binary digit that could be messages.
operated upon with the logic operators and subroutines. Also • Recipe vs. charge values of all system raw materials.
• Temperature history of a particular batch.
provided for were English-language comments, needed to
• Solvents added to reacted product
after a batch.
• Values of laboratory tests taken
during a batch.
• Calculated values based on mea-
surements taken during a batch.
The operator interface becomes very
important in a highly complex system.
The system must not only report and
localize alarms but must also direct the
operator to the alarm source, including
tabular displays, faceplates and full
process graphics. When more than one
control center shares access to the
same process data, coordination of
the internal intel.faces and the operator
display becomes crucial for overall
integrity.

Batch language requirements


High-level batch languages currently
used provide a data type called a
switch, which is more comprehensive
than a simple two-state device. These
languages can also offer a main pro-
gram for deterministic logic and arith-
metic, and advanced interrupt (also
called service-logic) routines.
The switch in the batch system has
associated information that includes its
current state, I)revious state at some
prior clock interval, preferred initial
state at the beginning of the logic pro-
gram, and desired or reference state.
This comprehensive definition of a
switch greatly simplifies many kinds of
logic tests on switches in the program,
with a minimum of program code.
A suspension or "WAIT" statement
in the language suspends further pro-
gram activit}+ until a particular event
occurs, permitting synchronization of
activity. A "WHEN" statement pro-
vides logic control on the occurrence of
a particular event. If such an event
occurs, the WHEN statement directs
the logic flow to a particular section of
Figure 6 -Batch-reactor configuration has contlnuous and on-off controls
the program.

8a CHFjMICAL F,NGINEERING/APRIL 29, 1985


Interrupt logic may be enabled during the main program. controllers originate from a remote operator workstation.
Once enabled, the occurrence of a particular event (e.g., Also, hybrid controllers, programmable controllers and the
switch change of state, operator entry, or certain program- genel.al-purpose computer may be manipulated from the
ming statements) causes the main logic flow to be interrupt- same video keyboard, presenting their data in the same video
ed, and an interrupt routine to be executed. On completion of display. More importantly, distributed systems allow the
this routine, the main logic flow may be resumed. addition of modules (and, therefore, application functions)
A special case is "HOLD" logic (usually triggered by an incrementally over extended periods of time, instead of in a
operator entry or program statement). It forces the digital single large project. A major benefit of this kind of control
and analog variables in the process to a predefined state, and architecture is the greater flexibility for both applications
suspends further activity pending operator shutdown or and financing [8].
restart.
Specifying an application in these languages requires Operator interface
some new techniques, as outlined by Ghosh [fl and Rosenof The operator workstation is evolving further, using many
[q. A checklist for features to be sought in vendor-supplied forms of input to the CRT. These include: touching the face
language is given by Ghosh [7]. of the CRT (touchscreens) [9], membrane pushbuttons with
interchangeable legends, and joysticks and lightpens. Voice
Example: Batch-reactor control output from system to operator has been available for
Simple batch chemical-reactor control requires charging sever.al years. A typical implementation is the solid-state
specified amounts of chemicals into a reactor vessel, sealing voice synthesizer with RS-232 interface, programmed in a
the reactor, and applying a specific time profile of tempera- fashion similar that of an alarm typewriter. Great progress
ture to control the reaction. The desired result is the produc- has been made in voice recognition, and printed-circuit cards
tion of a particular. yield of a given material, with a minimum are now offered that add this function to an existing system
of byproducts. [JO].
Fig. 6a shows a common batch-I.eactor configuration, with The next article in this CE Refresher will appear in the
continuous and on-off controls. Four ingredients are admit- June 24 issue.
ted to the reactor through on-off valves. The flow of each Steven Danatos, Edi,tor
ingredient is totalized to a target weight. A pressure control-
ler controls flow of evolved gas through a vent to maintain a References
pressure setpoint. 1. g£]#Vme.]a8od%£?3,Psr„°s5,rfumaF:8!e3,Cp°p?t3r&la3:S: A Technology Update, /%-
A temperature controller manipulates either steam or
2. 8;.::.n 'j gE8.4MA'7„EeqrT.u;£#e8:„S,hr###e%[gg,jspiT f72j5#ib8?t§:nsBsi:egg,'
cooling-water flow to control the reaction temperature dur-
Calif., June 6-8, 1984.
ing a specified time history (Fig. 6b). The time history
3 gay6re:r:'9u:g ,Jt. !rr.6cTet: e#%:I.°nrAfpsph,arAeg £££!r%:tts„e;:C#£sphT|Pgtaonn?
includes heating to raise the temperature at a specified rate
to a peak value, holding the peak temperature for a specified
4. F5]£:f6iT§yt.t'eifp6:%[£Sr°or,yEU„;:: Jig.k;993:npt.r8|3l:rs to Form Distributed
time, reducing the reaction temperature at a specified rate to
a curing temperature, holding the curing temperature up to 5.8:o°:.hjj?k',R#h„aL£:8icop£Cttu,#a#rnfospfaro°pb]:#5„adn„dstTnnft,rcso°#`;°tnes;
a maximum curing time, and cooling to a pumpout tempera-
Sgrste77ts, Lafayette, Ind., Oct. 1983.
ture. Once the pumpout temperature is reached, the product 6. B:::nm°efhtHaHtio:;' cB#,roa,mE*:#rect*°]dg8Simplifies Sequential Control
may be dumped, the reactor cleaned, and a new batch begun.
7. Ghosh, A., Checklist for Batch Process Computer Control, Cfeem. E7tg.,
The recipe is the list of critical parameters that determine Feb. 25, 1980, pp. 88-91.
the properties of the batch. In this particular case, the recipe 8. g?srtdrt#te€. Ja'nFr%it8;FteE.s PC^aen#. DE°#,' ]Ga.nT2'[¥]£§!8?bop:e8S65_%.based
includes:
• Weights of each of the four materials to be charged. 9. 5]PsTr?6:i6dEto Jb e#aet%rdatLeodopHteor#rcohii::]s , %:#% , CE°#,°L[:%±e]mg8;;
• Pressure setpoint to be maintained during the batch. pp. 93-94.
• Rate of increase of the temperature setpoint. 10. :]£chhai:'e ¥ite!Paeceec,hcB#:gr#gn.-,- £ann.EL%E]3:dj::7oF#re for the Man-
• Peak temperature setpoint.
• Rate of decrease of temperature setpoint. The authors
• Curing temperature setpoint.
• Cooling or pumpout temperature setpoint.
• Maximum reaction time.
• Cooling time.
li:i:::;t:ir#;;:;:ie;:am;;so:i:e;T:i:i::ti;i::oBg:::i;si:::r;o:,:iri::;:
For many batch reactors, this can be done in a single
general-purpose computer. A recent approach involves put- ::e:£n:e:ei£;gf¥tfre8[#Ecafiaenin,S¥#§t]es°:fnrTeeg:]§::°:I:dg%eji:ent:e:e:
ting the continuous controllers, totalizers, and ramp genera- in Massachusetts.
tors in a hybrid controller, and storing the recipe in a general- ---_- _

purpose computer that downloads setpoints and ramp rates


at the appropriate time. This is a distributed system.

Distributed systems
Distributed systems have made it possible to position con- Germany.
trollers near their corresponding process units and input/
output interfaces, while having command inputs for these

CHEMICAL ENGINEERING/APRIL 29,1985 83


FaulTIToleranT processes
Adequate and reliable data on process interruptions, coupled with
speedy diagnosis and repair of the faults, are essential for
maintaining uninterrupted production in chemical process plants.

Availability (AVAIL) is the average fraction of the time that


Ernest M. Cohen, The Foxboro Co.I
the unit is making acceptable product.
AVAIL -_ 1 -r!{MTTR) + (MTTsrN (MTBFi
have nominal, actual and maximum total production
during a year. Nominal production is the total produc- Actual production is directly dependent on availability and
Iiving with
tion faults given
obtained, is necessary in process
an average numberplants that
of interrup- maximum production. Increasing availability (hence, actual
tions (also called process faults). Actual production may be production) is a cooperative effort between the operations
less or more than nominal production, depending on the and maintenance departments, executed by reducing the
actual number of faults that occur during the year. Maxi- effects of process faults.
mum production is that attainable with no process faults - When a process fault arises, the following events must
an ideal never attained. OCcur:
In practice, an attainable production goal is to significant- 1. The staff must localize the sources of the fault.
ly improve nominal production - partly through better man- 2. The exact fault must be diagnosed, either automatically
agement of process faults. In this article, we will examine or manually.
the use of modern instrumentation systems to manage 3. The fault must be repaired immediately, if possible,
faults. within a known MTTR, and the process section should be
An evaluation of the consequences of faults comes from a restarted before upstream or downstream units must be cut
survey of large-ammonia-plant shutdowns [J]. Data for such back.
plants are given in the table. Part (a) shows that the uptime 4. The failed unit must be isolated for as long as possible if
or onstream factor is 87.997o. Part (b) shows that the uptime immediate repair is not possible.
(subject to efforts by plant personnel) or service factor for 5. Production goals must be redefined to accommodate the
the best plant is 96.9%. Part (c) shows a classification of shut-down machinery.
downtime by cause of fault. Similar analyses are pertinent to 6. An opportune time must be determined to repair failed
all industries and have been published for other types of machinery.
chemical plants and pulp mills. 7. An opportune time must be determined for restarting.

Dealing with faults Localize faults quickly


For the device, process unit, plant section or plant level, the Alarms are often the first sign of process faults. However,
following definitions represent long-term averages that interpreting these can be difficult. With complicated, high-
characterize function. The values may be calculated at vari- speed machinery, many alarms may be generated within
ous levels for a single device, a process unit, or an entire milliseconds of each other. The modern distributed-control
plant. It is especially meaningful to calculate these numbers system uses a hierarchical organization of alarms that facili-
for the sections of a plant that are separated by buffer tates analysis.
storage. Hierarchical alarms provide several levels of alarm, for
The mean time between failures (MTBF) is the average progressively smaller sets of equipment. One such design
time between faults that stop the production of usable (Fig. 1), used in a video-display-based workstation, operates
product. The mean time to repair (MTTR) is the sum of the as follows:
average time to diagnose a fault, and the average time to A plant is divided into eight areas, each of which is
repair the unit after a fault. Since time to diagnose is assigned a red and a yellow alarm status lamp, and eight
frequently much greater than time to repair, fast diagnosis named groups of machinery. A plant-alarm summary display
becomes important. The mean time to startup (MTTS) is the gives each area by name, and the names of the eight groups
average time to start up the process unit after a failure. associated with that area. An alarm in any equipment in a
*To meet the author, see Chem. Eng., Apr. 29, 1985, p. 66. given area will set the red-area alarm status lamp on, and
optionally sound a horn.
Pressing a button will call up the plant-alarm summary
display, with the names of the areas in alarm having the
i:a;:rgFE;#;ii:9:::3i!9:::1:9§8'44f#]i!:ih;e£§7#:a;!f:i9;18:2j;5:j]:;:;:#Li;8I3;g#§,:i#;i; relevant group name intensified and flashing. Pressing the

e4 CHEMICAL ENGINEERING/SEPTEMBER 16, 1985


signal button will result in selection of one of the areas in on the display in a special color that is dedicated to alarm
alarm, to show an area-alarm summary display. This gives states (see Fig. 2).
the names of each group in the area, and the names of the Both types of hierarchical displays localize the fault to a
individual control loops in the group. section of the process. The operator must have knowledge of
Then, the names of loops in alarm will show their tag the process and events immediately prior to the alarm to
names intensified, followed by a two-letter code indicating deduce what fault might be present. In processes having few
recycles and few interlocks, this is easy. When many recycles
Table -Shutdowns in large (550-1,650 std tons/d) ammonia plants are present, or many parallel machines are interlocked,
localization to a particular machine requil.es a joint investiga-
a. Overall plant performance (39 plants) tion by the operator in the control room and maintenance
Large-tomage plants (annual avg./plant) persons near the process units.
North America (1977-1981) When many machines are present, the situation becomes
Days of downtime 44.3 o/o
time-consuming because maintenance personnel may have to
No of shutdowns 8.4o/o travel a considerable distance to get to the offending device.
Onstream factor 87. 9°/a The ability of the operator to tell maintenance exactly where
to go is critical in limiting the time to diagnose (TTD).
b. Service factor (37 plants) Automated aids to diagnosis are helpful when many
Large-tonnage plants (days/yr down) machines al.e interlocked, or individual machines are compli-
North America (1977-1981 ) cated and fast, or when process complexity is greatest.
Average 91.2°/a (32.1 d/yr) Equipment-shutdown-analysis software can be employed
Best 50°/a of plants 93.5% (23.8 d/yr) where many interlocks are implemented in programmable
Best 20°/a of plants 94.7% (19.5 d/yr) controllers [fl. This software analyzes the equipment-shut-
Best plant 96.9°/a (11.3 d/yr) down (ESD) circuit interconnections to determine which
machine started the shutdown.
c. Downtime classification (39 plants) This process is divided into areas, each of which is started
Large-ton nage plants (days/yr/ plant) up and shut down with a single pushbutton input. Each
North America (1977-1981 ) machine's function is monitor.ed by an ESD circuit whose
Major equipment failures 15.3 d/yr/plant output is 0 if operation is correct, and 1 if an abnormal
Preventive maintenance 13.7 d/yr /plan condition exists. One of the inputs to an ESD circuit is
Instrument failures 1.1 d/yr/plant frequently the output of a downstream ESD circuit. These
Electrical failures 0.8 d/yr /plant interconnections form an ESD tl`ee (Fig. 3). When an element
in the tree fails, the upstream elements are shut down to
protect the equipment. The operator observes a large nun-
the type of alarm. A signal pushbutton then selects a particu- ber of machines going down within seconds, and finds it
1ar group; and a group full-value display shows a faceplate, difficult to determine which machine malfunctioned.
or representation of all control-loop data, with the tag names A general-purpose computer connected to the same net-
of loops in alarm intensified. work as that containing the programmable controllers can
Information as to which loop is in alarm is available monitor the ESD circuits when a shutdown occurs, and
graphically as well. A plant-overview deviation display examine the tree structure for the prime source or element
shows a vertical bar representing deviation from setpoint for causing the shutdown. The software system shown in Fig. 4
all loops in a given area. Acceptable deviations show inside a is used to identify the prime source. Here, the operator is
rectangular box, and alarms extend outside the box. Similar- able to identify the prime source within 30 seconds and can
ly, an area full-value display shows a vertical bar for each of direct maintenance to the cause of the failure.
the eight loops in each group of the area, with an intensified Alarms that may be milliseconds apart are resolved by a
loop-status letter above each bar to indicate the presence of sequence-of-events monitor (SEM). The device monitors a
an alarm. selected set of contacts at high speed. In the event of any of a
In addition, a printed record appears on the alarm type- certain set of predefined alarms, the local memory is trams-
writer, indicating the criticality of the variable, loop-tag ferred to the general-purpose computer to which it is at-
identification, type of alarm, limits, value of signal, and other tached. A video display and/or a printed report shows the
identifying information. If a number of alarms followed each time and tag of a critical set of alarms in the order in which
other, this listing would show the time order. they occurred. This is practical for up to 500 alarms and is a
Graphic displays provide another variation of hierarchical valuable aid in diagnosis. Fig. 5 shows an SEM output.
display, and use a matrix of membrane pushbuttons next to An average reduction of only five minutes in the localiza-
the operator display. When an alarm appears, a light flashes tion of a fault can save tens of thousands of dollars in
(as does the pushbutton light) that corresponds to the pro- production over the period of a year in high-capacity continu-
cess unit. Optionally, a horn sounds. When the operator ous plants.
presses the pushbutton, the display (often a process and
instrumentation drawing of the unit) appears on the video Diagnose the exact fault
screen. The symbols that are dedicated to particular mea- The operator/maintenance team is the critical entity in man-
surements and devices either flash, are intensified, or appear aging process faults. The operator must direct maintenance

CHEMICAL ENGINEERING/SEFTEMBER 16, 1985 8e


Figure 1 -Hierarchical alarm systems such as this, provide several levels of alarm, for progressively smaller sets of equipment

8e CHEMICAL ENGINEERING/SEPTEMBER 16, 1985


to the site of a fault as quickly as possible. When a redundant maintenance personnel adjacent to process equipment. Of-
process device is present to back up a failed device, the ten, it is not obvious why a machine has stopped. The reason
operator must decide with maintenance whether (and if so, must be deduced from the events just prior to failure (infor-
when) to connect it to the process. mation that the operator has). When a. hypothesis is proposed
Fast, clear communication (usually by two-way radio) is for the cause of the problem, it must be immediately deter-
necessary between the operator in the control room and the mined by maintenance in the vicinity of the machinery. That
certain faults can cause other faults,
cascading into a more serious shut-
down, puts a premium on finding the
right answer quickly. The alarm-dis-
play report` is a likely place to begin
looking for the source of the immediate
problem.
Historical-trend reports are needed
when the immediate alarms do not
tell the whole story. Small leaks of gas
or liquid often go unnoticed until a
failure. After the failure, a trend of
pressure or level going back several
days will quite often point to the prob-
lem. Historical-event reports combine
the alarm histories with all significant
changes to the process, including oper-
Figure 2 -Alarm states are indicated by flashing symbols or by colors dedicated to alarms ator setpoint changes, lab-test entries,
charge values, and so on.
The post-trip review report used
when an SEM is a part of the system is
a special type of alarm summary. When
-is
a predefined trip event (for example,
contact closures, or occurrence of an
- qHRE iaRE

- §REaiRE
.:
S#

•:.....
ffiE
alarm) takes place, the actual (sampled)
values of up to 61 variables, both dis-
i:.

crete (such as alarms) and continuous


I

ii
ff

(such as analog measurements), are


*

#
printed out for a period of 15 minutes
## before the trip to 15 minutes after it.
8g
+

60 The samples are taken at 1-min inter-


........ •......'

8
a- .:. •.:....;;: vals between 15 minutes and 5 minutes
before the trip, at 15-s intervals be-
.''.'':
..:.. .

§se ¥#9£
tween 5 minutes before and 5 minutes
•.

ffi :..
•'.
3Hrm '`''''`:;

after, and again at 1-min intervals from


5 minutes to 15 minutes after the trip.
ne I....
tire
..a
-a g

:.:.:,,:
This gives a comprehensive, time-based
picture of pre- and post-trip events, aid- i

a
ing the operator in assessing the cause
4a ffi
of the trip and damage that might have
iHHEEfr I
dEBffl •..
ffiae
.:.....
- been caused.

.:.

8 es# ¥£ Repair The fault, if possible


After diagnosis, two questions become
# 3 §

i 3 critical: Is repair possible at all?, and


a §3
Can repair be done in time to prevent a
cascade of faults? The mechanic may
be able to answer the first question
from experience. However, the second
question really calls for some historical
information.
An automated maintenance-record
Figure 3 -lnterconnected ESD (equipment shutdown) clrcuits form an ESD tree system provides information on the
state of repair of the machinery in a

CHEMICAL ENGINEERING/SEPTEMBER 16, 1985 87


form that makes quick reports possible. These reports are Lost-time reports
the basis for determining MTTR and MTTS.
A preventive-maintenance program used in a waste-treat- Knowing the consequences of faults in terms of lost produc-
ment plant (illustrated in Fig. 6) provides an example of what
is possible. A preventive-maintenance data file was created :ieopnoft:`#ee::#:i3haiiio:i:lie::=sTbalJeorb;e:aci:inApiot::-t:;se-
whose contents included a device-running contact status; a tem, lists the common records of faults, with inputs from
preventive-maintenance (PM) constant
that indicates the interval between
PMs; the integrated run time until the
next PM; the actual accumulated run
time, categorized as to motor, pump,
gearboxes, valves; the equipment tags;
and the date of the last failure, the
equipment identification, and the date
for each of seven separate failure
events.
An entry into the operator keyboard
produces a PM report showing for each
device the elapsed running time since
PM, and date of failures for up to seven
incidents. The devices in the four cate-
gories are:
1. Motors in pumps, mixers, ozone
equipment, and valves; and level, flow
and pressure transmitters.
2. Pumps either freestanding or in
ozone equipment.
3. Gearboxes in mixers.
4. Valves, i.e., valve bodies in on-off
and modulating valves.
The time difference between the last
contact-off status and the last failure is Figure 4 -Software can be used to analyze shutdown tree to determine which element failed
the time out-of-service.

Figure 5 -Sequence-of-events-monitor (SEM) printout enables operator to determine which event(s) Initiated a shutdown, and which followed

8e CHEMICAL ENGINEERING/SEPTEMBER 16, 1985


maintenance and operations. For each fault, this report tanks fill up and the downstream ones empty. The time until
shows the plant section, date, crew, shift, elapsed downtime,either of these happens is the time that maintenance person-
curtailed time (time waiting for restart), and tons of lost nel has to I.epair that process section before other sections
production. In addition, the department responsible for must shut down. Provided that the section has enough
downtime is encoded, as well as the plant section to which the overcapacity to "catch up," a cascade of shutdowns can be
production loss is to be assigned. The report is sorted by prevented merely by setting the tank levels based on cumu]a-
dive MTTR. The procedure and calcula-
tions to do this are given in Yorke [2}].
Preventive-maintenance (PM) report Of course, this is possible only in an
5/311/82 integrated plant-wide control system,
Elapsed with both current operating data and
F3untime to runtime Failu re Eq ulp ment historical maintenance information be-
Tag lD next pM last pM date ide ntify ing available to operations staff.
P-78A 25 2,540 60
5/23/82 Pump When lo repair and restai+
2/22/82 Pump When immediate repair is not possible,
3/2/81 Pump and the operator has moved the process
M-53 36 4,271 729 Mixers to an alternative operating point, the
7/1 /81 Mixers proper time to repair the disabled de-
Cell-5 74 520 480 5/1 /82 Ozone eq. vice may be considered. Other failures
72 340 660 4/21 /82 Ozone eq. that completely disable the process sec-
Dry A 76 220 40 5/5/82 Ozone eq. tion provide an opportunity to fix all
faults at once. A curtailment in produc-
tion may pl.ovide an opportunity to re-
Figure 6 -Preventive-maintenance report generated in a waste-treatment plant pair. Scheduled maintenance some-
times may be moved up to take
advantage of failures. With the proper
operating-department and maintenance-department sec- automated maintenance records, it is possible to make these
tions. This is another powerful diagnostic tool, used after the decisions based on MTTR and lost production rather than on
fault, to help set priorities for the next scheduled shutdown. intuition.
Generally, restarts should occur when MTTS does not
Alternative operating points affect production. Typically this happens when a parallel
When a plant section does not shut down completely but is unit is being restarted at the same time that an entire plant
disabled, it is sometimes possible to continue running but in a section is being restarted.
degraded mode-e.g., at a reduced production rate. To One should make operations fault-I.esistant by identifying
maintain product quality, an infrequently used operating fault-prone locations, assessing their operating status (espe-
point must be called on. If the computer system has stored cially, MTBF, MTTR, MTTS, AVAIL) and the production
sets of equivalent operating conditions that will produce a losses associated with faults. This is possible if one has an
given product quality, the operator may select one. If these easy-to-use method of entry of maintenance data, a means of
equivalent conditions are not stored, the operator is reduced coordinating them with operations data, and succinct reports
to using past notes or experimenting. that can be quickly obtained on a video screen or a printer.
Collecting equivalent sets of operating points is a labori- The final article of this CE Refresher will appear in the
ous operation, which can be facilitated, however, by using Dec. 9 issue.
historical-trending programs. These collect trends of a se- Steven Danatos, Edi,t,or
lected set of variables, and then do calculations to get the
means, standard deviations, etc., which are outputted in a References
printed report. The more elaborate of this class of programs 1. yof#LasTS'usje?.'|S?dc#e°#3;.W¢ryg.#uasrecsh °]fg8£rmonia plant shut-
can select the data corresponding to periods when the pro-
2. §#;#iz„aGji:n:fog77t:;;:7#y£°¢##£#e9#}°tn#e°c#:„f£,r A¥sr#to/¥%):
cess is in steady state, so that the means calculated corre-
spond more closely to setpoint values. An additional level of
3. £:::.nA#e#.;aE„q%]op%mt;.%{ 8£#g.Znce4%%]nysj3e;go:cDaf];i.r±tEg# system,
sophistication uses a static mathematical model of the pro-
cess to extrapolate other equivalent sets of setpoints that
will produce the same product quality.
The author
Preventing propagation of shutdowns is one of the most I -__ __
important ways to preserve nominal production. When a
shutdown occurs in a process section whose cumulative
frequency-of-repair is known, levels in the tanks upstream
and downstream can be set to minimize the number of
instances when that section shuts down the rest of the plant.
When a plant section is shut down, the upstream buffer Massachusetts.

CHEMICAL ENGINEERING/SEPTEMBER 16, 1985 80


Optimizing
p ooess and eoohomic aims
ofope:aht,i:gnc:osiufp:#s:,tRi:::uEpfi:3#5agnifi:entifi:nu!rf!i?:flftffn:8:t;£Foocn:rsos'units,

James A. Rogeis, The Foxboro Co.

objectives, it is essential that it meet them


through the conversion of resources into goods
0nce a company establishes a set of economic
and/or services.
There are two types of economic objectives, one called
strategic; the other, operating. Our approach here will be to
discuss only the operating type in detail. The strategic
objective is primarily concerned with external rather than
internal problems of a company, and specifically with selec-
tion of a company's product mix and target markets.
In most companies, operating objectives inevitably absorb
the bulk of the company's energy and attention. Everyone in
the organization is concerned with the myriad of recurring
operating problems. Management continually seeks to im-
prove efficiency, to cut cost, to sell more, and to advertise
better. Problems are generated at all levels of management;
those beyond the scope of lower management become the
concern of top management. The number of such decisions is
great and constant, particularly because of the need for daily
monitoring and control.
There are three types of basic resources: physical, i.e.,
plants and inventories; labor; and monetary, i.e., money and
credit. All are used in the conversion process; the plant
becomes obsolete, people retire, and money gets spent. With Figure 1 -When a disturbance occurs,
the control variable should be changed
this viewpoint, the overall problem for a company is to
configure and direct the resource-conversion process as effi-
ciently as possible, or, in more conventional language, to ideal control system would take into consideration all signifi-
maximize profitability of current operations. cant interactions as well as the economic impact among
The area that has been most often neglected for improving process variables, and establish a corrective control action
profitability is the control system. The design of a control that would achieve optimum process performance.
system and its level of use determine how well a unit in a By expanding the economic objective of profitability into
process plant is run. Therefore, this directly affects the three factors -efficiency, pl.oduct quality, and safety -
profitability of that unit. In this article, we will focus on the one obtains a more useful set of operating objectives.
use of a control system and show how doing this can lead to The major problems concern how to apply effective opti-
higher profitability -by solving two different problems, mizing control for operating conditions, resource allocation,
optimizing and information, via a control system. and scheduling of operational units among functional areas
in product lines [J,fl. The key decisions involve: inventory
What is The optimization problem? control for mass and energy in order to maintain steady-
The operating objective for a multivariable chemical process state operation; establishing optimal setpoints for efficient
is to achieve economically superior plant performance. The operating conditions; and setting of resource allocation and
scheduling to meet production plans.
Optimizing control is difficult to accomplish because dis-
turbances continually upset the process. Disturbances are
§i::rt:;;I:::8§8#p;I;7;9j[:;g;#9:§i]§nit3|j;:ii;:gp;pr;6;7;,:pP;a::i:7E]:g]ii;nle;:¥#¢::iij?P:273:i::p;a:jr!L variables that adversely affect the process operations and

90 CHEMICAL ENGINEERING/DECEMBER 9/23, 1985


over which there is no direct control. Values for such vari- 100 on an arbitrary scale. The graphical relationship among
ables are often determined by some known or unknown C, I) and I is an expression of the system's mathematical.
agency external to the process boundary. Control action is model.
required to overcome adverse effects. Uncontrolled distur- At time £], the disturbance variable has the value of I)1.
bances may be associated with raw materials, ambient condi- This` means that the process has been forced to operate
tions, fuel and energy costs, and other economic factors. somewhere along the line AD]. The optimum point on the line
Economic disturbances may vary with time, and will, is at Point Z because no other point has a higher value for
therefore, dictate a different mode of process operation. A throughput. In order to operate the process at the maximum
change in product specifications or a management demand throughput, the control variable must be set at Ci.
for grade change in an olefins plant or a production-rate At some later time, £2, the disturbance changes to I)2. If it
change in a specialty-chemical plant can be interpreted as an goes undetected or is ignored, the process will operate at
economic disturbance. These present new control problems Point I/. However, if optimal control action is taken, Cwill be
changed from C] to C2 and, therefore,
the process will be brought to Point V,
which is the optimum at the current
value, I)2. The difference between
Point U (throughput level of about 83)
and Point V (level of 90) illustrates the
economic justification for optimizing
control. (The dotted line Wzvy is the
locus of optimum points for the given
range of I).)
In general, processes are affected by
several principal disturbances that re-
quire manipulation of several different
control variables. Thus, the optimiza-
tion problem is complicated. Failure to
detect disturbance changes, and inabil-
ity or inaccuracy in determining and
controlling promptly to new optimums,
produce performance inferior to that of
optimizing control.
Fig. 2 shows four idealized per for-
mance curves:
1. Theoretical optimum (varies with
time because of both gradual and
abrupt disturbances).
2. Optimizing control (follows the op-
Figure 2 -Area under curves indicates improvement possible from better control timum closely).
3. Advanced regulatory control (rela-
tively freqrient and fairly accurate con-
when they occur, and remedial control action must be taken. trol adjustments).
It is never possible to identify all disturbance variables of 4. Conventional regulatory control (relatively infrequent
the process, let alone measure them. Numerous factors or or inaccurate control adjustments).
conditions are constantly changing. Some are important, The area under each curve in Fig. 2 represents the incen-
most are not. The best procedure, when practical, is to tive for the next higher level of control, e.g., advanced
include all major process disturbances and ignore those control vs. conventional control [J].
involved with second-order effects. This approximation will There have been many attempts to optimize the plant
render the minor disturbances indistinguishable from error. areas and process-unit areas by using dynamic process mod-
Revisions in return-on-equity and market forecasts are els and static economic models. It is assumed that with
economic disturbances at higher decision levels. These call accurate process measurements and models, a determination
for additional scheduling computations and integrated con- can be made of how a plant should be operating to achieve
trol actions. The optimization problem created when a distur- economically superior performance. It has been found that
bance occurs is shown schematically in Fig. 1. this approach is lacking because business information is not
For a hypothetical process with two independent vari- effectively incorporated.
ables -control variable (C) and disturbance (I)) -the ob-
jective of improved efficiency is to optimize an operating
What is the information problem?
condition, i.e., to maximize throughput (r). The contour lines The business-information function uses the real-time pro-
of the throughput functions are plotted on the CD plane, cessing system. With this, the organization should do busi-
with the maximum throughput corresponding to the value of ness planning more intelligently - since accurate real-time

CHEMICAL ENGINEERING/DECEMBER 9/23, 1985 91


process data should enable better plant operation. However, • Maximize throughput and minimize operating costs, giv-
in reality, most information available in plant real-time data- en the cost and availability of energy and raw materials.
base systems is useless. Generally, the useful information is • Meet product-quality specifications and order-require-
external to these highly structured databases, since business ments, because of competitive pressures to improve product
decision-making is long-term in nature, less structured, and quality, flexibility and quantities.
relies on external business factors. • Protect personnel and equipment, because of stringent
Therefore, to achieve the optimum economic performance, environmental and safety regulations. (This is accomplished
the key is to supply some of the business information that is by designing an independent emergency-shutdown system.)
available externally on a weekly, monthly or yearly basis in a Tools for implementing an integrated control strategy
time period relative to the real-time operation of the industri- (i.e., computer architecture, distributed processing, and data-
al process. However, the information-flow requirements of bases) have become increasingly powerful, reliable and avail-
such a system can become very complex. able. A multilevel hierarchical approach provides a frame-
Fig. 3 graphically displays one major
aspect of these requirements. The tri-
angle on the righthand side represents Business information (transaction data)
information flow from process mea-
surements up to business manage- # # fr # # # S
ment. As information is processed in
the upward direction, it can be reduced ConpRTvy / / +'EcomdrS#!*oNNS;ness
because the higher levels in the plant
need information on processing trends,
rather than detailed process data.
There are several techniques for elimi-
nating redundant data from the plarit/area / / Pr3f tyal¢ryH:i:°N:ProcessunitOperatingdecisions
streams of information being transmit-
ted [4].
Often, when the information flow in
plant-management systems is being
discussed, the upward flow is all that is
considered. To achieve optimum eco-
nomic performance, the downward
flow of information through the organi- Variable -~'~
- Stability decisions
zation must also be carefully consid-
ered. The triangle on the lefthand side
of Fig. 3 represents the flow of infor-
mation down through the plant-man- tttt[t!
agement structure. At the highest lev- Process information (real-time data)

els are the company objectives as well


as the major business-planning func- Figure 3 -Information passes both up and down the management levels
tions. These objectives must be put into
operational form as they flow down. It
is essential that they be combined and analyzed all the way work for organizing and implementing an integrated control
down to the unit processing level, so that unit processing can strategy for large complex processes (see the table). The
be optimized. If unit processing is not optimized, production functions at each level are to make decisions that concern
in the plant cannot be optimized either. only the level immediately below it. However, the informa-
The challenge for achieving both the planning and the tion on which these decisions are based is drawn from all
operating goals is to integrate these for optimum economic four levels. As the hierarchy is ascended, the following
performance. Integration is only part of the solution; timeli- OCcur:
ness in determining a change, and promptness in controlling • Control objectives become more economic and less
it, also are critical. physical.
• Control solutions become more complex.
Optimization and information solution •Control variables become more random and less
For solving the optimizing and information problem, technol- deterministic.
ogy is available in the form of an integrated control-strategy • Control-action frequency decreases.
approach. This offers the engineer and engineering manager The lowest level, variable management, uses conventional
many techniques for reaching a company's operating objec- regulatory control (see table), which includes simple control
tives, and is applicable to most industrial processes -partic- and cascade feedback loops [J, fl. The purpose of convention-
ularly chemical manufacturing. Use of this approach is al regulatory control is to hold the proper variables near
spurred by the desire for efficiency, product quality, and steady state to allow people or advanced control (automatic
safety. This means operating the plant to: constraint optimizers) to meet operating objectives. Conven-

Oa CHEMICAL ENGINEERING/DECEMBER 9/23, 1985


tional regulatory control, in a broad sense, interchanges mathematical sense. The practical approach relies on the
dependent and independent variables. With proper conven- experience of operations personnel to improve performance.
tional I.egulatory control (using cascade feedback principles), They use an ``experience" model, which reflects their engi-
a cumbersome physical model having inappropriate indepen- neering and process knowledge. Often, this approach proves
dent variables can be greatly simplified to one with more to be more effective than using a mathematical model, but
convenient higher-level generalized variables. (See Fig. 4.) produces a great deal of inconsistency because the experi-
The process-unit-management level employs advanced ence of the individuals varies.
regulatory control (see table). This includes feed forward The alternative is a mathematical approach that develops
control [7], adaptive-controller tuning [8], nonlinear control predictions from a model of the real process, but processes
are often so complex that a realistic, time-varying mathemat-
[9], multivariable (decoupling) control [jo,JJ], control of cal-
culated variables, and control against constraints [j9], for ical representation is unrealizable. The quality of predictions
controlling flow, temperature, pl.essure, level and composi- from a model has a direct impact on the reliability of benefits

Table -Multilevel-hierarchical-control
approach includes four or more levels

OBJECTIVE: TOOL:
Economic business Information control
Includes:
• Management information systems/
data processing
• General ledger
• Payroll
• Personnel
• Accounting
• Investment management

OBJECTIVE: TOOL:
Production Optimizing control
Includes:
• Non-constraint optimization
• Dynamic process models
• Static economic models
• Inventory management
• Quality decision management
• Maintenance management
• Production scheduling
• Plesource allocation

OBJECTIVE: TOOL:
Operating Advanced regulatory control
Includes:
• Constraint optimization
• Static process models
• Material energy balance
• Feedforward control
• Decoupling control
• Quality control

OBJECTIVE: TOOL:
Stability Conventional regulatory control
Includes:
• Simple control loops
Figure 4 - Cascade-feedback • Simple cascade loops
interchanges can reduce control complexity

tion. The cascade-feedback principle extends itself as these estimated from optimization. Complex models, requiring
advanced regulatory-control strategies provide convenient many inputs, are difficult to fit with real process data.
independent variables (setpoints) for human operators or the Inaccul.acies from process measurements are all combined,
next-higher level, plant/area management. and the model can lose its predictive accuracy [Jfl. Further,
That level uses optimizing control in both a practical and a inaccurate predictions result in larger safety margins and

CHEMICAL ENGINEERING/DECEMBER 9/23,1985


®3
reduced benefits. The value of the mathematical approach industrial process with respect to its economic goals. Multi-
may be in allowing less-experienced personnel to address the 1evel hierarchical control provides a framework for organiz-
optimizing problem. However, this involves dealing with the ing this integration, and offers a rational basis for design
complexity and vastness of keeping track of many variables. and implementation. An initially complex control strategy is
An application of the mathematical model is as a decision- replaced by local control strategies that are mol.e easily
support tool that permits responsible personnel to look for designed and implemented. Compensations for interaction
problems by examining the relationship between real and effects are made through the coordinating efforts of highel.-
mathematical variables. level controls.
However, process characteristics can often be accounted The multilevel hierarchical approach allows an ordering
for without complex modeling. with respect to time, degree of aggregation, frequency of
The proponents of multivariable techniques (for conven- control action, and other attributes to be considered during
tional regulatory control, advanced regulatory control, and the systems-design stage. The multilevel hierarchical struc-
optimization) often tend to overstate the need to simulta- ture also plays an important role with respect to organizing
neously handle many variables rather than decompose them the flow of information through the system and providing
into a series of single-independent-variable conventional- the mechanism for effective utilization of feedbacks for
control and advanced-regulatory-control strategies for opti- control and decision-making.
mizations that are appropriately linked together in time [J4]. Finally, the integrated multilevel hierarchical approach is
If steady-state analysis shows that the objective function compatible with modern concepts of distributed microproces-
is always optimized at certain constraints, there is no reason sor computer control, and with the use of minicomputers in a
to model the process. For most processes, the optimum coordinated system of control for achieving long-term profit-
conditions lie beyond constraints. ability in the process plant.
The most simple and effective form of optimizing control Steven Danatos, Editorr
usually operates the process against predetermined opti- References
mum constraints. Constraint-optimizing control methods are 1. gij]:S*:# I.o9kt, `iE;a:illation Control," 2nd ed„ Chap. 1 and 11, MCGraw-
very useful when problems can be decomposed into single
2. t#::gri6ac:eas;e!t¥gsu:i:dfi,8i:pa#ry;1::gttefo°Zt;*LZ#°Andf„°ar„::#Pcuot#rc#o°|;?
independent variables, and when constraints predominate.
Usually, decomposition into such variables provides better
3. fracvGa:;wTfi:ii, #wm§# L88g.trol for Industrial Processes," chap. 5,
understanding, commissioning, and reliability without loss
of performance. In other words, decomposition follows the 4. g;ericr,%n;,Cri¥;6hD]ag%7,Cp:T8;:Ssion by Redundancy Reduction, /RE
cascade principle previously described.
5. 8::€r°oni', t;fo¥Z.Bfi:`g.#nacye%S,' ]Tge8r3TiT°6]8.8y and Techniques for Process
If constraint-optimizing control is combined with noncon-
straint optimization for single independent variables based 6. Col.don, L. M., Feedback Control Modes, Cfaem. E7zg., Aug. 8,1983, p. 79.
on simplified process models, a very powerful control tool 7. Badavas, P. C., Feed forward Methods for Process Control Systems,
Cfaem. E7tg., Oct. 15, 1984, p. 103.
can be employed for profitable applications. In particular, 8. F:g.a8,a;b8E., 8..jgB.:rect-Synthesis and Adaptive Controls, Che77i. E7ig.,
note that:
1. The multilevel hierarchical structure with the decompo- 9. #%;:nit:;.Ji'4,F]ege8d3?ape.k23ng:thods for PI.ocess Control Systems, Cfoem.

sition approach can extend to a hierarchy of four, or more, 10. Rys, R. A., Advanced Control Methods, Cfoe77.. E7tg., Aug. 20, 1984, p.
levels. Each higher level coordinates the actions of a lower 151.

level, according to cascade-feedback principles. 1 1. §g#C:aT;£:::g':s:ilp,E]Co;:nf::.#e;:::p]!np:::#Pb];ntgh :nEpgtn;:1:f,1:8 £:#ird°::


2. The operations personnel are quite familiar with the
12. Kurth, T. C., How to Use Feedback Loops to Meet Process Conditions,
principle and, therefore, accept it more readily. Cfeem. E7tg., Apr. 30, 1984, p. 77.
3. There is much compatibility between the hierarchical- 13. 8oaisukinns 'cBLpr.6[T8;f#'e%?dfz8£:::;nrlb:% 9}°o#:tyhps§5:°r Advan ced
decomposition approach and the use of distributed micro-
14 Eao:V;l`t;§§];eggv=;ig€:e;t¥it;g;e:#Sn¥mTc°;r`:Zi:inspe£#oP#'{:%,mveotih°]dg:
processors in an integrated system for plant control [J5].
4. A single-shot approach for integrating these operations
is impractical. Instead, an approach must be taken that 15. MCGee, N. F., Microprocessor Controllers, Cfaem. E7tg., Feb. 4, 1985, p.
67.
allows integration to progress over time, starting with those
areas that offer the greatest potential for improved
The author
profitability.

Summary !egb;ilo?k,29692i£,'Fteo,x¥6!1i:ge4:;:6?8:.VNaenEc:if:C:o#:o:
An integrated control system using a multilevel hierarchical
structure with a decomposition approach is concerned with #h#d;Eos%:a:tafeey¥:e:n:i;#:o:£t;;*;#sar¥:;Xj;£Ee:nFE;;sre:eh:_
the integration of many decision-making and contl.ol func- Instrument Soc. of America, and is a registered Pro-
fessional Engineer in Massachusetts.
tions. These play a role in the effective operation of an

®4 CHEMICAL ENGINEERING/DECEMBER 9/23, 1985


If you have questions about information contained in any of these reprints, contact your
]acal Foxboro sales representative or J. H. Rogers, Manager of Advanced Control
Technology (617) 549-6580.
Compliments of
The Foxboro Company

'CunoR
PUB 623A IOM 5/86
Printed in u.S.A. ® Registered Tradornark

You might also like