Getting the Best Performance from Challenging Control Loops
Jacques F. Smuts - OptiControls Inc, League City, Texas; jsmuts@[Link]
KEYWORDS
PID Controls, Oscillations, Disturbances, Tuning, Stiction, Dead Band
ABSTRACT
While many control loops are easy to tune and present almost no control problems, a few control loops
can be very problematic and never seem to control right. These loops exhibit process control problems
like oscillations, large deviations from set point, sluggish response to disturbances, spurious upsets,
and inability to follow a set point. Control engineers and technicians can spend many hours or even
days trying to improve the performance of these challenging control loops, but the results often remain
unsatisfactory.
This paper explores various reasons why the optimization of some control loops can be very
challenging, and explains how to get the best performance from these loops. It covers several problems
originating from within the control loop, such as poor controller tuning, final control element
problems, and instrumentation problems. It also covers problems originating outside of the control
loop such as interactions and disturbances. The paper presents the appropriate corrective actions to
take for solving each of the problems and improving control performance. These include tuning
methods, control strategy design, calibration and maintenance practices, and equipment design.
To help with the diagnosis of control problems, the paper presents a fault tree that divides control
problems into different categories. It provides steps to follow for differentiating between different
problems and diagnosing the causes of poor control. It also presents corrective actions that should be
taken to solve each of the problems. Examples from real process plants are given throughout the paper
to illustrate many of the typical problems experienced and show improvement results possible through
proper problem analysis and corrective action.
HIGH-LEVEL SYMPTOMS OF POOR CONTROL
Poor control loop performance often makes itself evident as excessively large deviations between the
process variable and its set point. An operator may notice the deviations on his process trend displays
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
or process alarm system and place the controller into manual mode in an attempt to stabilize the
control loop. As a result, poor control loop performance may be indicated by excessive deviations
between process variable and set point or controllers being in manual mode.
CONTROLLERS IN MANUAL
Some plants have 30 percent or more of their control loops in manual control mode. Being in manual
mode does not necessarily mean the control loop performs poorly in automatic control mode. Many
control loops are associated with redundant equipment, or certain operating modes, and can justifiably
be in manual mode as a result. However, if a control loop is supposed to be in automatic control mode,
but it is in manual mode, further investigation is required. It is important to talk to the operator to find
out why the loop is in manual. Historical process trends of times when the loop was in automatic
control can be reviewed to gain more insight into the problem, or the loop can be placed into automatic
control mode to observe how it responds.
DEVIATIONS FROM SET POINT
If a controller is in automatic control mode and its process variable always remains acceptably close to
its set point, there should be no need for concern. However, excessively large deviations between
process variable and set point indicate poor control. The problem may be a constant or intermittent,
and could originate from within or outside the control loop, but if large deviations occur, it provides
grounds for further investigation. Large deviations between process variable and set point can be
caused by a rapidly changing set point, process disturbances, loop nonlinearities, interactions, control
element saturation, or poor controller tuning. The exact cause can be pinpointed through systematic
analysis, using a fault analysis tree like the one below.
Figure 1. A fault analysis tree for systematically determining the cause of poor control.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
CYCLICAL OR RANDOM DEVIATIONS
The first step in analyzing the control problem would be to look at the shape of the deviations on a
trend plot: whether they are cyclical (oscillating) or random. This will determine the path for further
analysis. Although sophisticated frequency analysis tools can be used to determine if deviations are
cyclical or random, it can also be done quite simply by looking at historical time trends of the process
variable. Once it has been determined if the deviations from set point are cyclical or not, the next level
of analysis can be done.
Figure 2. Process variables behaving cyclically and randomly.
CYCLICAL DEVIATIONS
Cyclical deviations (or oscillations) can appear as combinations of sine, saw-tooth, or square-wave
patterns, sometimes intermixed with some randomness. Oscillations can originate from within the
control loop or it could be caused by external factors. It could also be as a result of a cyclical
interaction between two or more control loops. To narrow down the cause of the oscillation, the
controller should be placed into manual mode to see if the oscillation would stop. If the oscillation
persists when the controller is in manual mode, it originates from outside the loop.
EXTERNALLY-CAUSED OSCILLATIONS
An oscillation with its origin outside the control loop can influence the control loop through its set
point or through the process.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
OSCILLATING SET POINT
Oscillations entering the loop through the set point are easy to find simply look at where the loops
set point is driven from. For example, if a steam flow control loop oscillates because its set point
(coming from a temperature controller) is oscillating, the problem likely lies with the temperature
controller. To verify, place the temperature controller in manual and see if the flow controller stops
oscillating. The fault analysis should then be applied to the temperature controller.
OSCILLATING PROCESS
One oscillating loop can cause several other loops on the same plant to oscillate with it. The loops will
all oscillate in harmony with the same period of oscillation.
Historical trends or process analysis software can be used to identify all the loops oscillating with the
same period. The problem loop can then be isolated through knowledge of the process and its
interactions, by looking at phase-shifts between oscillations, or by placing likely culprit loops in
manual one at a time. If the loop driving the oscillations is placed into manual control mode, the
oscillations will cease on all loops. That loop should then be analyzed further.
Note that this scenario is different from a cyclical interaction in which two or more control loops
interact directly with each other in a cyclical fashion. In the case of a cyclical interaction, any of the
participating control loops placed into manual will cause all loops to stop oscillating. This will be
discussed later.
INTERNALLY-CAUSED OSCILLATIONS
Oscillations generated by a control loop itself can be caused by faulty final control element (e.g.
control valve or damper) or by tuning. Generally, if the oscillation is caused by poor tuning, the
process variable will oscillate with a reasonably smooth sine-wave pattern, and the oscillation will
often grow in amplitude until either the controller output or the process variable periodically runs into
its limits. If the oscillation is caused by final control element problems, the trends are more likely to be
shaped like a square wave or saw tooth wave. However, this is a guideline and not a definitive test. If
the control loop drives a final control element, the performance of the latter should be checked first by
doing simple valve diagnostic tests before attempting to tune the controller. This is especially true if
the control loop used to work properly, and is now oscillating without any changes to the controller
settings.
The most common equipment-based causes of oscillations are control valve (or damper) related. Note
that the discussions below sometimes mention only control valves for the sake of brevity. However,
dampers can cause the same problems with the same symptoms as control valves. So where only
control valves are mentioned, the same arguments will also apply to dampers.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
STICTION
A common problem found in final control elements is stiction. This is short for static friction, and
means that the valve internals are sticky. If the stem of a valve with stiction comes to rest, it tends to
stick in that position. Additional force is then required to overcome the stiction.
The integral control mode will continue to change a controllers output in an attempt to get the process
variable to its set point. While the valve is sticking, the process remains deviated from set point but
additional pressure builds up in the valve actuator. If enough pressure has built up to overcome the
static friction, the valve breaks free, and travels to the new controller output which is now far beyond
its original value. This causes the process to overshoot its set point. Then the valve sticks at the new
position, the controller output reverses its direction of travel and the whole process repeats in the
opposite direction. This causes an oscillation, called a stick-slip cycle. If loop oscillations are caused
by stiction, the controller outputs cycle often resembles a saw-tooth wave, while the process variable
may look like a square wave or an irregular sine wave.
Figure 3. A flow loop with a stick-slip cycle. (PV is process variable, CO is controller output, and SP is set point.)
Stiction might be caused by an over-tight valve stem seal, by sticky valve internals, by an undersized
actuator, or a faulty positioner. Stiction can be detected by placing the controller in manual mode and
making small changes (0.5% is recommended) in controller output and monitoring the process variable
for a resulting change. If the control valve seems to accumulate a few of the controller output changes
before the process variable shows movement, it has stiction.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
Figure 4. A stiction test revealing the presence of control valve stiction.
POSITIONER OVERSHOOT
Because of the widespread adoption of positioners for accurately positioning control valves and
dampers, one problem that is more common now than a decade ago, is that of positioner overshoot.
Positioners are fast feedback controllers mounted on the final control element to measure the valve
stem or damper vane position and manipulate the actuator until the desired valve position is achieved.
Most positioners can be tuned. Some are tuned too aggressively for the valve or damper they are
controlling. This causes the device to overshoot its target position after a change in controller output.
The positioner and control valve form a closed-loop system that may cause the valve or damper
position to hunt around or even oscillate. The positioner could simply be defective and cause the valve
to overshoot. Or the valve can be sticky and the positioner is simply trying to overcome the stiction. If
the controller on a fast-responding loop like a flow control loop is tuned aggressively, the combination
with positioner overshoot can cause severe oscillations in the control loop. Positioner overshoot can be
detected on fast-responding loops by placing the controller in manual and changing the controller
output by two to five percent.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
Figure 5. A faulty positioner causing the process variable to overshoot and hunt around after a change in controller output.
DEAD BAND AND LEVEL LOOPS
If a level controller drives a valve directly (i.e. no cascade control), and the valve has dead band, the
loop will continuously oscillate. Valve dead band will be discussed later. A level control loop will also
oscillate if the controller has an internal dead band around the set point. This is sometimes done to
prevent the controller from reacting to process noise, but it should not be done on level loops because
of the continuous oscillation it causes.
TUNING
A loop that is tuned too aggressively (overly fast response) can quickly develop oscillations. Step tests
should be done on the process to determine the dominant process characteristics: process gain, dead
time, and time constant. A step test is done by placing the controller into manual mode and changing
its output by a few percent (between two and five percent is normally sufficient). Three or more step
tests should be done to compare the results, throw out outliers, and use the average.
Proven, broad-spectrum tuning rules like the Cohen-Coon or Lambda tuning rules should be used to
calculate new controller settings. However, many tuning rules are too aggressive in their original form,
and it is recommended to use only half of the calculated controller gain. Best practices prescribe using
tuning software for analyzing step-test data and calculating new controller settings.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
Figure 6. A step test showing measurements to be taken for controller tuning.
CYCLICAL INTERACTION
Interaction between loops with similar dynamics can cause the two loops to fight each other. One
example of cyclical interactions is when liquid pressure and flow are controlled on the same line. This
is often done with a pressure-reducing controller controlling one valve, and a downstream flow
controller controlling another valve. Both controllers affect the flow and the pressure, and a cyclical
interaction between the two loops can easily occur.
Cyclical interaction is aggravated by aggressive tuning, for example when using the Ziegler-Nichols or
Cohen-Coon tuning rules.
To solve problems with cyclical interactions, control loops have to be tuned less aggressively. Using
the Lambda method results in very stable control loops. One can think of highly interactive control
loops as a tub filled with water. If you drop a stone in the tub, lots of waves result that take a while to
stabilize. Using the Lambda tuning method is like replacing the water with oil. Now if you drop the
stone into the tub, the oil just absorbs the disturbance and no persistent wave action results. The pulp
and paper industry has highly interactive processes, and it has had great success using the Lambda
tuning method.
RANDOM DEVIATIONS
In contrast to oscillations that are periodic, poor control can also make itself evident in large but
random deviations between the process variable and set point. These could be measurement noise,
process disturbances or rapid set point changes. To understand how a control loop is capable of
handling disturbances, we need to look at the speed of response of a control loop.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
LOOP SETTLING TIME
There are several measurements for loop response; settling time will be used here. Settling time can be
defined as the duration of time during which a deviation between set point and process variable is more
than 5% of the size of the deviation. The settling time of a control loop cannot be infinitely short. If a
control loop is tuned sluggishly, it will have a long settling time. If the tuning is improved, the settling
time will be reduced, but only up to a point. If the tuning is made any faster, the loop will become
cyclical and the settling time will increase.
Figure 7. The settling time of a control loop has a minimum limit.
The minimum settling time of a control loop is determined mostly by the amount of dead time in the
process. For a flow loop, the settling time is about three times the dead time, for a temperature loop its
between three and four dead times, and for a level loop it is about four dead times.
MEASUREMENT NOISE
Measurement noise is random, rapidly-changing deviations from set point. The rate at which this
happens is so much shorter than the loop settling time that it is impossible for the controller to
eliminate noise or even reduce its amplitude. A controller responding aggressively to noise will likely
increase the average deviation size. The amplitude of noise can be reduced through filtering the
process variable with a first-order lag filter or a moving-average filter. It is important to note that a
filter increases the apparent dead time of a loop and therefore increases its settling time. Filtering
should be applied only when needed, and then as little of it as possible.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
DISTURBANCES
A process disturbance can push the process variable away from its set point. Disturbances are often the
nemesis of good loop performance. As described above, feedback control is limited in how fast it can
eliminate the effects of a disturbance and bring the process back to set point. If a disturbance occurs
much slower than the settling time of a control loop, feedback control should be able to significantly
reduce its amplitude. If not, it may be a problem with the final control element or the tuning of the
controller. One should first check for final control element problems before tuning the controller.
DEAD BAND
Dead band (sometimes called hysteresis), reduces the effectiveness with which a controller can
counteract disturbances. Every time the direction of a disturbance changes, the controller output has to
traverse the entire dead band before the final control element begins moving. Dead band can be
detected very reliably with a simple process test consisting of two controller output steps in one
direction and one step in the opposite direction with the controller in manual mode. The second and
last steps should be the same size. If the process variable does not reach the same level after the first
and third steps, it indicates the presence of dead band. Dead band is a mechanical problem and cannot
be addressed with tuning.
Figure 8. A dead-band test revealing the presence of dead band.
SATURATION AND RATE-OF-CHANGE LIMITS
A control loop may also appear to have sluggish response if the controller output becomes saturated at
its upper or lower limit. If the controller output is constrained by a rate-of-change limiter, it also may
cause sluggish response regardless of how well the controller is tuned. Alarms can warn of these
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
conditions or historical time-trends of the controller output and process variable can be reviewed to
find their presence.
TUNING
Once the final control element has a clean bill of health, the controller tuning should be reviewed to
see if it is perhaps sluggish tuning that reduces the controllers effectiveness in counteracting
disturbances.
The controller-tuning advice given earlier applies here too. Note that the Lambda tuning method
results in stable control loops, but often cause a sluggish response to disturbances, especially on slow
temperature loops. Cohen-Coon tuning provides faster disturbance rejection.
Although correct tuning methods can go a long way in minimizing the effects of disturbances,
disturbances sometimes happen so rapidly that feedback control alone is unable to reduce their effects
to reasonable levels. The fastest possible feedback control action is limited by the makeup of the
process dynamics. Once this limit has been reached, other solutions must be sought to obtain further
improvement in performance. It is sad to hear of personnel spending days and even weeks tweaking a
control loop that is already at the limit of its performance capability.
CASCADE AND FEEDFORWARD CONTROLS
If a disturbance occurs faster than the control loop can respond, there is very little the controller can do
to reduce its amplitude. In cases like this the feedback controller can be greatly augmented with
cascade and feedforward control.
Cascade control should be applied whenever a slow-responding loop like temperature or level controls
liquid or gas flow. The flow should be controlled directly with a flow controller of which the set point
is set by the level or temperature controller.
Feedforward control should be applied when large, measureable disturbances affect the process
variable. The feedforward controller uses the disturbance measurement to generate an equal but
opposing control action to minimize the effect of the disturbance on the process variable.
INTERMITTENT PROBLEMS
Some control problems seem to come and go with time intermittent problems. These problems are
more difficult to track down and solve, but it helps to know what causes to look for.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]
NONLINEAR VALVE CHARACTERISTIC
Many control valves and most dampers have a nonlinear installed characteristic. This means that the
flow characteristic of the device changes depending on how much open it is. If tuning is done with the
valve or damper at the one end of its travel, the settings might not work at the other end and could
cause oscillations or sluggish behavior. If this is the case, a function generator (X-Y curve) can be
placed in the path of the controller output to cancel out the control valve or damper nonlinearity.
NONLINEAR PROCESS
Many processes react differently based on operating conditions, product type, production rate, etc. In
many cases the differences in process characteristics are large enough to affect control loop
performance. For example, the process gain of a heat exchanger is much less at high product flow rates
compared to low flow rates. These changes in process characteristics often require different tuning
settings for optimal control at various operating conditions. However, this is seldom implemented,
leaving the control loop with poor response for most of its operating range. On systems with varying
process characteristics, controller tuning should be altered automatically based on the operating
conditions. This is accomplished quite effectively by implementing gain scheduling. Gain scheduling
uses the operating condition as an input to one or more function generators to dynamically adjust the
controller gain, and sometimes also the integral time, and derivative time if used.
CONCLUSION
Obtaining robust and optimally-performing control loops can be challenging at times. Control loops
can perform sub-optimally due to a variety of reasons and controller tuning alone is in many cases not
the ultimate solution for poor control performance. Through a simple but systematic analysis of the
control problem, the root cause of poor control can be established and the problem can be resolved or
at least minimized in the most effective way.
Distributed with permission of author(s) by ISA, 2011
Presented at 2011 ISA Automation Week; [Link]