You are on page 1of 7

2.

24 Real-Time Performance Assessment


J. P. GERRY (2005)

INTRODUCTION a key variable to be assessed, whereas in refineries the aver-


age temperature error can be a key of assessment.
To optimize the regulatory control loops in a plant, one first The definition of loop health can also vary from loop to
needs to assess the current performance. How are the loops loop. For example, a flow loop may have different perfor-
operating now? Could they be performing better? By assess- mance objectives than the level in a surge tank. A flexible
ing the performance in real time, a number of benefits can definition is the best since it allows for improvements and
be obtained: refinements over time.
Loop health can be evaluated by considering the key
• Knowledge, on a real-time basis, of the health of the
plant, unit operations, and loops can be acquired. assessments, baselines, thresholds, and economic significance.
• Poor-performing loops can easily be flagged.
• Continual performance assessment allows one to Assessment Criteria of Loop Health Assessments are cal-
record the history of performance, so that differences culated over “assessment intervals.” This interval should be
in performance between time periods can be com- defined for each unit operation and should be at least 6 times
pared. For example, one can evaluate the effect of the longer than the longest cycle in the unit or at least 100 times
following variables on performance: the longest process dead time. Generally, the assessment
• Seasonal changes interval should be 4 hours or longer.
• Grade changes Loop health can be evaluated and assessed on the basis
• Feed changes of many criteria, and the reader should be familiar with the
• Shift changes related terminology, described below in alphabetic order. Any
• Equipment changes of the following criteria or their combinations can be used
as key assessments, which refer to criteria that are relevant
The tools of optimization should be readily accessible in the
to the health of a control loop:
performance monitor.

• Average error is the average of the absolute error that


PLANT ASSESSMENT AND MONITORING has existed between the setpoint and measurement
during the assessment period.
Process variable, set-point, and controller output data should • The Harris Index (normalized) is defined as the ratio
be collected from the plant for every control loop. In order between the error (PV − SP) variance during this
to maintain accuracy, these data should be collected while assessment over the variance achievable by a mini-
all software compression is turned off. Ideally, the data mum variance controller. The larger the value of the
should be collected at intervals of roughly 1/10 of the dead Harris Index, the poorer the performance of the loop.
time of the loop. Generally once per second is adequate, but The Harris Index will vary as a number between one
the collection frequency of some loops such as temperature and infinity. A value of one is perfection or minimum
can be longer. As of this writing, standard connection meth- variance control. As the quality of loop performance
ods are available using OPC or Ole for Process Control, drops, the index rises. However, sluggish tuning usu-
which is a standard developed by the OPC Foundation ally causes the Harris Index to drop. Cyclic response
(www. opcfoundation.org). or overly aggressive tunings will make the Harris
Loop Performance and Health Assessment Index rise. The Harris Index is considered to be high
when it is between 50 and 100, but it can get much
The overall performance of a control loop or loop health may higher. However, these numbers may vary from loop
be different for different businesses. It can also vary from to loop and also depend on the occurrence of set-point
plant to plant depending on the business objectives. For changes and load upsets. The amount and pattern of
example, in paper mills, the variability of the paper is often noise can greatly influence the Harris Index.

311

© 2006 by Béla Lipták


312 Control Theory

• Noise band is determined on the basis of statistical • Variability is the relative value of variance (defined in
analysis of the noise, which during the assessment the next paragraph) in the process variable during the
period is superimposed on the process variable, assessment period. It is expressed as a percentage of
expressed in percentage. The noise band is the stan- the mean and so allows comparison between the level
dard deviation (sigma) of the process variable during of variability in different processes.
the quietest time. The minor variations in the process
variable that are not caused by the controller, by set Variability = 2(Standard Deviation)(100/Abs(Mean))
point changes, or by load upsets are considered to be 2.24(1)
normal process noise. Such noise may be caused by
where Abs = absolute value; Standard Deviation =
electrical interference, magnetic fields, flow turbu-
sqrt(Variance); and Variance = sum(Mean – x(i)) /
2
lence loops, waves in a tank in level loops, etc. Gen-
(npts − 1).
erally, the larger the noise band, the poorer the
• Variance measures the spread or dispersion of the
performance of the loop. The cycling of a process
process variable during the assessment period. “Vari-
variable in an unstable and fast control loop might also
ance, normalized” is dimensionless because it has
appear as noise to this assessment.
been normalized with 100 divided by the span of the
• Oscillation detection confidence value is a value between
PV, squared:
0 and 100%, where 100% indicates full confidence that
the loop is continuously oscillating. This assessment is
Variance = [sum(Mean − x(i)) (100) ]/[(npts − 1)
2 2
performed by analyzing the pattern of the error signal.
× (span of PV) ]
2
• Oscillation diagnosis confidence value is between 0 2.24(2)
and 100%, where 100% indicates full confidence that
the oscillation is caused by one of the three causes:
Related Assessments Criteria
valve stiction or hysteresis, bad tuning, or causes exter-
nal to this loop. The following assessments provide beneficial diagnostics but
• Output standard deviation is the standard deviation in are not contributors to loop health.
the controller output during the assessment period.
• Output at a limit is the percentage of time that the • Three largest oscillation periods and strengths. This
controller output was at a limit. value is obtained by spectral analysis of the error data
• Probable Performance Increase Index is a measure of during the assessment time. The three oscillations of
the performance increase, which is possible by using the the process variable having the most power (and their
fastest PID tuning with a safety factor of 2.5. It is the amplitudes) provide the six assessments. The power
“percent probable” performance increase in integrated should be evaluated using a cluster of peaks and not
absolute error (IAE) terms, when the loop is responding just a single peak, and the zero frequency component
to a load upset, assuming the controller does not over- should be ignored. This assessment can provide some
shoot with either settings. This index shows the improve- very beneficial diagnostics but is not viewed as a con-
ment in process response to upsets with the new tuning. tributor to loop health. For example, if the cause of
This index is often proportional to the savings that the loop oscillation is unknown, by sorting the loops by
plant can achieve from improved control loop tuning. oscillation period, one can identify the loops having
• Set-point crossings is the number of times the process the same period of oscillation — even when it is “hid-
variable (PV) crosses the set point in one day. den” in other information. This can help in pinpointing
• Time in Normal is the term for the time period in the cause of oscillation.
percentage that the loop spends in the Normal mode. • Process model parameters. These are parameters of
Most often the Normal mode is the Automatic mode, the process model — dead time, gain, lag times, and
but not always. integration time, including second-order coefficients
• Valve travel is the total distance the controller output that allow for imaginary roots. The process model is
moves during the assessment period, which is normal- usually not a key assessment of control loop health.
ized to be on a 24-hour basis. It is found by summing • Model quality. This is an assessment of the accuracy
the absolute value of the controller output changes that of the fit between the actual process and its model.
occur between each sample. The quality of the process model is usually not a key
• Valve reversals is defined as the number of times the assessment criterion in the evaluation of the health of
controller output changes direction during an assess- the control loop.
ment period normalized to a 24-hour basis. It is a value • Robustness. Robustness is a relative number giving an
between zero and infinity. This value is of interest indication of how robust or sensitive the loop is to
because more reversals usually result in more valve process upsets or changes. The larger this number, the
wear. more robust the loop. There is always a tradeoff

© 2006 by Béla Lipták


2.24 Real-Time Performance Assessment 313

between fast tuning and robustness. The robustness bidirectional. For example, a threshold may be set above the
number is discussed in detail in Section 2.26 in this baseline and/or below the baseline.
handbook. The robustness number is normalized since For example, a threshold for variability might be calcu-
it uses gain and dead time ratios. It can be used as a lated by looking at a number of assessments from a period
comparison between loops. when the plant was running well. In that case, one should find
• The relative response time (RRT) is a relative indicator the standard deviation and average value of the variability
of the speed of the control loop. The smaller the rel- during these assessments. The standard deviation is also called
ative response, the faster the loop. The higher the the sigma value. X number of sigma values over the average
relative response value, the slower the loop. This num- value is called the X sigma value. So setting the threshold of
ber is used for comparison purposes. The RRT is variability to six standard deviations over the average would
dependent on the controller’s speed of response and be setting it to the six sigma value. This is one strategy for
can be changed by adjusting the PID settings. If one setting thresholds.
uses fast tuning, the controlled process variable often Thresholds could also be set to the minimum or maxi-
will oscillate around the set point after it is changed. mum value of an assessment over the baseline time.
The period of this oscillation in the time response is
roughly the same as the RRT. The value of the RRT Loop Health Calculation The calculation of loop health
is found by: should be flexible enough to allow for different interpretations
1. Calculating the closed loop frequency response to a depending on the nature of the plant or process. It should
load upset allow for different categories of loops to have different health
2. Finding the frequency where the amplitude ratio requirements. For example, a temperature loop may need to
peaks be fast and responsive whereas a level loop may need to allow
3. Converting this to a time period for large level variations before causing a change in flow.
Interacting loops and RRT. In order to prevent inter- Loop health calculation should also allow different
action, the PID tuning settings should be so adjusted loops to have different priority ratings based on the relative
that the RRTs of the loops differ by a factor of three. importance of the loops.
For example, if there are three interacting loops and
the fastest one has a period of 1 second, the tuning of
Key Assessment Criteria
the other two controllers should be readjusted so that
their RRTs are at least three and nine. In order to quantitatively evaluate the performance of a loop,
Cascade Loops and RRT. For cascade loops, the one has to select assessments that in combination define the
master (outer) loop should have an RRT that is three loop’s health. These are called key assessments. The key
times slower than that of the inner loop. This relation- assessments should be those that are important to the oper-
ship is obtained by first tuning the slave (inner) loop, ation of the plant, and any assessment can therefore be
and then the master. If the cascade master loop has an selected to be a key assessment.
RRT that is less than 3 times that of the slave (inner)
loop, the master should be de-tuned so that its RRT Percent towards Threshold Every assessment can poten-
will be 3 times that of the slave. tially provide a contribution to describing the health of the
loop. For each assessment during every assessment period, a
Baseline and Threshold Values The baseline time is a win- “Percent towards Threshold” is found:
dow or period of time when the plant is running well. The
baseline time can be used as a reference for comparison. The Percent towards Threshold = (Current Value − Baseline)/
length of the baseline period should exceed six and should (Threshold − Baseline)100
preferably equal 30 assessment periods. For statistical meth- 2.24(3)
ods of evaluation, at least 30 assessments are suggested.
Every assessment can have a baseline value. A baseline Next, an “Average Percent towards Threshold” is found. The
value represents either the ideal value or the value from a Average Percent towards Threshold is the average of all the
time when the plant was running well. For example, a base- Percent towards Threshold values for all the key assessments
line value for average error might be zero, since zero error during an assessment period. Using this calculation, lower
is ideal. The baseline for process dead time might be the values of Average Percent towards Threshold mean a better-
average or minimum dead time that is present during the performing loop. Larger values mean a poorer-performing
baseline time period. loop. A patent is pending for this technology, which is used
Thresholds represent values of the assessed criteria that in performance assessment.
one would prefer not to exceed. Thresholds represent limits
or boundaries between which the assessment criteria would Loop Health = (Average Percent towards Threshold)/
remain if the plant is running well. All thresholds should be (Economic Significance) 2.24(4)

© 2006 by Béla Lipták


314 Control Theory

Economic Significance is a divisor that incorporates a factor of Results should be shown in both graphical and tabular
economic importance to the health of each loop. For example, form. All of these results should be up to date on a real-time
if two loops are exhibiting the same “Average Percent towards basis as much as possible.
Threshold” the one with the higher economic significance
(lower divisor) will bubble to the top of a priority list of loops. Plant Overview

An overview window should provide an assessment of loop


Historization of Assessments The assessments for every
health in the various sections of the plant and also in the overall
loop should be historized. This allows for performance com-
plant. Each plant area should have a history of assessments in
parisons at different time periods. This information can be
that area. For example, Figure 2.24a shows a table and plot of
very valuable in troubleshooting and optimizing operations.
loop health for 1 week. The two lines at the top and bottom
represent two plant sections (unit operations), while the line
Accessibility and Presentation of Results Currently, the most
in the middle describes the health of the entire plant. In the
accessible method for distribution of the results seems to be via
table below the chart, the health of the two unit operations
a browser-based interface. This means the performance moni-
making up the plant is listed as the third and fourth rows.
toring system should be a Web server; it should be able to serve
up the performance results in an easy-to-read format for plant
Prioritization of Problem Areas
personnel. Most plant personnel are already familiar with nav-
igation in browser interfaces and Web pages. Any personnel The interface for users of assessment software should allow
with access to a computer that is connected to the plant local for default and for the generation of priority lists, which are
area network (LAN) will have access to such information. configured by the user. The user should be able to sort the
A network bridge allows this same connectivity on the control loops in the plant on the basis of any combination of
office LAN. A network bridge uses two network interface cards assessment criteria chosen by the user.
in the server machine running the plant monitoring software.
Using two cards enables the plant LAN to be totally separate Priority List of Loops Having Most Impact One of the lists
from the office LAN. A Web server running the monitoring that is generated includes, in their order of priority, the loops
software can send assessment results to the office LAN. needing attention. Both the engineering and plant maintenance

Boiler feed 45%


39% 39% 39% 38% 39% 39% 38%
preheater 40%
Condensate 35%
unit
30%
Entire plant 25% 25% 26% 25% 25% 24% 25%
25%
20%
15%
10%
5%
0%
4/7/2003 4/9/2003 4/11/2003 4/13/2003
Time
Latest data on the left Latest data on the right

Location 4/7/2003 4/8/2003 4/9/2003 4/10/2003 4/11/2003 4/12/2003 4/13/2003


Entire plant 33.3% 33.7% 33.6% 33.0% 33.7% 32.8% 32.5%
Boiler Feed Preheater 24.9% 24.9% 25.9% 25.1% 25.4% 23.9% 24.6%
Condensate Unit 38.9% 39.5% 38.8% 38.3% 39.2% 38.8% 37.8%

FIG. 2.24a
1
Visual and tabular overview of the health of unit operations in a plant.

© 2006 by Béla Lipták


2.24 Real-Time Performance Assessment 315

Biggest payback loops

Refresh PlantTriage home page Snoozed loops Help

Loop name Unit operation Description Average economic assessment Snooze loop

FIC100 Condensate unit Condensate return 34.0% Snooze

TIC101 Boiler feed preheater Preheat temperature 33.5% Snooze


LIC100 Condensate unit Knockout drum level 23.0% Snooze

FIC101 Boiler feed preheater Preheat steam condensate flow 14.7% Snooze
PIC100 Condensate unit LP Steam makeup 0.5% Snooze

Average economic assessment


FIC100 34.0%
TIC101 33.5%
LIC100 23.0%
FIC101 14.7%
PIC100 0.5%
0 5 10 15 20 25 30 35

FIG. 2.24b
1
A sample list with the control loops listed in order of their need of engineering or maintenance attention.

departments should find this list useful. The order of the loops Custom Loop Lists The loop performance assessment soft-
in the list should reflect their potential for improving the pro- ware allows users to create their own lists of loops. The rows
ductivity of the operation. To allow the user to obtain more and columns are selected by the user and can contain both
information about the various unit operations, a clickable “drill custom loop lists and custom assessment lists as they are
down” can be provided, with more detail about the area in chosen by the user. Once set up, the software allows the user
question. Figure 2.24b shows an example of such a page. to recall the custom loop lists.

Listing Oscillating Loops The system should also include


a list of oscillation assessments. Time of oscillation detection, Recording Tuning Changes PID tuning values and changes
diagnosis, period, and strength are some of the valuable data are recorded by the performance monitor. A time window of
collected. Sorting by “oscillation detection” shows all loops the history of these changes is available for viewing in the
that are oscillating. browser-based Web interface. This feature also allows the
Sorting by oscillation period groups the loops by their user to enter notes or events. When several tuning adjust-
periods of oscillation. If the oscillation of several loops is ments are made to a PID loop, this information can be used
caused by the same problem, they are likely to have similar as an indication that a problem exists with that loop.
oscillation periods. The top three oscillation periods for each Figure 2.24c shows events that were automatically picked
loop should be included in the sort so that any secondary up by the performance monitor and also shows some user-
oscillations can be picked up. entered events. The data in the rows containing PID settings
were detected and automatically entered by the system.
Model and Quality Loop List A list containing columns of
Model Quality and Probable Performance Increase Index Performance Monitoring Software Capabilities
makes it possible to improve the tuning of many PID loops
all at once. In the list, high model quality means that good A brief listing is provided here of the types of testing and
identification exists and hence good PID tuning is possible reporting that should be available as part of the performance
and is readily available for that loop. For those loops with a monitoring software:
high model quality and where the Probable Performance
Increase shows that better tuning is available, the improve- • Stiction and hysteresis testing and reporting. Many
ments can be made to several loops simultaneously. plant oscillation problems are caused from poorly
In such cases, the user, using the analysis software for each operating valves
loop, would request the data used by the performance moni- • Valve wear analysis. This analysis compares valve
toring software to arrive at the model. As is discussed in more travel and reversals with differing PID tuning values
detail later, the analysis software displays the optimal PID and with and without the use of various types of process
tuning along with time simulations and robustness analysis. variable filters.

© 2006 by Béla Lipták


316 Control Theory

Event log

Event Log over a 2 week period.


Ending Thursday April 3 2003

Loop, unit
operation or Loop types Pattern match
group

All loops All

Refresh PlantTriage home page Help New event Submit

Time Loop P I D Author Summary


Sort Sort Sort Sort Sort Sort Sort

8:05 AM Tank
LIC100 Joe.Blogs@expertune.com
4/3/2003 Shakedown
5:40 AM Moved
PIC100 john.gerry@expertune.com
3/30/2003 probe
10:33 AM P. I
FIC101 100 0.2 0 john.gerry@expertune.com
3/28/2003 changed
Re-
10:27 AM
LIC100 Joe.Blogs@expertune.com calibrated
3/28/2003
xmitter
10:24 AM unknown - please P. I
3/28/2003 PIC100 12 0.64 0
acknowledge changed
9:24 AM unknown - please P. I
3/28/2003 TIC101 350 1.4 0
acknowledge changed

FIG. 2.24c
1
Part of an event log listing.

• PV filter analysis. This analysis finds the largest pos- • Simulating the loop performance using a variety of
sible filter that will reduce valve movement but will tuning settings and PV filter values. Simulations must
not inhibit speed of response or reduce robustness. include both set-point response simulations and
• Unfiltering capability. The software should have the responses to load upsets. The simulations should dis-
ability to remove the effects of a PV filter so the play and compare the performance with both the exist-
unfiltered PV can be viewed graphically by the user ing and the proposed settings. Simulations should
and the analysis software. describe the response of the loop to noise and should
• Model identification and display. The model is used compare the response of the model of the process to
by the analysis software for time simulations, robust- that of the real process.
ness analysis, and valve wear analysis. The model • Linearization analysis including characterization and
should be identified with minimal prior user knowl- characterizer builder. This allows linearization of non-
edge of the dynamics; the software should find the linear loops through the use of a characterizers. This
best model structure and dead time without the user software should be suitable to both the types of pro-
estimating these. The model can be used by model cesses whose linearity changes with load and those
predictive controllers. whose nonlinearity depends on setpoint (pH).
• Optimal PID tuning. A variety of methods including • Power spectral density. This allows a detailed anal-
set point or lambda tuning should be used to minimize ysis plot of power in a signal at all frequencies or
load upsets (minimum IAE). periods.
• Robustness. There is always a tradeoff between tuning • Statistical analysis including a histogram. This allows
for fast response and sensitivity to process load for the generation of statistical comparisons to evalu-
changes. Robustness analysis is critical to providing ate and document the benefits of work done.
practical PID tuning settings for the loop. Robustness • Time line analysis. This analysis describes time rela-
analysis should reflect the changes that occur in pro- tionships between components.
cess dead time and process gain. • Integrated reporting.

© 2006 by Béla Lipták


2.24 Real-Time Performance Assessment 317

Oscillating 110
Osc - Hardware 100
Osc - Tuning 90
Osc - Load 80
70
60
50
40
30
20
10
0
1/26/2003 2/14/2003 3/6/2003 3/25/2003
Time

FIG. 2.24d
The oscillation of a flow loop on a paper machine was detected by the assessments of “oscillation detection” and “oscillation from tuning,”
1
signaling a need to re-tune the PID when the paper grade is changed.

CASE STUDIES diagnosis that the oscillation was caused by the PID control-
ler with the wrong tuning settings.
Model Predictive Controlled Refinery Later, it was learned that the plant changed the paper
grade, which would have required re-tuning and resulted in
In a refinery, model predictive controllers (MPC) tend to oscillation. It was noted that oscillation can be avoided if the
move the process toward its constraints as far as possible to tuning settings are automatically changed whenever that
maximize the resulting economic gain. How close the con- grade of paper is run.
troller can push the process to its constraints is related to set-
point crossings, average absolute error, and variability. These
criteria of assessments infer the ability of the control loop to CONCLUSIONS
closely approach the constraint. High-quality performance is
indicated when this assessment value is low and if the base- Assessing the performance of a plant in real time requires
lines are lower than thresholds. the careful selection of the appropriate software, but the
In order for an MPC controller to work, the loops tied to benefits are also compelling. Process plants will benefit from
it need to be in their normal operating mode and not at a considering this option.
limit. The assessments “time in Normal” and “output at a
limit” assess the performance of these. Time in Normal
should be maximized and should have a high baseline and Reference
low threshold whereas output at a limit should be minimized:
it should have a low baseline. 1. These visual aids are provided courtesy of ExperTune Inc.
In a fluidic catalytic cracker the side valves are very
expensive — costing upwards $100,000 each, and the entire unit
Bibliography
must be shut down when the valve needs service. Therefore,
the assessments of “valve reversal” and “valve travels” can be Gerry, J., “Prioritizing and Optimizing Problem Loops Using a Loop
used to measure the wear and tear on these valves. The lower Monitoring System,” ISA, Chicago, October 2002.
the assessment values the better. Baselines should be low and Gerry, J., “Process Monitoring and Loop Prioritization Can Reap Big
thresholds high for these assessments. Payback and Benefit Process Plants,” ISA, Chicago, October 2002.
Mansfield, E., Basic Statistics with Applications, New York: W.W. Norton,
1986.
Grade Change in a Pulp Mill Paulonis, M., and Cox, J., “A Practical Approach for Large-Scale Controller
Performance Assessment, Diagnosis, and Improvement,” Journal of
Figure 2.24d describes the assessment of the oscillation of a Process Control, Vol. 13, 2003.
flow loop in a pulp mill. The two lines reflect the assessments Ruel, M., “Learn How to Assess and Improve Control Loop Performance,”
of “oscillation detection” and “oscillation from tuning” or a ISA, Chicago, October 2002.

© 2006 by Béla Lipták

You might also like