ARTICLE IN PRESS

Control Engineering Practice 13 (2005) 1383–1390

www.elsevier.com/locate/conengprac

Industrial implementation of on-line performance monitoring tools

Tore Hägglund
Department of Automatic Control, Lund Institute of Technology, Box 118, S-221 00 Lund, Sweden
Received 4 November 2003; accepted 20 December 2004
Available online 23 February 2005

Abstract

This paper treats implementation of on-line procedures for automatic performance monitoring. Two on-line detection procedures,
one that detects oscillations in control loops and one that detects sluggish control loops, are first summarized. The implementation
aspects of these procedures are then discussed and experiences of their use in various industrial products are demonstrated.
r 2005 Elsevier Ltd. All rights reserved.

Keywords: Supervision; Monitoring; Detection; Diagnosis; Performance; Implementation

1. Introduction paper industry, see e.g., Perrier and Roche (1992),

Lynch and Dumont (1996), and Owen, Read, Blekken-
Many control loops in process control plants perform horst, and Roche (1996). It has also been applied in the
poorly. The reasons may be equipment problems such as chemical industry, see e.g., Stanfelj, Marlin, and
sensor faults or stiction in control valves, or bad MacGregor (1993) and Thornhill, Oettinger, and
controller tuning. There is an increasing understanding Fedenczuk (1999). Conclusions about the control loop
of the fact that badly performing control loops cause performance can also be deduced from spectral analysis.
losses in production as well as quality. See, e.g., Examples are given in Desborough and Harris (1992)
Bialkowski (1993), Ender (1993), and Paulonis and and Tyler and Morari (1996).
Cox (2003). Therefore, there is an industrial interest in There are several reasons for bad control loop
performance monitoring tools that detect and make performance. One reason is stiction in the control valve,
operators and maintenance staff aware of badly see Bialkowski (1993). This results in stick-slip motion
performing control loops. A good overview of perfor- and oscillations. These oscillations can be detected by
mance monitoring tools is given in Qin (1998). methods like the one presented in Hägglund (1995) and
There is also an increasing interest in off-line Thornhill and Hägglund (1997).
procedures and plant auditing. The Harris index, see Another important reason is improper controller
Desborough and Harris (1992), has received lots of tuning. Most control loops in the process industry are
attention. In this method, the control loop performance conservatively tuned, resulting in sluggish responses to
is compared with an ‘‘optimal’’ performance, where load disturbances, and therefore unnecessarily large and
optimal in this case means minimum-variance control. A long deviations from the set point. This method of
modified version of the Harris index, where a less operating the process plants results in decreased product
aggressive optimal control design is considered, is given quality.
in Horch and Isaksson (1999). The Harris index and Why are the controllers conservatively tuned? The
modifications of it have been applied in the pulp and main reason is the lack of knowledge about the
importance of efficient controller tuning and how it
Tel.: +46 46 222 8798; fax: +46 4613 8118. can impact the operation efficiency. The engineers tune
E-mail address: [email protected]. the controllers until they are considered ‘‘good enough’’.

0967-0661/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2004.12.006
 ARTICLE IN PRESS
1384 T. Hägglund / Control Engineering Practice 13 (2005) 1383–1390

They have not got the time to optimize the control. Process output y and set point ysp
Many controllers are tuned once they are installed, and
then never again. To retain stability when operating
conditions change in non-linear plants, the controllers
are tuned for the ‘‘worst case’’. A better solution would,
of course, be to use gain scheduling and perhaps
adaptation. When a controller is retuned, it is mostly
because the process conditions cause oscillatory control.
In other words, when the controllers are retuned, they
are detuned. When the process conditions change to
sluggish control, the controller is normally not retuned
again. Sluggish control loops can be detected by the
method presented in Hägglund (1999).
This paper will first briefly review the two monitoring
procedures for detecting oscillating loops and sluggish
loops, respectively. These procedures can be implemen-
ted in several ways, on line or off line, in the controller Time
or DCS system that performs the control or in an
Fig. 1. Process output y and set point ysp (constant) during a period of
external computer. The main goal of this paper is to good control followed by a load disturbance.
explore these implementation aspects, and present the
requirements that are discovered through the experience long period without zero crossings occurs. This means
of their use in industrial control systems. that the IAE becomes large. When the IAE exceeds a
certain limit, IAElim ; it is therefore likely that a load
disturbance has occurred.
2. Monitoring tools The underlying idea of the oscillation-detection
procedure is to conclude that an oscillation is present
This section gives a summary of the two performance if the rate of load-disturbance detections becomes high.
monitoring tools. The behaviour of the control performance is monitored
over a supervision time T sup : If the number of detected
2.1. Oscillation detection load disturbances exceeds a certain limit, nlim ; during
this time, it can be concluded that an oscillation is
There are several reasons for oscillations in control present.
loops. They may be caused by excessively high controller The oscillation detection procedure has three para-
gains or oscillating load disturbances, but the most meters that must be set: IAElim ; T sup ; and nlim : In
common reason for oscillations is friction in the valve. A Hägglund (1995), the value of nlim is set to nlim ¼ 10: The
procedure for detecting oscillations in control loops was supervision time should be proportional to the time
presented in Hägglund (1995). constant of the loop. In Hägglund (1995) it is suggested
The principle behind the detection procedure is to to set T sup ¼ 5nlim T u ; where T u is the ultimate period
study the magnitude of the integrated absolute error obtained from a relay auto-tuning experiment, see
(IAE) between successive zero crossings of the control Åström and Hägglund (1995). If such an experiment is
error, i.e., not performed, it is suggested that T u be replaced with
Z ti
the integral time T i : Finally, the parameter IAElim is set
IAE ¼ jeðtÞj dt, (1) to IAElim ¼ 2a=ou ; where a as the lower limit of the
ti1
acceptable oscillation amplitude at the ultimate fre-
where eðtÞ ¼ ysp ðtÞ  yðtÞ and ti1 and ti are two quency ou : A suggested value of a is 1% of the signal
consecutive instances of zero crossings. range. Again, if ou is not available, it is replaced by
It is assumed that the controller has integral action, so oi ¼ 2p=T i :
that the average control error is zero. If no integral For on-line applications, it is more convenient to
action is present, the average value of the measurement perform the calculations recursively. Such a procedure is
signal can be obtained using a low-pass filter. summarized in the following program:
During periods of good control, the magnitude of the
control error is small, and the times between the zero INITIALIZATION
crossings are relatively short. See Fig. 1: This means that a ¼ 1 ½%
during good control, the IAE values calculated accord- n_lim = 1;
ing to Eq. (1) become small. When a load disturbance if ultimate_frequency is available then
occurs, the magnitude of eðtÞ increases, and a relatively begin
 ARTICLE IN PRESS
T. Hägglund / Control Engineering Practice 13 (2005) 1383–1390 1385

iae_lim = 2 * a / omega_u; Set point ysp and measured variable y

t_sup = 5 * n_lim * t_u; 0
end else –0.4
begin –0.8
iae_lim = 2 * a / omega_i;
0 10 20 30 40 50
t_sup = 5 * n_lim * t_i;
end; Control signal u
1
gamma = 1 - h / t_sup;
0.5

0
LOAD DETECTION 0 10 20 30 40 50
if sign(e) = sign(e_old) then
begin Fig. 2. Good and bad control of load disturbances.
iae = iae + abs(e) * h;
load = ;; Both responses have an initial phase where the two
end else signals go in opposite directions, i.e., DuDyo0; where
begin Du and Dy are the increments of the two signals. What
if iae 4iae_lim then load = 1 characterizes the sluggish response is, that after this
else load = ;; initial phase there is a very long time period where the
iae = abs(e) * h; correlation between the two signal increments is
end; positive. This observation forms the base for the Idle
index, see Hägglund (1999), which expresses the relation
between the times of positive and negative correlation
OSCILLATION DETECTION between the signal increments.
x = gamma * x + load; From now on, it is assumed that the sign of the static
if x 4n_lim then process gain is known, and for simplicity that it is
begin positive. Further, it is assumed that the control loop is
oscillation = true; subjected to load disturbances only. If there are setpoint
x = ;; changes present, these responses should be excluded
end; from the analysis.
The factor g determines the time horizon in the filter. Its To form the Idle index, the time periods when the
relation to the supervision time is correlations between the signal increments are positive
and negative, respectively, are calculated first. The
h following quantities are updated every sampling instant:
g¼1 , (2)
T sup (
tpos þ h if DuDy40;
where h is the sampling interval. tpos ¼
tpos if DuDyp0;
If the signal range of the process output is known, and (
if the time constant of the loop is possible to estimate, at tneg þ h if DuDyo0;
tneg ¼
least as T i ; then it is possible to run the oscillation tneg if DuDyX0;
detection procedure automatically without any further
process information. This property is of vital impor- where h is the sampling period. The idle index I i is then
tance in process control applications. A modified defined as
version of the method was suggested in Thornhill and
tpos  tneg
Hägglund (1997), where no knowledge of the signal Ii ¼ . (3)
ranges was required. tpos þ tneg
Note that I i is bounded to the interval ½1; 1: A positive
value of I i close to 1 means that the control is sluggish.
2.2. The Idle index The Idle index for the sluggish response in Fig. 2 is
I i ¼ 0:82: A negative value of I i close to 1 may be
Many control loops are conservatively tuned, result- obtained in a well-tuned control loop. The Idle index for
ing in sluggish control. Fig. 2 shows two responses to the good response in Fig. 2 is I i ¼ 0:63: However,
load disturbances in the form of step changes at the negative Idle indices close to 1 are also obtained in
process input. One response is good, with a quick oscillatory control loops. Therefore, it is desirable to
recovery without any overshoot. The second response, combine the Idle index calculation with an oscillation
however, is very sluggish. detection procedure in order to detect these systems. Idle
 ARTICLE IN PRESS
1386 T. Hägglund / Control Engineering Practice 13 (2005) 1383–1390

indices close to zero indicate that the controller tuning is perform the monitoring. In this section, this knowledge
reasonably good. is therefore considered first, and the possible implemen-
The conclusions are drawn under the assumption that tations are then discussed. The implementation issues
the load disturbances are step changes or at least abrupt are also treated in Dumont, Kammer, Allison, Ettaleb,
changes. This is a reasonable assumption in many and Roche (2002).
situations, since load changes are often caused by
sudden changes in production. However, if the load 3.1. Knowledge required
disturbances are varying slowly, the Idle index may
become positive and close to one even in situations when 3.1.1. Signals
the control is not sluggish. To avoid this, it might be Most performance monitoring tools require that the
advantageous to calculate the Idle index only during measurement signal is available, and several tools also
periods when there are abrupt load changes. This can be require that the setpoint and the control signal are
accomplished using load detection procedures, see provided.
Hägglund and Åström (2000). The oscillation detection procedure uses knowledge
A recursive version of the Idle index calculation is about the measurement signal only, whereas the Idle
given by the following procedure, which is updated index tool also requires that the control signal and the
every sampling instant: setpoint are available. The setpoint is needed to exclude
periods when the excitation is caused by setpoint
if DuDy40 then s ¼ 1
changes.
else if DuDyo0 then s ¼ 1 It is important that the representations of the signals
(4)
else s ¼ 0; are appropriate. The signals that enter the control block
if sa0 then I i ¼ gI i þ ð1  gÞs: in the DCS system are normally not the same as the
signals that are available on the AD/DA converters.
The factor g is defined in Eq. (2), where the supervision There might be filters and other functions inside the
time is T sup ¼ tpos þ tneg : DCS system that alters the signals between the AD/DA
The procedure is sensitive to noise, since increments converters and the control block. It is also important
of the signals are studied. It is therefore important to that the signals are sampled with a sampling period that
filter the signals. To do this, it is necessary to have some is short enough. Ideally, the sampling rate should be as
information about the process dynamics to find a high as the sampling rate of the control block in the
suitable filter time-constant. DCS system.
It is also desirable to avoid calculations near steady-
state, when the signal-to-noise ratio is small. A natural 3.1.2. Parameters
way to ensure this is to perform the calculations only There are several parameters that are useful in the
when monitoring tools. For example, the integral time is often
jej4e0 , (5) the only parameter available that describes the time
constant of the loop. Such a time is needed to determine
where e is the control error, and e0 is a threshold based the supervision time both in the oscillation detection
on a noise-level estimate or fixed to a few percent of the procedure and in the Idle index calculation. It is also
signal range. needed to determine suitable time constants in filters.
The supervision time T sup ¼ tpos þ tneg used in Eq. (3) This is of particular importance in the Idle index
or the corresponding factor g used in the recursive calculations.
calculations, Eq. (4), must be determined. As for the The signal ranges are also useful parameters. For
oscillation detection procedure, it is reasonable to example, the signal range of the measurement signal is
determine it as a factor times T u if this time is available, needed to determine the amplitude a leading to IAElim in
otherwise as a factor times T i : If at least one of these the oscillation detection procedure. It is also useful for
times are available, and if the signal ranges of u and y determination of the threshold e0 in the Idle index
are known, it is possible to make the Idle index calculation, see Eq. (5). The signal ranges are also used
calculation automatic without any further process to determine when signals become saturated.
information.
3.1.3. Operating conditions
It is important that the monitoring is performed only
3. Implemented where? during normal operating conditions. Data obtained
during manual control, during periods when the
The on-line performance monitoring tools may be controller is running in tracking mode, or during periods
implemented in various ways. The suitability of these when the signals are saturated should, e.g., be excluded.
implementations depends on the knowledge required to Therefore, it is important that the controller mode of
 ARTICLE IN PRESS
T. Hägglund / Control Engineering Practice 13 (2005) 1383–1390 1387

operation is provided in the performance monitoring the experience obtained from this implementation is still
tools. limited.
In Paulonis and Cox (2003), a performance monitor-
ing system at Eastman Chemical Company is presented,
3.2. Implementation possibilities
where both the oscillation detection procedure and the
Idle index procedure are implemented. In this system
Since there is so much knowledge required to
over 14,000 loops in 40 plants at nine sites worldwide are
implement even such simple performance monitoring
supervised. The user interface is web based.
tools as those two described in this paper, these tools
The detection procedures have also been implemented
cannot be implemented in systems that only have access
in various DCS systems, e.g., from Hartman and Braun,
to signals. They must be implemented either in the DCS
Siemens, and Honeywell, especially in the Pulp and
system that performs the control, or in an external
Paper industry. These systems are more or less suitable
system that has access to the information needed
for this kind of implementation. Some have free-
through, e.g., an OPC server.
programmable blocks, but in most of the applications
An ideal implementation platform would be the DCS
the programming has been performed using the stan-
system. These systems are, however, limited in several
dard blocks already available.
aspects. First of all, the computation time left for
One example of such an implementation is made in
monitoring functions is often small because of all time-
the paper mill AssiDomän Frövi in Frövi, Sweden, see
critical control functions that run on the system.
Ericsson, Hägglund, and Rudfjäll (2002). The oscillation
Another limitation is the programming environment.
detection procedure is implemented in a High-Perfor-
Implementation of monitoring functions in the DCS
mance Process Manager (HPM) in the Honeywell
system is normally restricted to a rather primitive block-
TDC3000 system. The detection procedure is currently
programming language.
running on-line and supervises over 91% of the control
A more convenient solution is to implement the
loops, which corresponds to about 1200 loops, 24 hours
monitoring functions in a separate computer system that
a day in the carton board mill.
is connected to the DCS systems through OPC servers.
Since so many loops are supervised, all detections
This requires that the DCS systems have this facility,
must be handled and presented in a suitable way. Each
and that they are able to deliver the data at a rate that is
loop is characterized by its ‘‘oscillation index’’, which is
high enough. Since a typical process plant often has
an index that is increased every time an oscillation is
several DCS systems, another advantage with this
detected. The code presented in Section 2 is comple-
strategy is that it provides a possibility to handle and
mented with the following calculations:
present information from all loops in the plant in a
unified framework. Furthermore, the programming OSCILLATION INDEX CALCULATION
environment is often more convenient in these external if oscillation = true then
systems than in the DCS systems. begin
oscindex = oscindex + Ti/6;;
end;
4. Implementations
The reason why the oscillation index is increased
The oscillation detection procedure was first imple- by the integral time (in minutes) is that it makes it
mented in the single-station controllers ECA400 and possible to compare oscillation indices of different
ECA600 from ABB. The method is fully automatic in loops with different time constants. A rapidly osc-
the sense that all parameters are fixed to the values illating loop has a short integral time and a high
suggested in Section 2. When an oscillation is detected, frequency of oscillation detections, whereas a slow loop
the front starts to flash and continues to flash until the has a long integral time and a low rate of oscillation
alarm is acknowledged. The detection procedure also detections.
has an additional function. If the ECA600 controller is The oscillation index is reset to zero when the
running in adaptive mode, the adaptation is switched off controller is tuned or when a maintenance is performed.
as soon as an oscillation is detected. This is to avoid In this way, it is easy to get a quick historical overview
detuning the controller in case of stick-slip motion of the different loops by inspecting the history of their
caused by friction in the valve. An application of the use oscillation index.
of the monitoring procedure in the ECA600 controller is The most oscillatory loops are presented on a
described in Section 5. ‘‘top-ten list’’ in the mill-wide information system
Both the oscillation detection procedure and the Idle for proper action by the instrument department.
index calculation are implemented in the ABB In- Two examples from the implementation are given in
dustrialIT system, but this system is recently released and Section 5.
 ARTICLE IN PRESS
1388 T. Hägglund / Control Engineering Practice 13 (2005) 1383–1390

5. Examples successive zero crossings of the control error, as well as

IAElim : Since the ultimate period was not available,
The oscillation detection procedure, implemented in IAElim was calculated from the integral time T i as
the ECA-600 controller has been applied on various IAElim ¼ 2=oi ¼ T i =p 7:6: The IAE values are sig-
industrial plants with good results. The following nificantly larger than IAElim ; as can be expected because
example is taken from a pulp concentration control of the high oscillation amplitude.
section in a paper mill, where pulp is diluted with water Finally, the fourth graph shows the rate of load
to a desired concentration. The concentration controller detections x and the rate limit nlim ; both calculated
was a PI controller with gain K ¼ 0:33 and integral time according to the program in Section 2. The rate x
T i ¼ 24 s: exceeds the rate limit nlim after about 3 min, and the
Fig. 3 shows 10 min of data from the loop. The first detection procedure gives an alarm. The rate x
graph shows the process output, i.e., the pulp concen- converges to about 80, eight times larger than the rate
tration, in %. Because of high friction in the water valve, limit. However, in the implemented version x is
the process oscillates with an amplitude of a few percent. reinitialized to zero every time x exceeds nlim :
The first graph also shows an estimate of the set point, The oscillations are easily noticed in Fig. 3. However,
since this variable was not recorded. The estimate is process operators seldom have access to these kind of
simply obtained by a low-pass filtering of the process graphs, but are often left with a bar graph with a low
output. resolution, or even just a numerical display. The present
The second graph shows the control signal in %. The oscillation had been present for a long time without
third graph shows the IAE value calculated between being discovered by the process operators.
The last two examples are taken from the implemen-
tation in Frövi, see Ericsson et al. (2002). The first is a
Process output y and estimated set point ysp flow control loop, where the flow of retention chemicals
50 added directly to the short circulation of the top layer of
the board machine is controlled. The chemicals control
45
the retention, which in turn affects, e.g., the basis
40 weight. It is, of course, important to keep the flow at a
desired level.
35 Data from the flow control loop are presented in
0 200 400 600
Fig. 4. The loop was oscillating, and in this case
Control signal u the oscillations were detected almost immediately, and
12 the control loop was switched into manual mode. The
maintenance staff investigated the oscillating loop and
10 found, by performing repeated small step changes in the
control signal, that the pilot valve in the actuator
8

0 200 400 600 l/s %

IAE and IAElim 24 Maintenance
23
Flow [l/s]
22
40 21 Automatic mode
20
20 19
Control signal [%]
18
0
0 200 400 600 800 1000 1200 1400
0 200 400 600
Rate of load detections x and rate limit nlim 5 Oscillation index
4
20 3 Manual mode
2
10 1
0
0 0 200 400 600 800 1000 1200 1400
0 200 400 600
Fig. 4. Stick-slip motion in a flow control loop. The upper diagram
Fig. 3. Oscillation-detection procedure applied to a pulp concentra- shows the process output (solid) and the control signal (dashed). The
tion control loop. lower diagram shows the oscillation index.
 ARTICLE IN PRESS
T. Hägglund / Control Engineering Practice 13 (2005) 1383–1390 1389

25 procedures are probably general and valid for many

20
other monitoring procedures.
The monitoring tools cannot be implemented in
y [%]

15 systems that only have access to the signals of the

10 control loop. There is so much more information needed
about the control loops in addition to the values of the
5
0 100 200 300 400 500 600 700 800 signals that are monitored. The operating condition,
controller parameters, and values of additional signals
8.5
are examples of such information. This means that the
8
monitoring tools should be implemented in the DCS
7.5
systems or in systems that are connected to the DCS
u [%]

7
6.5
systems through OPC servers.
6
Some examples of successful applications of perfor-
5.5 mance monitoring tools have also been given in the
0 100 200 300 400 500 600 700 800 paper. The use of this technique is increasing rapidly,
and several papers have verified that this is a good way
Fig. 5. Stick-slip motion in a differential pressure loop. The upper
diagram shows the measurement signal (y) and the lower diagram to increase productivity and reduce down time in
shows the control signal (u). process control plants.

suffered from stick-slip motion. The action taken was to

clean the actuator and switch the mode to auto again.
References
As shown in Fig. 4 the oscillation index was
increasing due to repeated detections. The increase Åström, K. J., & Hägglund, T. (1995). PID Controllers: Theory,
stopped when the mode was switched to manual. After Design, and Tuning. Research Triangle Park, North Carolina:
the maintenance was undertaken the index was reset to Instrument Society of America.
zero. The oscillations and the index increase stopped Bialkowski, W. L. (1993). Dreams versus reality: a view from both
after this maintenance. sides of the gap. Pulp and Paper Canada, 94(11).
Desborough, L., & Harris, T. (1992). Performance assessment
The second example shows control of a differential measures for univariate feedback control. Canadian Journal of
pressure of a steam cylinder in the drying section of the Chemical Engineering, 70, 1186–1197.
board machine. Disturbances in the differential pressure Dumont, G.A., Kammer, L., Allison, B.J., Ettaleb, L., & Roche, A.A.
affect both the steam pressure and the flow of condensed (2002). Control system performance monitoring: new developments
water in the entire steam group. and practical issues. In IFAC World Congress. Barcelona, Spain.
Ender, D. B. (1993). Process control performance: not as good as you
Data from the loop are presented in Fig. 5. This think. Control Engineering, 40(10), 180–190.
control loop was oscillating, which was discovered from Ericsson, S., Hägglund, T., & Rudfjäll, P. (2002). Implementation of
inspection of the ‘‘top-ten list’’. The loop was investi- an on-line detection procedure for oscillating control loops. In
gated, and in this case the actuator was also suffering Control Systems 2002. Stockholm, Sweden.
from stick-slip motion. Fig. 5 shows repeated manual Hägglund, T. (1995). A control-loop performance monitor. Control
Engineering Practice, 3, 1543–1551.
step changes of the control signal without any response Hägglund, T. (1999). Automatic detection of sluggish control loops.
in the process value. In this case no maintenance could Control Engineering Practice, 7, 1505–1511.
be made without stopping the entire board machine and Hägglund, T., & Åström, K. J. (2000). Supervision of adaptive control
the controller was left in manual mode until the next algorithms. Automatica, 36, 1171–1180.
Horch, A., & Isaksson, A. (1999). A modified index for control
stop of production.
performance assessment. Journal of Process Control, 9, 475–483.
Lynch, C. B., & Dumont, G. A. (1996). Control loop performance
monitoring. IEEE Transaction on Control System and Technology,
6. Conclusions 4, 185–192.
Owen, J.G., Read, D., Blekkenhorst, H., & Roche, A.A. (1996). A mill
Several procedures for automatic performance mon- prototype for automatic monitoring of control loop performance.
In Proceedings on Control System ‘96. Halifax, Nova Scotia,
itoring have been developed in the last decade. These Canada.
procedures are now implemented and applied in Paulonis, M. A., & Cox, J. W. (2003). A practical approach for large-
industry. This paper has focused on various aspects of scale controller performance assessment, diagnosis, and improve-
these implementations. ment. Journal of Process Control, 13, 155–168.
Two performance monitoring tools have been dis- Perrier, M., & Roche, A.A. (1992). Towards mill-wide evaluation of
control loop performance. In Proceedings on Control System ‘92.
cussed in this paper, one for detecting oscillating loops Whistler, BC.
and one that detects sluggish loops. The conclusions Qin, S. (1998). Control performance monitoring—a review and
drawn about the implementation aspects of these assessment. Computers and Chemical Engineering, 23, 173–186.
 ARTICLE IN PRESS
1390 T. Hägglund / Control Engineering Practice 13 (2005) 1383–1390

Stanfelj, N., Marlin, T., & MacGregor, J. (1993). Monitoring and Thornhill, N. F., Oettinger, M., & Fedenczuk, P. (1999). Refinery-wide
diagnosing process control performance: The single-loop case. control loop performance assessment. Journal of Process Control,
Industrial and Engineering Chemistry Research, 32, 301–314. 9, 109–124.
Thornhill, N. F., & Hägglund, T. (1997). Detection and diagnosis of Tyler, M. L., & Morari, M. (1996). Performance monitoring of
oscillation in control loops. Control Engineering Practice, 5, control systems using likelihood methods. Automatica, 32,
1343–1354. 1145–1162.