Anda di halaman 1dari 24

INTERNATIONAL JOURNAL OF ADAPTIVE CONTROL AND SIGNAL PROCESSING

Int. J. Adapt. Control Signal Process. 2003; 17: 685708 (DOI: 10.1002/acs.772)

A relative performance monitor for process controllers


Q. Li, J.R. Whiteley and R.R. Rhinehartn
School of Chemical Engineering, Oklahoma State University, 423 Engineering North, Stillwater, OK 74078, USA

SUMMARY
A monitor is developed to automatically detect poor control performance. It provides a measure (relative
performance index}RPI) of a control-loop performance relative to a reference model of acceptable
control. The reference model simulates the controlled variable output of a user-dened, acceptably tuned
control loop. The inputs to the reference model are the setpoints (same as the true plant) and the
disturbances (estimated from the measurements). The monitor uses routine plant operation data only.
Pending ability to obtain temporally accurate process models, and the validity of process measurements,
simulations and experiments show that the monitor can detect the poor control performance caused by
improper controller parameter values or changes in plant characteristics, and can distinguish it from poor
performance caused by external disturbances. Copyright # 2003 John Wiley & Sons, Ltd.
KEY WORDS:

control-loop performance; controller performance; process monitoring; performance


assessment

1. INTRODUCTION
The performance of a process controller often changes during plant operation. An initially welltuned controller may become undesirably sluggish or aggressive due to many reasons, such as
changes in process gain, process dynamics, valve stiction or constraints. A controller with poor
performance increases manufacturing costs, lowers product quality and even risks process
safety. Therefore, monitoring controller performance is important, and has become a routine
task for process control engineers. However, detecting a poorly performing controller requires
expertise and experience, and is very time consuming. In practice, many poorly performing
controllers often exist in plants unnoticed for a quite long time before being detected.
Therefore, it would be nice to have an automatic monitoring tool to indicate when a control
loop has signicant changes in its performance relative to the performance desired by operators.
The monitor should not disturb routine plant operation, and it should use only the routine plant
operation data. The monitor should detect a poorly performing controller, and suggest when

Correspondence to: R. Russell Rhinehart, School of Chemical Engineering, Oklahoma State University, 423
Engineering North, Stillwater, OK 74078, USA.

Contract/grant sponsor: Measurement and Control Engineering Center (MCEC)

Copyright # 2003 John Wiley & Sons, Ltd.

Accepted June 2003

686

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

maintenance is needed for the controller (checking for valve stiction, adjusting process model
and retuning the controller).
Currently, most cited data-based controller performance monitoring techniques are based on
the work of Harris [1], in which the controlled variable variance under minimum variance
control (MVC) is used as a lower bound benchmark to evaluate the performance of single-loop
controllers. The ratio of the minimum variance to the variance of the controlled variable is
dened as normalized performance index [2]. Other similar measures have also been proposed,
such as the closed-loop potential index [3], the relative variance index that compares actual
control to both MVC and open-loop control [4], and a modied performance index based on the
desired pole locations and MVC [5, 6]. Reviews on the MVC-based techniques can be found in
References [79]. Many methods are also proposed to detect sluggish control [10] or oscillations
[1115]. Applications of the performance assessment schemes in the process industries can be
found in References [1521]. The minimum variance benchmark is widely used in industry for
performance assessment because it can be used to determine the improvement potential in
variance reduction by only requiring an estimate of process delay and routine closed-loop
operation data.
Although the MVC benchmark provides a theoretical lower bound on controlled variable
variances, it is not a practical benchmark that every good controller should try to achieve. It is
because (1) the minimum variance usually cannot be achieved unless the process model and
disturbance model are perfectly known, which is practically impossible and (2) operating too
close to the minimum variance often means excessively large moves of controlled variables,
which is not acceptable in practice. As a consequence, a well-tuned controller in practice
operates with some distance from the minimum variance point. A user has to decide the optimal
distance for good control, which is unique to the balance of issues for each loop. One cannot say
that the closer the actual variance is to the minimum variance, the better the controller
performs.
There are other passive, data-based monitoring techniques. An automated on-line goodness
of control performance monitor was proposed by Rhinehart [2224]. The method uses a
computationally simple, robust statistic, called the r-statistic, which is dened as the ratio of the
expected variance of the deviation of the controlled variable from the setpoint to the expected
variance based on the deviation between two consecutive process measurements. Diculties in
the r-statistic method are either how to choose the right range of r-values for acceptable and
unacceptable performance or how to select a sampling rate to eliminate auto-correlation.
An online automated control performance monitor based on statistical dierences in runlength (RL) distributions was proposed [25]. The RL index is dened as the time period (number
of sampling periods) between two consecutive zero-crossings (sign changes) of the controller
actuating error signals (setpoint minus controlled variable). The histogram of RL index under
dierent control performance, such as sluggish control, aggressive control (oscillations) and
good control, are signicantly dierent and therefore are used to detect poor control
performance. The monitor does not require either process knowledge or process model, and it
uses only routine plant operation data. One diculty in RL-distribution method is how to
choose a representative data to build the reference RL-distribution, which represents acceptable
control performance.
Most control performance monitoring techniques do not dierentiate poor control
performance caused by external disturbances from that caused by the control loop itself
(improper controller tuning parameters, constraint or control valve stiction). Obviously, it is
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

687

A RELATIVE PERFORMANCE MONITOR

important to dierentiate these two cases. If the controller or control valve is the reason of the
poor control, maintenance is called for, but if the external disturbance is the cause, loop
maintenance should not be triggered. This work develops a technique that is insensitive to
disturbance behaviour, and identies problems within the control loop (e.g. tuning, constraints,
valve stiction).
The idea of using a reference model has been used in model reference adaptive control [26], in
which the controller parameters are adjusted online to minimize some function of the dierences
(errors) between the measured controlled variable output and the output of the reference model,
which represents the desired closed-loop response. This is extended, here, for control
performance monitoring.
The concept is to have a reference model to simulate how a chosen good control system would
respond to the same setpoint sequence and disturbance sequence imposed on an actual plant,
and to compare the actual control performance relative to the performance dened by the
model. In this work, a relative performance monitor is proposed to measure the current controlloop performance relative to that of a reference model. The reference model represents the
behaviour of an adequately tuned control loop under the same (setpoint and disturbance) inputs
as the actual plant experiences.

2. A RELATIVE PERFORMANCE MONITOR


The basic idea of the proposed relative performance monitor (RPM) is to compare the
performance of a control loop under monitoring to that of a reference model to measure
the relative performance of the actual control loop to the good performance represented by the
reference model. Since the controller actuating error (i.e. setpoint controlled variable) is a
good indication of control-loop performance, we compare the actuating error of a control loop
to that of a reference model to determine the relative performance of the control loop and the
reference model. Figure 1 shows this basic idea. System A in Figure 1 represents a control loop
under monitoring. The inputs to system A are the setpoint SP and disturbances d and the output
is the actuating error eA : Using eA as plant output is equivalent to using the controlled variable
CVA as the plant output because eA SP  CVA and the SP is known. The same inputs, SP and
d; are fed into the reference model R; and the output of the reference model is eR eR
SP  CVR : A comparison of the error sequences of eA and eR through a proposed relative
performance index gives the relative performance of a control loop and the reference model.
A simplied version of a typical feedback control loop consists of a controller and a process,
as shown in Figure 2. In Figure 2, we arrange the block diagram such that SP and d are the
SP
d

System A
(controller &
process)

eA

RPI

RPI

Reference
Model, R
eR

Figure 1. Basic idea of the relative performance monitor.


Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

688

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

System A
SP
d
+

MVA
CA

CVA

eA

PA
+

Figure 2. Block diagram for a typical feedback control system (system A).

inputs and the actuating error e is the output. PA represents the actual plant A to be controlled,
CA the controller, CVA the controlled variable, SP the setpoint, MVA the manipulated variable,
and d the eects of disturbances and noise, which are modelled as additive to the controlled
variable CVA :
For the control loop in Figure 2, we have the following Laplace transform relation:
PA sCA s
1
SPs
ds
1
CVA s
1 PA sCA s
1 PA sCA s
Since the actuating error eA SP  CVA ; we have


PA sCA s
1
eA s 1 
ds
SPs 
1 PA sCA s
1 PA sCA s
eA s

1
SPs  ds
1 PA sCA s

The term PA sCA S=1 PA sCA s represents the closed-loop response of system A to the
setpoint input. If we use ASP s to represent the closed-loop response of a control loop to the
setpoint input, we have,
CVA s
PA sCA s

3
ASP s
SPs
1 PA sCA s
or
1  ASP s

1
1 PA sCA s

So, the actuating error of a control loop A can be calculated as


eA s 1  ASP sSPs  ds

In a similar way, by choosing a desired closed-loop setpoint response relation, RSP s, and
using RSP s as the reference model of good control, we can calculate the actuating error output
of the reference model as
eR s 1  RSP sSPs  ds

The reference model R in Figure 1 represents the desired good relationship between the inputs
(the setpoint and disturbance) and the output (the actuating error or the controlled variable).
Therefore, the CVR output SP2eR of the reference model simulates the good closed-loop
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

689

behaviour under the same input (setpoint and disturbance) conditions as experienced by system
A; the actual control loop. So, a comparison of system A output, eA or CVA ; and the reference
model output, eR or CVR ; gives the performance of a control loop A relative to that of the
reference model.
A relative performance index (RPI, dened later) is calculated from the actuating errors of a
control loop and the reference model. The RPI provides a measure of the relative performance
of the control loop and the reference model.

3. REFERENCE MODELS
There are many ways to dene a reference model. The reference model can be as simple as a step
response function, or as complex as a simulated adaptive control system.
Here we discuss two simple forms of the reference model: (1) the parametric model form, such
as a rst-order or second-order response model, and (2) the non-parametric model form, such as
a nite impulse response (FIR) model, or a step response model.
A parametric model has a small number of parameters that need to be specied. For example,
a rst-order parametric response model has the following form:
RSP s

1
tR s 1

which has only one parameter, the time constant, tR ; since the gain must be equal to 1.0 to
remove oset. A second-order parametric model has the following form:
RSP s

1
t2 s2 2zts 1

which has two parameters that need to be specied, the damping factor z and the time constant t
(again, the gain is set to 1.0 to remove oset). When z > 1 (overdamped case) or z 1 (critically
damped case), there is no overshoot in model response to a step setpoint change. When z51
(underdamped case) for a step change, we have analytical solutions for the following
performance indices:
!
pz
Overshoot : OS exp p
9
1  z2
Decay ratio :

!
2pz
DR OS2 exp p
1  z2

Period :

2pz
P p
1  z2

10

11

By choosing z and t we can determine the desired response to a setpoint change.


The rst-order and the second-order models can represent most of the desired response types,
and they require only one or two parameters to be specied, so it is very easy to specify a
parametric reference model. Further, no plant tests are required to obtain a parametric reference
model, so this method is useful when no plant step tests are allowed.
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

690

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

The other simple form of the reference model is the FIR model, which is usually obtained
through plant tests after the control loop is well tuned by whichever criteria the user wishes.
After the controller tuning, introduce a step change DSP in setpoint, and record the CV response
samples, s0 ; s1 ; s2 ; . . . ; sn ; until the CV reaches the new steady state (assuming at sample n). Then,
the reference model RSP ; in discrete time, can be represented as
RSP q rSP0 rSP1 q1 rSP2 q2    rSPn qn

12

where q is a forward shift operator, and q1 is the backward shift operator, such that for a
sample xk at time k; qxk xk 1 and q1 xk xk21; the model coecients rSPi ;
i 0; 1; 2; . . . ; n; are determined by,
si  si1
rSPi
; i 1; 2; . . . ; n and hSP0 0
13
DSP
and n is the number of sampling periods that it takes the CV to reach a new steady state after a
setpoint change.
In discrete time, the actuating error sequence eR k can be derived in a similar way as deriving
Equation (6), so we have
eR k 1  RSP qSPk  dk

14

We can prove that 12RSP q is actually a control loops CV response to the disturbance
input. For a control loop as shown in Figure 2, we have
ASP q

CVA k
PA qCA q

SPk
1 PA qCA q

15

CVA k
1
PA qCA q

1
dk
1 PA qCA q
1 PA qCA q
Therefore,
CVA k
1  ASP q
dk

16

This also indicates that once the CV response of a feedback control loop to the setpoint input is
xed, its response to the disturbance input is also xed.
Let Rd q denote the CV response of the reference model R to the disturbance input, we have,
Rd q rd0 rd1 q1    rdn qn 1  RSP q

17

Since
1  RSP q 1  rSP0  rSP1 q1      rSPn qn
and rSP0 0; we have,
rd0 1
rdi rSPi
Copyright # 2003 John Wiley & Sons, Ltd.

si1  si
;
DSP

i 1; 2; . . . ; n

18

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

691

So, the time-domain-equivalent way to calculate the actuating errors as in Equation (14) is
n
X
eR k
rdi SPk  i  dk  i
19
i0

where rdi are dened in Equation (18).


The major advantage of using an FIR reference model is that an FIR model can represent the
closed-loop responses of any linear control system with any order or complex dynamics.
4. RELATIVE PERFORMANCE INDEX
The performance of a control loop could be measured by many metrics, such as mean-squared
error (MSE), mean absolute error (MAE), variance, minimum variance-based indices, r-statistic
or the index based on RL distribution dierences.
The exponentially weighted-moving-average of squared error (EWMASE) or absolute error
(EWMAAE) could also be used if we want to put more weighting on recent data than old data
or we do not want to take too much computer resources (memory or CPU time) for processing
the data. The EWMASE metric M1 is calculated recursively as follows:
M1 k l * M1 k  1 1  l * e2 k

20

where M1 k and M1 k  1 are the EWMASE metric values calculated at time k and k21; l is a
constant between 0 and 1.0, and the pth sample of e2 in the past carries a weight of 12llp ; e is
the dierence between the setpoint and the controlled variable, i.e.
eA SP  CVA

21

eR SP  CVR

22

where subscript A represents the actual control loop, and subscript R represents the reference
model.
From eA ; we can calculate a performance metric MeA ; such as MSE or EWMASE, for the
control loop under monitoring. And similarly, from eR ; we can also calculate the same
performance metric MeR for the reference model.
We dene an RPI based on a chosen performance metric M: The RPI based on metric M is
dened as the ratio of the metric value for the reference model MeR ; and the metric value for
the control loop MeA ;
MeR
23
RPIM
MeA
If we choose M1 (i.e. EWMASE) as the performance metric, the RPI based on M1 is
M1R
RPIM1
M1A

24

where
M1R k l * M1R k  1 1  l * e2R k

25

M1A k l * M1A k  1 1  l * e2A k

26

The RPI based on other metrics (such as MSE) could also be dened in a similar way.
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

692

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

The RPI provides a measure of the relative performance of the control loop and the reference
model, and (1RPI) represents the improvement potential in control-loop performance if
retuning the loop to the reference model level of performance. The RPI value can be interpreted
as follows: (1) An RPI value close to 1.0 means the control-loop performance is close to that
of the reference model; (2) an RPI value 51.0 indicates the loop performance is much worse
than that of the reference model, and something should be done on the control loop, such as,
re-estimating plant model, retuning controller parameters, checking for valve stiction, etc. In
this case, the chosen performance metric (such as EWMASE or MSE) can be reduced by
(1RPI)n100% under similar input conditions if we retune the control loop to reach the same
performance level represented by the reference model. (3) An RPI value >1:0 means the control
loop has a better performance than the reference model. If the RPI values are much greater than
1.0, we may need to update the reference model.
If we want the monitor to automatically ag poor control, we can specify a critical or
threshold value PRIc, such that when the RPI value exceeds the threshold value, the monitor
automatically ags. A good choice of the threshold value depends not only on the noise level of
the data but also on the users tolerance level on the deviation of a control-loop performance
from that of the reference model. A user can make a choice of the threshold value based on the
interpretation that the variance or MSE or EWMASE can be improved by (1RPI)n100% if
retuning the control loop to the performance level represented by the reference model. For
example, RPI=0.75 means that the variance (or similar measures) can be reduced by 25% if the
controller is retuned to reach the performance level of the reference model. RPIc=0.75 could be
used as an initial threshold value to ag the performance monitor if 25% improvement is worth
the controller-retuning eorts.
Statistical tests could also be used to establish the threshold values for controller monitor
agging. The authors have used the F -test to test the hypothesis of equal variances
of actual output and reference model output. The critical value to reject the hypothesis was
used as the threshold for monitor agging. Although statistics-based approaches provide
powerful methods to determine the threshold values, it adds complexity, and will not be
demonstrated here.

5. DISTURBANCE ESTIMATION
The inputs to the reference model, as well as the actual closed control loop, are the setpoint and
disturbances. We know exactly the setpoint input sequence to the actual control system, but we
do not know the disturbance (including noise) sequence, and therefore we have to estimate it.
The disturbance estimation with adequate accuracy is the key to the monitors ability to
distinguish the poor control performance caused by external disturbances and other poor
performance caused by problems within the control loop, such as plant characteristic changes,
poor controller tuning parameters, valve stiction or hitting constraints.
We choose to use the prediction error sequence of the plant model as estimates of the eective
disturbance and noise, which are added to the true CV. The model prediction error is the
dierence between the actual measured CV value and the model predicted CV value.
Let the plant model be f k; uk; y; i.e.
ymod k f k; uk; y
Copyright # 2003 John Wiley & Sons, Ltd.

27

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

693

where ymod k is the plant model output at discrete time k; uk the plant input at time k; and y
the set of model parameters. Then the model prediction error sequence is just yk  ymod k:
Therefore, the disturbance sequence dk can be estimated by
dk yk  f k; uk; y

28

where f k; uk; y can be a linear or non-linear plant model.


If we know the plant characteristics do not change much during plant operation, we can use a
xed plant model f k; uk; y to estimate the disturbances. If no plant models are available, we
can obtain an estimated plant model through a step change in plant input while the controller is
oine. The open-loop response data can be used to build an FIR plant model as described
previously in Equations (12) and (13).
If the plant model parameters change signicantly during plant operation and the xed plant
model cannot reect the changes, it may be necessary to adaptively estimate and update the
model parameters. If this is the case, we assume a plant model structure, such as ARX,
ARMAX, BoxJenkins, statespace or a non-linear neural networks model and estimate the
model parameters adaptively as the plant changes during operation. For the purpose of
disturbance estimation, we may choose a simple plant model structure and estimate its
parameters under closed-loop condition.
Here, we choose a simple rst-order ARX plant model structure
yk

b
uk  kd dk
1 aq1

29

where yk; uk and dk are the plant output CV, input MV and eective disturbances (plus
noise), respectively, at discrete time k; kd is the time delay, a and b are constants, and q1 is the
backward shift operator, such that q1 uk uk  1: The rst-order ARX, which is the
discrete time version of the popular rst-order-plus-time-delay (FOPTD) model, is chosen here
because of its simplicity in parameter estimation and its adequacy for our disturbance
estimation for control performance monitoring purpose. There are many system identication
methods to estimate a system order and parameter values [27].
After estimating the plant parameters, a; b and kd ; the eective disturbance can be estimated
as
dk yk 

b
uk  kd
1 aq1

30

We choose the recursive least-square (RLS) estimation method because it works well with
time variant or non-linear processes, which are common in the process industry.
The RLS algorithm has the following form [27]:
y# k y# k  1 Kkyk  y#k

31

where y# k is a vector of the model parameters estimated at discrete time k, yk is the observed
process output at time k, y#k is a prediction of yk based on observations up to time k  1 and
the estimated model at time k  1:
The gain Kk in Equation (31) has the following form:
Kk Qkck
Copyright # 2003 John Wiley & Sons, Ltd.

32

Int. J. Adapt. Control Signal Process. 2003; 17:685708

694

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

where ck is the gradient of y#kjy with respect to y; y#kjy is the prediction of yk based on the
model described by parameters y; and Qk is a matrix that controls the adaptation gain and
direction.
For linear regression model structures, such as AR and ARX, the process output prediction
y#kjy can be written as
y#kjy jT kyk  1
where jk is the regression vector, which consists of past values of observed inputs and outputs.
Therefore,
d
ck
y#kjy jk
33
dy
The matrix Qk can be determined by minimizing the following cost function:
k
X

lkj yj  y#j2

34

j1

where the forgetting factor l is a constant between 0 and 1, and typically between 0.970.995
[27]. The forgetting factor is used to discount old measurements exponentially so that an older
observation will carry a less weight in the cost function than more recent data. The squared
error that is p samples away in the past from the current time k carries a weight of lp in the cost
function in Equation (34), which is to be minimized to obtain the model parameters.
For linear regression models, such as ARX, the cost function Equation (34) can be minimized
exactly with the following choice of Qk [27]:
P k  1
Qk P k
35
l jkT P k  1jk
P k

1
P k  1jT P k  1
P k  1 

l
l jkT P k  1jk

36

For the rst-order ARX process model shown in Equation (29), the model parameter vector is
y a bT

37

jk yk  1uk  kd T

38

The regression vector is


Note that the time delay kd in terms of number of sampling periods must be specied before
we can construct the regression vector y: If we know that the range of change of the process
delay during operation is smaller than one sampling period, we can use a xed process delay
during monitoring. Otherwise, we need to treat the delay as a variable and estimate it. One way
to estimate the delay is to assume a possible range of change of the process delay during
operation in terms of number of sampling periods, and for each time delay value, construct the
corresponding regression vector and estimate the corresponding model parameters. The delay
with the smallest prediction error is chosen as the estimated delay.
As the new input output data arrive, we can recursively estimate the plant model parameters
(kd ; a and b), which can adapt themselves as the plant characteristics change during plant
operation. The closer the estimated model is to the true plant, the more accurate are the
disturbances estimated.
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

695

6. APPLICATION PROCEDURE OF THE RELATIVE PERFORMANCE MONITOR


(1) Specify a reference model RSP, either through a plant step test or by choosing parameters
for a parametric model.
(a) If plant step tests are allowed, make a step change in SP after tuning the controller at
the plant nominal operating point. Observe the closed-loop CV response. Calculate
the reference model parameters rdi from the CV response data using Equation (18).
(b) If plant step tests are not an option, assume the desired response to setpoint input is
rst-order or second-order, and choose the parameter td as in Equation (7) or the
parameters t and z as in Equation (8).
(2) Choose an approach to estimate the disturbance.
(a) One approach is to use a xed and predetermined plant model. If no predetermined
model is available, it can be identied by a step test. Make a step change in plant
input at the nominal operating point while the controller is oine, record the plant
output response, and calculate a step response-type model parameter values in the
same way as shown in Equation (13). After a plant model is available, estimate the
eective disturbance sequence using Equation (30).
(b) Another approach to estimate the disturbance is to use an adaptable plant model
whose parameters are recursively estimated online. Then, choose a forgetting factor
l; usually between 0.90 and 0.995, according to the desired emphasis of recent
measurements relative to past ones. After estimating the plant model parameters
using Equations (31)(36), estimate the eective disturbances at each sample time
using Equation (30).
(3) Feed the actual, known setpoint sequence and the estimated eective disturbance
sequence into the reference model to obtain the actuating error output of the reference
model, either using Equation (19) for the non-parametric reference model case (with the
plant step test data) or using Equation (6) for the parametric reference model case (with
chosen model parameters).
(4) Calculate the RPI values using Equations (24)(26), for example.
(5) Choose a threshold value RPIc (for example, RPIc=0.75) for agging poor control
performance based on the users tolerance level on the performance deviation of
the control loop from the reference model. If an RPI value is greater than RPIc, the
relative performance monitor ags poor control performance and suggests loop
maintenance.

7. EVALUATION AND DISCUSSION OF THE RELATIVE PERFORMANCE


MONITOR
The relative performance monitor is demonstrated through computer simulations and
experiments on a water ow control loop.
7.1. Simulation results and discussion
We simulate an SISO rst-order-plus-time-delay (FOPTD) plant controlled by a PI controller
using Matlab/Simulink. In all simulations, the plant and the PI controller are simulated using
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

696

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

continuous-time model, which can be handled by Matlab built-in functions (with variable step
size). The relative performance monitor takes data samples and performs computations at every
discrete time unit.
The plant is
Tp

dy
yt Kp ut  Td
dt

39

where Kp is the process gain, Tp is the process time constant and Td is the time delay. Choose
Kp=1, Tp=10 and Td=1, and then the plant becomes
10

dy
yt ut  1
dt

40

The disturbance, which is added to the plant output, has at least two additive components: (1)
measurement noise represented by a zero-mean Gaussian noise with variance equal to 1, and (2)
a disturbance process driven by a zero-mean Gaussian noise wt:
15

dy
yt 2wt
dt

41

After tuning, the PI controller has the following parameters: controller gain Kc=5
(dimensionless), and integral time constant Tc=10 (time units), determined from the ITAE
tuning method for setpoint changes and the plant model. Any controller tuning method could
be used to tune the controller to the users satisfaction.
Make a step change in SP, record the CV response and calculate the reference model RSP for
the setpoint input. Figure 3 shows the plant closed-loop response to the setpoint change, and the
calculated reference model parameters rSPi and rdi :
The solid line in the third plot of Figure 3 is the response to the same step setpoint change of a
second-order parametric reference model below:
RSP s

1
1:12 s2 2 * 0:7 * 1:1s 1

42

with a damping factor z 0:7 and a time constant t 1:1: We can see that, in this case, the
response of the second-order reference model Equation (8) with appropriate parameters z and t
is close to the actual control-loop response. In many cases, we can nd a parametric reference
model such as Equation (8) or (7) to represent the desired reference model if no plant tests are
allowed.
Here, we choose to use a xed and predetermined plant model to estimate the disturbance (the
rst approach). Since all models have errors, to be realistic, here, we assume the model
parameters are 20% dierent from the true values. In simulations, we use the following plant
model to estimate the disturbances:
12

dy
yt 1:2ut  1
dt

43

whose gain and time constant are both 20% larger than the true values of plant described by
Equation (40).
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

697

Figure 3. Closed-loop step responses to setpoint change, and the reference model parameters rSPi and rdi :
The solid line in the third plot is the response of a second-order parametric reference model to the setpoint
change. (Sampling period=1 time unit.)

7.2. Simulation of good control performance


First, run a simulation test with the acceptably tuned controller, and with the Gaussian noise
and rst-order process disturbance. The results after a warm-up period are shown in Figure 4.
The actual and estimated disturbances are shown in Figure 5.
Figure 4 shows the plant SP, plant output CVA, plant input MVA, reference model
output CVR and RPI index values. A setpoint change from 0 to 20 is introduced at sample
time 150. We can see that the RPI values are overall very close to 1.0 before and after the
setpoint change.
Figure 5 shows the actual disturbance sequence and the estimated disturbance sequence using
the pre-estimated plant model shown in Equation (43). We can see that although the estimated
plant model parameters are 20% larger than the true values, the estimated disturbance is very
close to the true value. Note here, the estimated disturbance is obtained from the dierence
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

698

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

Figure 4. Plant output CVA, plant input MVA, reference model output CVR and RPI when the control
loop is tuned. (Sampling period=1 time unit.)

between the actual plant output and the plant model output after removing the moving average
mean value from the dierence because we assume the disturbance is zero mean. Otherwise, the
estimated disturbances will have an oset from the true disturbances after setpoint changes, and
the oset is due to the plant model errors.
7.3. Simulation of poor control performance due to a too aggressive controller
To see how the monitor detects oscillations caused by a too aggressive controller, change the
controller gain Kc from 5 to 10 and the integral time Ti from 10 to 2, and run the simulator with
other conditions the same as before. The results are shown in Figures 6 and 7.
We can see that except for a short time immediately after the setpoint change, most of the
time the RPI values are smaller than 0.5, which means the squared errors (EWMASE or MSE)
could be reduced by at least 50% if the control loop is retuned to reach the same performance
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

699

Figure 5. The actual and estimated disturbances before and after a setpoint change at time 150 (a welltuned control loop). (Sampling period=1 time unit.)

level represented by the reference model. Immediately after the setpoint change, the RPI values
are increased for a short time because a very aggressive controller responds to a setpoint change
much quicker than the reference model, and the quick controller response reduces the errors due
to setpoint change, and improves performance. But, the overshoot due to the aggressive
controller causes the errors to increase, and the RPI values then drop to the previous level.
From Figure 7, we can see that the disturbance estimates are close to the actual values even
when there are oscillations due to an aggressive controller.
7.4. Simulation of poor control performance due to external oscillatory disturbances
To dierentiate the oscillations caused internally (by poor controller parameters) and externally
(by oscillatory disturbances), we run a simulation test with a sinusoidal signal below added to
the previous disturbances (Gaussian noise plus a rst-order process disturbance)
5 sin2pk=0:4
The amplitude is 5 and the frequency is 0.4 radian/sample, or a period of about 15 samples. The
results are shown in Figures 8 and 9.
From Figure 8, we can see that the RPI values are close to or above 1.0, which means the
relative performance monitor does not indicate poor control when the oscillations are only due
to external oscillatory disturbances. Even a well-tuned controller will oscillate under external
oscillatory disturbance, so does the reference model. Remember an RPI value greater than 1.0
indicates the actual performance is better than the reference model. The disturbances estimated
from a predetermined plant model Equation (43) with a 20% parameter error appear very close
to the true value of the oscillatory disturbances as shown in Figure 9.
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

700

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

Figure 6. Plant output CVA, plant input MVA, reference model output CVR and RPI when there are
oscillations due to a too aggressive controller (poor tuning parameters). (Sampling period=1 time unit.)

The ability of the relative performance monitor to dierentiate oscillations caused internally
and externally is very useful. We do not want to adjust a controller unless the control loop itself
is causing the poor performance. The monitor does not indicate oscillations caused by external
disturbance, and therefore makes it easy to identify the root cause of oscillations.
We can see that the accuracy of the disturbances estimated from a predetermined imperfect
plant model, in the simulations, is good enough for our control performance monitoring
purpose.

7.5. Simulation of control performance under drifting disturbances


To see the eect of the drifting (non-stationary) disturbances on the disturbance estimation and
the performance monitor output, an integral of the Gaussian noise (random walk) is added to
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

701

Figure 7. The actual and estimated disturbances when there are oscillations due to a too aggressive
controller. (Sampling period=1 time unit.)

the original disturbances (Gaussian noise plus a rst-order process disturbances). The actual
and estimated disturbances are shown in Figure 10.
From Figure 10, we can see that when the disturbances contain a slowly drifting component,
the mean value (more exactly, the zero or very low frequency components) of the actual
disturbances will drift around, and sometimes to a value far away from zero. It is very hard to
estimate this drifting component accurately because an imperfect plant model is used to estimate
the disturbances, and the model errors will accumulate. Note from Figure 10 that except the
drifting mean value (the very low frequency components), the estimated disturbances have a
pattern very similar to the actual disturbance.
Since most well-performing controllers have sucient integral action and thus can remove
slowly changing drifting (low frequency) disturbance fast enough, these slowly drifting
disturbances have almost no eect on the performance of this good controller. In other words,
the performance measures (such as the variances, MSE or EWMASE) of a well-performing
controller with reasonable integral action should be almost the same no matter whether there
exist the lowly drifting disturbances.
Therefore, although the estimated disturbances, which are fed into the reference model, may
have dierent slow-frequency components from the actual disturbances, these slowly changing
errors in the disturbance estimation will not have much eect on the output variances or MSE
of the reference model because the reference model almost always represents a well-performing
controller with sucient integral action. Figure 11 shows the results of the performance
monitor when there exist the drifting disturbances shown in Figure 10. We can see that the
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

702

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

Figure 8. Plant output CVA, plant input MVA, reference model output CVR and RPI when there are
oscillations due to external oscillatory disturbances. (Sampling period=1 time unit.)

performance monitor outputs with signicant drifting is close to the case without the drifting
shown in Figure 4.
7.6. Experimental results and discussion
Experiments on a water ow control loop are also used to evaluate the relative performance
monitor. In the water ow control loop, the controlled variable is the water ow rate, and the
manipulated variable is the signal to the valve. Control is executed by a Camile Tg 2000 system,
using 420 mA signals}to an i/p device operating a ow control valve in a 12 inch line, and from an
orice ow transducer. The inputoutput relation exhibits the rst-order plus time-delay dynamics
and non-linear characteristics. A PI controller is used to control the water ow rate at setpoint.
The nominal operating point of the water ow rate was 35 kg/h. After tuning the controller,
make a step change in the setpoint, and record the closed-loop response. Since the controller is
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

703

Figure 9. The actual and estimated disturbances when there are oscillations due to external oscillatory
disturbances. (Sampling period=1 time unit.)

Figure 10. The actual and estimated disturbances when drifting disturbances are added.
(Sampling period=1 time unit.)
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

704

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

Figure 11. Plant output CVA, plant input MVA, reference model output CVR and RPI when drifting
disturbances are added. (Sampling period=1 time unit.)

tuned acceptably well as judged by the operator, this response to a setpoint step becomes the
reference response. The experimental data, CVA, and the estimated reference model parameters,
rSPi ; are shown in Figure 12.
7.7. Experiments on poor control performance due to plant changes
To see if the monitor can indicate the poor control performance caused by changes in
plant characteristics, we operate the process at dierent ow rates. Since the process is nonlinear, a change in operating point (ow rate) means a change in plant characteristic, or
a change in the parameters in the FOPTD linear plant model, which we use to describe
the process.
To estimate the disturbance, we use the second approach, i.e. use an adaptable plant model
to estimate the disturbance. Since the plant characteristics experience signicant changes in
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

705

Figure 12. Experimental closed-loop step response data and the reference model RSP parameters rSpi :
(Sampling period=0.1 s.)

this experiment, an adaptable plant model is appropriate and a xed plant model will yield
large errors in the disturbance estimation. We choose the FOPTD plant model structure, with
a range of possible delays of 48 samples. The results from the second approach are shown
in Figure 13.
Starting from the nominal operating point, we make several step changes in ow rate setpoint,
and observe the changes in control performance due to plant changes while maintaining the
controller parameters unchanged. Figure 13 shows the experimental data CVA, MVA, the
reference model output CVR, the RPI values, and the estimated disturbances.
From Figure 13, we can see that when the plant operates near the nominal operating point
with a water ow rate of 35 kg/h, where the controller is tuned, the RPI values are close to 1.0,
even though there are setpoint changes to 30 kg/h, and then 25 kg/h. As the water ow rate
decreases and moves away from the nominal operating point, the plant characteristics change.
Step tests indicate that the steady-state gain is approximately 0.1 kg/h/% near the nominal
operating point 35 kg/h, but it becomes 0.5 kg/h/% near the operating point 20 kg/h, and 0.9 kg/
h/% near the operating point 5 kg/h.
As the plant changes, the control-loop performance gets worse, and the RPI values drop
far way from 1.0. We can see that the RPI values drops below 0.5 (indicating poor loop
performance) when the ow rate is below 20 kg/h, where the plant steady-state gain is at least
5 times that of the initial model obtained at the nominal operating point, and the poor control
performance with excessive oscillations occur. After the ow rate moves back to the region near
the nominal operating point, the control performance recovers and the RPI values are close to
1.0 again. Figure 13 also shows the estimated disturbances. We can see that, at steady state,
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

706

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

Figure 13. Water ow control experimental data CVA, MVA, reference model output CVR, RPI values,
and the estimated disturbance. (Sampling period=0.1 s.)

the disturbance estimates look very similar to the uctuations of the CV measurements, but when
there are excessive oscillations, the estimated disturbances have larger errors due to the plant
model errors.
7.8. Discussion summary
From the simulations and experiments, we can see that using a reference model to simulate a
well-tuned controllers response to setpoint and disturbance inputs has the following benets:
(1) it provides a practical, achievable standard for comparing control performances since the
user can specify the desired reference model or build the reference model by a plant step test just
after controller tuning, without the impractical shortcomings of the idealized minimum variance
standard; (2) it provides a relatively fair comparison because the reference model is subject to
exactly the same setpoint sequence and similar (depending on the estimate accuracy) disturbance
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

A RELATIVE PERFORMANCE MONITOR

707

sequence as the actual control loop, and the comparison standard is constantly adapting to the
plant input levels in setpoint and disturbances rather than comparing to a xed standard as in
many other methods; (3) the reference model directly simulates the controlled variable output
sequence, which is to be compared with the actual CV output sequence, so a variety of control
performance comparison methods can be used, such as, a simple visual comparison, variance
comparison or other comparisons in overshoot, settling time, mean absolute error, mean square
error or RL distributions.
Although not shown here, our tests reveal the relative performance monitor is able to indicate
poor control performance caused by too sluggish control or constraint-hitting conditions.
Further experiments are needed to claim the monitors ability to indicate poor performance
caused by valve stiction.
We suggest using a xed, reasonably valid plant model to estimate the disturbance whenever
the plant characteristics do not change much during operation, but when there are signicant
changes in plant characteristics during operation, an adaptable plant model should be used.
However, model identication or estimation under closed-loop condition is a very challenging
thing to do.
8. CONCLUSIONS
A relative control-loop performance monitor based on a reference model is developed, and
demonstrated with both simulations and experiments. The simulations demonstrate that the
monitor can distinguish poor control performance caused by external disturbances from those
caused by problems within the control loop, such as improper controller tuning parameters.
This ability is important for the control engineers to identify the root cause of poor control as
well as to decide which control loop needs maintenance. The experimental results demonstrate
that the monitor also can detect poor control performance due to changes in plant
characteristics.
The introduction of the reference model in the proposed relative performance monitor has the
following benets: (1) it provides a practical, achievable, exible comparison standard because
the user can specify the reference model, (2) it directly simulates a good control loops behaviour
under the actual setpoint input sequence (the same as the true plant) and estimated disturbance
input sequence, so the comparing standard is adaptive to the changes in the actual input setpoint
or disturbance, (3) the simulated plant output can be visually observed and compared directly
with the actual output using a variety of comparing methods and (4) the monitor uses routine
plant operation data only, and therefore does not disturb plant operations.
Ecacy of the approach requires a reasonably true process model, and is predicated on valid
measurements.
ACKNOWLEDGEMENTS

The authors appreciate both the nancial support and guidance from the industrial sponsors of the
Measurement and Control Engineering Center (MCEC).
REFERENCES
1. Harris TJ. Assessment of control loop performance. Canadian Journal of Chemical Engineering 1989; 67:856861.
2. Desborough LD, Harris TJ. Performance assessment measures for univariate feedback control. Canadian Journal of
Chemical Engineering 1992; 70:11861197.
Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

708

Q. LI, J. R. WHITELEY AND R. R. RHINEHART

3. Kozub DJ. Controller performance monitoring, diagnosis: experiences and challenges. In Proceedings of the 5th
International Conference on Chemical Process Control, Kantor JC et al. (eds). AIChE & CACHE: Tahoe, CA, 1996;
8396.
4. Bezergianni S, Georgakis C. Controller performance assessment based on minimum and open-loop output variance.
Control Engineering Practice 2000; 8:791797.
5. Horch A, Isaksson AJ. A modied index for control performance assessment. Journal of Process Control 1999;
9(6):475483.
6. Horch A. Condition monitoring of control loops. Ph.D. Dissertation, Department of signals, sensors and systems,
Royal Institute of Technology: Stockholm, Sweden, 2000.
7. Harris TJ, Seppala CT, Desborough LD. A review of performance monitoring and assessment techniques for
univariate and multivariate control systems. Journal of Process Control 1999; 9(1):117.
8. Qin SJ. Control performance monitoring}a review and assessment. Computers and Chemical Engineering 1999;
23(2):173186.
9. Huang B, Shah SL. Performance Assessment of Control Loops: Theory and Applications. Springer: New York, 1999.
10. Hagglund T. Automatic detection of sluggish control loops. Control Engineering Practice 1999; 7(12):15051511.
11. Taha O, Dumont GA, Davies MS. Detection and diagnosis of oscillations in control loops. In IEEE International
Conference on Decision and Control, Kobe, Japan, 1996; 24322437.
12. Thornhill NF, Hagglund T. Detection and diagnosis of oscillation in control loops. Control Engineering Practice
1997; 5(10):13431354.
13. Ettaleb L. Control loop performance assessment, oscillation detection. Ph.D. Dissertation, Department of Electrical
and Computer Engineering, University of British Columbia: Vancouver, Canada, 1999.
14. Miao T, Seborg DE. Automatic detection of excessively oscillatory feedback control loops. In Proceedings of 1999
IEEE International Conference on Control Applications, Hawaii, 1999; 359364.
15. Thornhill NF, Huang B, Zhang H. Detection of multiple oscillations in control loops. Journal of Process Control
2003; 13:91100.
16. Perrier M, Roche AA. Towards mill-wide evaluation of control loop performance. In Proceedings of Control
Systems92, Whistler, BC, Canada, 1992; 205209.
17. Desborough LD, Harris TJ. Control performance assessment. Pulp and Paper Canada 1994; 94(11):441.
18. Jofriet P, Seppala CT, Harvey M, Surgenor B, Harris TJ. An expert system for control loop performance. Pulp and
Paper Canada 1996; 97(6):T207T211.
19. Harris TJ, Seppala CT, Jofriet P, Surgenor B. Plant-wide feedback control performance assessment using an expertsystem framework. Control Engineering Practice 1996; 4(9):12971303.
20. Thornhill NF, Oettinger M, Fedenczuk P. Performance assessment and diagnosis of renery control loops.
A.I.Ch.E. Symposium Series 1998; 94(320):373379.
21. Thornhill NF, Oettinger M, Fedenczuk P. Renery-wide control loop performance assessment. Journal of Process
Control 1999; 9(2):109124.
22. Rhinehart RR. A watch dog for controller performance monitoring. In Proceedings of the 1995 American Control
Conference, Seattle, WA, 1995; 22392240.
23. Venkataramanan G, Shukla V, Saini R, Rhinehart RR. An automated on-line monitor of control system
performance. In Proceedings of the American Control Conference, Albuquerque, NM, 1997; 13551359.
24. Narayanaswamy G. Renement and experimental verication of a statistical-based goodness-of-control monitor.
M.S. Thesis, Department of Chemical Engineering. Texas Tech University: Lubbock, TX, 1998.
25. Li Q, Whiteley JR, Rhinehart RR. An automated performance monitor for process controllers. Control Engineering
Practice in press: accepted June 2003.
26. Butler H. Model Reference Adaptive Control. Prentice-Hall: Hemel Hempstead, UK., 1992.
27. Ljung L. System Identication (2 edn). Prentice-Hall PTR: Upper Saddle River, NJ, 1999.

Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Adapt. Control Signal Process. 2003; 17:685708

Anda mungkin juga menyukai