You are on page 1of 41

Control Performance Monitoring { A Review and

Assessment 
S. Joe Qin
Department of Chemical Engineering
University of Texas, Austin, TX 78712
Email: qin@che.utexas.edu; Fax: (512)471-7060
April 30, 1998
Keywords: control performance monitoring; minimum variance control; robustness; root-

cause diagnostics; plant-wide variability assessment; intelligent sensors and valves, model
based control

The original version of this paper was presented at the NSF/NIST Measurement and Control Workshop,
New Orleans, March 6-8, 1998.


S. Joe Qin
Department of Chemical Engineering
University of Texas, Austin, TX 78712

Abstract

In this paper we present an overview of current status of control performance monitoring using minimum variance principles. Extensions to PID-achievable performance
assessment, trade-o between performance and robustness, and trade-o between deterministic and stochastic performance objectives are discussed. Future directions are
pointed out for research and practice with regard to root-cause diagnosis, plant-wide
performance assessment, multivariable assessment, adequacy assessment of existing
control strategies, performance assessment of model predictive control, and the use
of intelligent eld devices and arti cial intelligence to form a systematic diagnostic
methodology. A brief tutorial on performance assessment is given in the appendix
with an industrial process example.

1 Introduction
There is a recent resurgence of interest in control loop performance assessment and diagnosis
due to the work of Harris (1989). In his work Harris proposed the use of closed-loop data
to evaluate and diagnose controller performance using minimum variance control (Astrom,
1967) as a benchmark. Astrom (1967) proposed the minimum variance control (MVC) principle and the use of auto-correlation to indicate how close the existing controller performance
is to that of a minimum variance controller. DeVries and Wu (1978) extended Astrom's work
to assess multivariable processes by estimating the variance of one-step ahead prediction error using normal process data. This work further used spectral analysis to indicate the cause
of a strong frequency component in a process variable. Harris (1989) proposed the use of
minimum variance controller as a lower bound to assess the performance of single loop controllers. The lower bound is estimated from closed-loop operating data accounting for the
process time delay, which was ignored in DeVries and Wu (1978). Further work by Desborough and Harris (1992, 1993) proposed (i) the use of a performance index which is the ratio
of the best achievable variance to the variance of the controlled variable under assessment;
and (ii) the use of analysis of variance (ANOVA) for feedforward/feedback control loops to
indicate whether a poor performance is due to the feedforward controller or the feedback
controller. Stanfelj, et al. (1993) discussed the use of cross-correlation analysis for feedforward/feedback control loops to diagnose the root cause of a poor performance. Although
the process time delay plays an important role in estimating the MVC-achievable variance,
the estimation of time delay from closed loop operating data is not addressed until Lynch
and Dumont (1996). In their paper Lynch and Dumont discussed (i) the use of the xed
model variable regressors proposed by Elnaggar (1990) to estimate the process time delay;
(ii) the use of a Laguerre network to model the controller output instead of an ARMA model
by Harris (1989); and (iii) to determine the degree of process nonlinearity by monitoring
the static input-output relation. Extensions of the performance assessment to non-minimum
phase processes and MIMO processes are discussed in Huang and Shah (1997) and Huang, et
al. (1997). Kendra and Cinar (1997) discussed the use of frequency analysis approaches for
3

the monitoring of loop performance. Tyler and Morari (1995) proposed the use of likelihood
ratio to determine if the control performance is acceptable or not. Kozub (1996) summarized
some recent work in this area up to early 1996.
The MVC based variance is only exactly achievable when a minimum variance controller
is used with a perfectly known process and a disturbance model, which requires at least
a Smith predictor control structure for processes with time delays. In practice, however,
more than 90% of industrial control loops are PID type without time delay compensation.
Therefore, no matter how the PID parameters are tuned, the MVC-based variance is not
exactly achievable for PID controllers when time delay is signi cant or the disturbance is
non-stationary. Eriksson and Isaksson (1994) addressed this point and proposed to use a PI
controller as a benchmark for a special type of disturbance model. Another realistic performance measure proposed by Ko and Edgar (1998) calculates a lower bound of the variance by
restricting the controller type to PID only and allows for more general disturbance models.
The PID-achievable lower bound is generally larger than that calculated from MVC, but it
is possibly achievable by a PID controller.
The aforementioned work in performance monitoring is concerned with the assessment of
the output variance due to unmeasured, stochastic disturbances, which are further assumed
to be generated from a dynamic system driven by white noise. For this reason, we refer
to this class of the performance monitoring methods as stochastic performance monitoring.
While these methods bring up an important aspect of the controller performance, they do not
provide any information about the traditionally concerned performance, such as step changes
in setpoint or disturbance variables, settling time, decay ratio, and stability margin of the
control system. We refer to this class of monitoring techniques as deterministic performance
monitoring. Astrom (1991) discussed some recent developments in this area. Shinskey (1994)
proposed the use of time delay to determine the best achievable deterministic performance.
Swanda and Seborg (1997) use settling time to evaluate the deterministiccontrol performance
of PID controllers. As indicated by the analysis given later in this paper, the two types of
performance usually cannot be best achieved simultaneously. In this paper, we focus on the
discussion of stochastic performance monitoring.
4

The organization of this paper is given as follows. Section 2 provides an overview of


major methodology development in control performance monitoring with an assessment.
Section 3 discusses several extensions to the performance monitoring problem. Some future
directions are pointed out in section 4. The last section provides conclusions to the paper.
A brief tutorial on control performance assessment is given in Appendix A for those who are
interested in implementing the assessment algorithm.

2 Assessment Based on Minimum Variance Principles


2.1 Minimum Variance Control

A discretized SISO process model can be described by the following z-transfer function
representation,
y(k) = Gu (z?1)u(k) + Gw (z?1)w(k)
(1)
where

Bu (z?1) z?d = G (z?1)z?d+1


Gu (z?1) = A
u
(z?1)
u

is the transfer function from the manipulated variable u to the controlled variable y with a
time delay (d ? 1). Gu (z?1)  BuA(uz(?z1?)1z)?1 is the process model without time delay. The noise
dynamics
?1
(z?1)
Gw (z?1) = AB(wz(?z1)r) q = A (zB?1w)(1
? z?1)q
w

has the form of ARIMA(p,q,r). This ARIMA(p,q,r) is able to account for nonstationary
disturbances even though w(k) is zero-mean white noise (MacGregor, et al., 1984; Box and
Jenkins, 1970). In this section we assume all the polynomials are stable, i.e., all poles and
zeros are inside the unit circle. We further assume that Au, Aw and Bw are monic, and
the variance of the noise w(k) is w2 . For non-minimum phase processes (i.e., unstable Bu),
MVC can be designed with minor modi cation and the minimum variance calculation can
be found in Huang and Shah (1997).
The minimum variance control rst derived by Astrom (1967) is feedback control which
achieves minimum output variance. Therefore, it can also be viewed in an internal model
control (IMC) structure or a Smith predictor structure. The block diagrams for the three
5

equivalent structures are shown in Figure 1. The controller transfer functions in the MVC
IMC
SP
feedback form, IMC form, and Smith predictor form are denoted as GMV
c , Gc , and Gc ,
respectively. The equivalence between MVC and IMC was revealed by Bergh and MacGregor
(1987) to analyze the robustness of MVC. The IMC form is convenient to analyze the case
where the process Gp is di erent from the model Gu .
Next we derive the minimum variance controller using the IMC structure shown in Figure 1(b), because the derivation appears to be more straightforward. The noise transfer
function Gw (z?1) can be rearranged in the following form by a long division of Aw(z?1 )rq
into Bw (z?1),
Bw = F + z?d G
A rq
A rq
w

or

Aw rqF + z?dG = Bw
(2)
which is known as a Diophantine equation about F (z?1) and G(z?1 ). The polynomial F is
monic because Aw and Bw are monic. Assuming a zero setpoint and no model mismatch,
i.e., Gp = Gu , the process output can be represented as follows using Figure 1(b) and the
Diophantine equation,
y(k) = Gw w(k) ? Gu GIMC
c Gw w(k )
?
d
Bu ?d
= (F + Az rGq )w(k) ? GIMC
c Gw A z w(k )
w
u
1
B
u Bw IMC
=
A rq (G ? A Gc )w(k ? d) + Fw(k)
w

The implication of the above relation is rather profound. Since the controller GIMC
has to
c
be causal, the rst term on the right hand side of the relation depends on data up to time
k ? d, while the second term depends only on data after k ? d. Therefore, no matter what
controller is used, the two terms are independent. As a result, the variance of the output
2
varfy(k)g  varfFw(k)g = (1 + f12 +    + fd2?1 )w2  MV

(3)

2 , is achieved by the minimum variance controller in the


where the minimum variance, MV
IMC form,
Au G
(4)
GIMC
=B
c
B
u w

The achieved output under minimum variance control is:

y(k) = Fw(k)

(5)

Therefore, the autocorrelation function of y is zero beyond the time delay (d ? 1). Astrom
(1967) rst used this property to assess how close a control performance is from minimum
variance.
For a non-zero setpoint, yr (k), the principle of superposition gives the resulting closed-loop
relation,
y(k) = BG z?dyr (k) + Fw(k)
(6)
w

Now we can easily nd the equivalent MVC controller in the standard feedback form in
Figure 1(a),
GIMC
Au G
c
GMV
(7)
c = 1 ? GIMC G = B FA rq
u

The controller transfer function in the Smith predictor form is,


GIMC
Au
G
c
GSP
(8)
c = 1 ? GIMC G = B B ? z ?1 G
u w
c
u
Consider a special case where the noise dynamics is IMA(0,1,1), i.e.,
?1
Gw (z?1) = 11??cz1z?1
After solving the Diophantine equation, the resulting MVC in three di erent forms is
Au
(1 ? c1)z?1
GMV
=
(9)
c
Bu 1 ? c1z?1 ? (1 ? c1)z?d
Au (1 ? c1)z?1
(10)
GIMC
=
c
Bu 1 ? c1z?1
Au (1 ? c1)z?1
(11)
GSP
=
c
Bu 1 ? z?1
and the closed loop relation is
? c1)z?d y (k) + Fw(k)
y(k) = (11 ?
(12)
c1z?1 r
Therefore, no matter what the process dynamics is, the closed-loop dynamics using MVC
is always rst order plus dead-time. All process poles and zeros are cancelled by the MVC.
7

This is why the MVC is aggressive and sensitive to process changes. Furthermore, the MVC
controller in Eq. 9 is identical in structure to the Dahlin's controller known in digital control
(Dahlin, 1968; Seborg, et al., 1989). The controller in this case is the inverse of the process
model without time delay plus an integrator. The controller in IMC form is the inverse of
the process model without time delay plus a rst order lter. If, in addition, the process
model is rst order, second order, rst order plus dead time (FOPDT), or second order plus
dead time (SOPDT), the resulting transfer functions for the MVC are standard PI, PID, or
Smith predictor with PI or PID controllers. The list of the controller transfer functions for
these special cases are given in Table 1.
From the above analysis, we summarize the relationship between MVC and other more
popular controller forms in practice.
1. If the disturbance is IMA(0,1,1) and the process model is rst order or second order
only, the MVC is a PI or PID controller, respectively.
2. If the disturbance is IMA(0,1,1) and the process model is FOPDT or SOPDT, the
MVC is a Smith predictor with a PI or PID controller, respectively.
3. If the process model is general and the noise model is IMA(0,1,1), the MVC is a
Dahlin's controller with closed-loop time constant  = ?lncT1 , where T is the sampling
interval.1
4. The MVC (GIMC
in Eq. 4) inverts the invertible part of the process dynamics and
c
chooses a ler that has to do with noise dynamics and process time delay only. In the
case of non-minimum phase processes, the unstable zeros are not inverted, similar to
the treatment in IMC design. Refer to Huang and Shah (1997) for further detail.
5. If the noise dynamics is ARMA, i.e., without an integrator, the resulting MVC (Eq. 7)
does not even have integral control. As a special case, if the disturbance is white noise,
the resulting MVC is open-loop, i.e., no feedback control.
If c1 = 0 in the IMA noise model, there is no corresponding Dahlin's controller, but the MVC can still
be designed.
1

The above observations are useful in understanding the pros and cons of using MVC for
controller performance assessment.

2.2 Control Performance Assessment

2.2.1 Feedback Control Only

The tasks for control performance assessment from closed-loop operation data include:
1. Estimate process time delay (e.g., Lynch and Dumont, 1996).
2. Identify the closed-loop model relating process output y to noise w. This model can be
ARMA (Harris, 1989) or other types of time-series models. The identi cation of this
time series model does not require plant testing, but care must be taken in selecting
the data.
2 from Eq. 3.
3. Estimate the minimum variance MV

4. Estimate the actual control error variance y2 from the data.
2 with  2 to see how far the actual performance is from the minimum
5. Compare MV
y
variance performance.

6. As an alternative strategy, calculate the autocorrelation of the output y to see if there


is signi cant correlation beyond the time delay.
The time delay estimation is little addressed in the performance monitoring literature,
but it is critical for estimating the minimum variance from the data. The identi cation
of the time series model is a rather straightforward task. In addition to comparing the
variance of the process output, the use of autocorrelation provides a cross-check for the
control performance assessment based on Eq. 5.
The control performance assessment method is practically appealing for its simplicity
and plant-friendly (i.e., no plant tests needed). A brief, step-by-step tutorial is given in
Appendix A for those who are interested in implementing the monitoring algorithm. Based
9

on the previous analysis and the nature of minimumvariance control, we provide the following
assessment about the method in general.
2 is invariant regardless of the control structure.
1. The theoretical minimum variance MV
It is solely determined by the process time delay and the dynamics of unmeasured
disturbances. In other words, the minimum variance is a lower bound, which may
or may not be achievable, for all types of controllers including PID, PID with Smith
predictors, feedforward control of measured disturbances, model based control, and
2 to assess the performance of other
multivariable control. Therefore, one can use MV
types of controllers.

2. There is no other way to exactly achieve minimum variance than minimum variance
control. The reason is that for linear processes there is only one minimum for the
variance. Making the output variance close to the minimum variance implies that the
actual controller is made close to a minimum variance controller.
2 for PID loop assessment is appropriate (i.e.,  2 is approximately
3. The use of MV
MV
achievable) for the following situations: (i) the process has negligible dead time, such
as ow loops; (ii) the process is low order; and (iii) the unmeasured disturbance is
fairly stationary. Industrial experience indicates that about 20% of control loops in
re nery fall in this category (Kozub, 1998). If the process time delay is signi cant,
minimum variance control requires time-delay compensation.

4. Several other factors that prohibit the minimumvariance from being achievable include:
(i) multivariable interaction; (ii) model plant mismatch, including process nonlinearity;
(iii) model error in noise dynamics; and (iv) process constraints. One could be in a
situation where no improvement can be made based on PID controllers, but the actual
variance is still far from the minimum variance. In this case the use of minimum
variance to assess PID loops can be misleading. More realistic performance bounds,
such as PID-achievable performance bound (Ko and Edgar, 1998), are needed.
10

5. One should be careful not to extremize the stochastic performance by sacri cing the
deterministic performance, such as responses to setpoint and deterministic load disturbance changes. For example, if the noise dynamics is ARMA, the minimum variance
is achieved with no integral control. This will greatly sacri ce the deterministic performance of the controller. Trade-o is necessary between deterministic and stochastic
performance.
6. For process with signi cant time delays, the use of Smith predictors (or model based
control in the general case) can reduce the variance, since MVC shares the same control
structure with Smith predictor. The use of model based control including dead time
compensation will in principle improve both deterministic and stochastic performance
of the controller. The implementation of Smith predictors as a way to improve control
performance is made easier here since the time delay is already estimated during the
performance assessment step.
7. The MVC-based performance assessment is not applicable to processes with varying
time delay, which are often encountered in the process industries. It is also not suitable
in situations where variability is allowed intentionally to reduce the disturbance to other
variables, such as the control of surge tanks.

2.2.2 Assessment of Feedback/Feedforward Control


The assessment of control loops with feedforward controllers is often more important because
these loops are typically more critical and dicult to control. A process subject to m
measured disturbances vi; i = 1; 2    ; m, can be described as follows,

y(k) = Gu u(k) +

m
X
Gv;i vi(k) + Gw w(k)
i=1

(13)

where Gv;i is the transfer function for the i-th feedforward path. For the interest of reducing process variability, feedforward control should always be used to reduce the source of
disturbances. An ine ective feedforward control will contribute a large variance due to the
measured disturbances. Therefore, one additional task in assessing feedback/feedforward
11

control loops is to diagnose whether a poor performance is due to feedback control or feedforward control.
Two methods have been reported for the performance assessment of feedback/feedforward
control loops:
1. Use of analysis of variance (ANOVA) (Desborough and Harris, 1993); and
2. Use of cross-correlation analysis to indicate major contributions to output variance
(Stanfelj, et al., 1993).
The ANOVA method by Desborough and Harris (1993) bears the following assumptions:

 The measured disturbances vi(k); i = 1; 2;    ; m are outputs of some dynamic systems


driven by white noise wi(k), i.e.,

i = 1; 2;    ; m

vi(k) = Hi(k)wi(k);

(14)

 The noise terms w(k) and wi(k); i = 1; 2;    ; m are independent over time and Gaussian.

 The noise terms w(k) and wi(k); i = 1; 2;    ; m are mutually independent.


Eliminating u(k) in Eq.13 with feedforward/feedback controllers,
m
X
u(k) = ?G y(k) + Gff v (k)
c

i=1

and substituting Eq.14 into Eq.13, the following closed loop relation is obtained,

y(k) = (z?1)w(k) +
=

m
X
(z?1)w (k)
i=1

1
m X
1
X
X
ij wi (k ? j )
j w(k ? j ) +
i=1 j =0

j =0

(15)
(16)

Therefore, the variance of the output is,

y2 = w2

2 X 2 X 2
ij
w;i
j+
j =0
i=1
j =0
1
X

12

(17)

Assuming the time delay in Gu is d ? 1 and those in Gv;i are li ? 1; i = 1; 2; : : : ; m,


Desborough and Harris (1993) break the output variance into ve terms,
2 mv
2 fb
2 ff
2 ff=fb
y2 = (2)mv
w + ( )v + ( )w + ( )v + ( )v

(18)

where di = min(d; li) and the de nitions of the variance terms are:

2 Pd?1 2
 (2)mv
w  w j =0 j : minimum variance due to unmeasured disturbance w(k )

Pm 2 Pd?1 2
 (2)mv
v  i=1 w;i j =di ij : minimum variance due to disturbance vi (k )

 (2)fbw  w2 P1j=d j2: variance due to w(k) under non-optimal feedback control

2 Pd+di ?1 2 : variance due to vi (k ) under non-optimal feedforward


 (2)ffv  Pmi=1 w;i
ij
j =d

control

2 P1
2
 (2)ff=fb
 Pmi=1 w;i
v
j =d+di ij : variance due to vi(k ) under non-optimal feedfor-

ward/feedback control

An ANOVA table can be used to assess the contribution of each variance. The sum (2)mv
w +
2
(2)mv
v = MV is the MVC based lower bound for the assessment of feedback/forward control
loops.
The identi cation of the process and disturbance models takes two steps. First closedloop model of ARMAX is identi ed from measured disturbances and output data. Then
each measured disturbance model (Eq.14) is identi ed as an ARIMA time series model.
Time delays (d and li) are required and the model orders also need to be determined. The
identi ed closed-loop model is then used to carry out an ANOVA table for the output
y based on the time delays in the feedforward and feedback paths. This method can yield
valuable information about the sources of variability provided that all measured disturbances
are mutually independent. Although Desborough and Harris (1993) suggest that routinely
operational data can be used in the method, plant tests are sometimes needed in order to
identify the models and break the correlation among the measured disturbances.
The cross-correlation method (Stanfelj, et al., 1993) for root-cause diagnosis of feedback/feedforward control loop performance does not require the closed-loop disturbance
13

models. This method appears to be e ective when the disturbances are not strongly interrelated. The success and limitation of this method in practice is reported in Kozub (1996).
2 must be calculated di erently in
It should be noted that the minimum variance MV
feedback/feedforward control than in feedback control alone. The major di erence is in the
estimation of the variance of the unmeasured disturbance w. An ARMAX model should be
used for identifying the closed loop model to include the e ect of measured disturbances, as
opposed to an ARMA model only.

3 Extensions
One of the signi cant points in the performance assessment work of Harris (1989) lies in the
fact that reducing process variability is critical to industrial processes. In particular, the
stochastic performance is brought to one's attention with an easily practicable approach.
However, it is not advisable to achieve minimum variance at the cost of giving up other important considerations including robustness to process changes, deterministic performance,
knowledge based diagnosis, and the diagnosis of sensor and actuator faults that cause significant deterioration of the control performance. In this section we discuss a few extensions of
the performance monitoring task, including: (i) estimate PID-achievable performance; (ii) robustness and performance trade-o ; (iii) deterministic and stochastic performance trade-o ;
and (iv) knowledge and pattern-recognition-based diagnosis for the health of the controllers,
actuators, and sensors.

3.1 Assessment for PID-Achievable Performance


Since most industrial controllers are of PID type, it is sensible to assess the control performance using a PID-achievable variance. Ko and Edgar (1998) proposed the following
procedure to estimate the PID-achievable performance:
1. Estimate the closed loop transfer function from the noise w to the process output based
on a PID controller.
14

2. Expand the closed loop transfer function in a impulse response form, i.e.,

y(k) =
where the impulse coecient

1
X
i w(k ? i) + w(k )
i=1

is dependent on the PID control parameters.

3. Calculate the variance of y by

y2 =

1
X
2 2
2
i w + w
i=1

4. The PID-achievable variance is derived from


2 = min  2
PID
PID y

Since the actual PID controller is typically di erent from a minimum variance controller,
the resulting closed loop response is of in nite impulse in general. However, one can always
truncate it to a nite impulse response as long as the closed loop system is stable.
Unfortunately, the above procedure is dicult to implement if the open loop process model
is not known. Assuming that the open loop model and the time delay of the process are
known, Ko and Edgar (1998) used the closed loop operation data to identify the disturbance
transfer function and then performed the above procedure. For example, assuming the
process model is given by:
Gu = 1 ? 01:8z?1 z?6

For four di erent disturbance models in Table 2, one can use the above procedure to estimate
the best PID-achievable variance. The PID-achievable variance and the minimum achievable
variance for the four cases are also given in Table 2. As expected, the PID-achievable variance
is always larger than the minimum variance, in one case by more than 100%. This example
indicates that there can be a large di erence between the PID-achievable variance and the
minimum variance.
It is observed from Table 2 that the minimumvariance is more dicult to achieve when the
disturbance is non-stationary (i.e., ARIMA instead of ARMA). On the other hand, processes
with little time delay makes it easier to achieve minimum variance using PID controllers.
15

Therefore, it is helpful to investigate the dependence of the ratio of PID-achievable variance


to MVC variance on time delay and disturbance stationarity via simulation. Consider the
following process,
Gu = 1 ? 01:8z?1 z?d
subject to the following disturbance model,

?1
Gw = (1 ?  z?1)(11 +? 00::42zz?1 )(1 ? 0:5z?1)

To test the e ect of time delay we vary d from 1 to 10. To test the e ect of the disturbance stationarity we vary 1 from 0 to 1. The ratio of the PI-achievable variance to
MVC variance is then calculated and plotted in Figure 2. It is observed that more nonstationary disturbances generally make the MVC variance more dicult to achieve by PI
controllers. However, larger time delay does not necessarily makes MVC variance more dif cult to achieve; the MVC variance is actually easier to achieve when the time delay is very
large. This observation is not intuitive. To further examine this observation, Figure 3 shows
the actual PI-achievable variance and MVC variance vs. time delay for 1 = 0:7. This gure
2 and  2 larger, but their di erence diminishes
shows that more time delay makes both MV
PID
because the MVC variance contains more terms due to larger time delay (Eq.3).
The above experiment shows that the minimum variance can be achievable for a PID
controller when the time delay is very small or very large, but it is not achievable for a PID
controller when the time delay is medium. Practical experience shows that about 20% loops
in re nery can achieve minimum variance using PID controllers (Kozub, 1998).

3.2 Robustness and Performance Trade-o


A well designed controller should not only demonstrate superb performance due to stochastic
disturbances, but also be robust to process changes and model-plant mismatch. The design
for robustness is well studied in the IMC literature (Morari and Za riou, 1989). To make
the controller robust to model-plant mismatch, it is desirable to design an IMC lter. A
larger time constant for the lter will result in a more robust controller in general. This time
constant is equivalent to the desired closed-loop time constant. In fact, this design method
16

is popular in the pulp and paper industry as a tuning method, known as Lambda tuning. If
we wish to design an IMC controller that will result in a closed-loop time constant  with
the following closed loop relation,
c~1)z?d y (k)
y(k) = (11 ?
(19)
? c~1z?1 r
where c~1 = exp(? T ) and T is the sampling interval, the resulting IMC controller is
1 ? c~1 Au
G~ IMC
= 1?
c
c~1z?1 Bu
Assuming further that the noise model is IMA(0,1,1),
1 ? c1
Gw = 1 ?
c1z?1
and there is no process model error, Appendix B shows that the best achievable variance is
2
2 =  2 + (~c1 ? c1)  2
IMC
(20)
MV
1 ? c~21 w
which is larger than the minimum variance.
2 can be used to assess the performance degradation due to the
The above variance IMC
requirement of robustness in design. Given that  is the desired closed loop time constant,
2 . For this reason we refer to  2 as the
the best achievable performance should be IMC
IMC
best IMC-achievable performance for a desired closed-loop time constant.
Equation 20 points out an important con ict between minimum variance and robustness.
From the IMC literature, it is known that a larger c~1 results in more robust control (Palmor
and Shinnar, 1979; Garcia and Morari, 1982). However, a larger c~1 dramatically increases
the IMC-achievable variance. Therefore, in the IMC design, using very large c~1 will increase
the variance.
Equation 20 also points out an theoretical lower bound for selecting the closed loop time
constant, , in IMC design. Having c~1 = c1 will result in the best performance; having
c~1 > c1 will improve the robustness. However, there is no bene t at all for having c~1 < c1.
Therefore, one should choose c~1 to be at least c1, or,
  ? lnTc
1
This lower bound is related to the dynamics of the unmeasured disturbance.
17

3.3 Stochastic and Deterministic Performance Trade-o


Although deterministic disturbances can be realized by stochastic disturbances with ARIMA
models, the performance measures for stochastic and deterministic disturbances are, in
essence, di erent. In the traditional deterministic performance assessment, one is concerned
with the speed of response, settling time, overshoot, and damping ratio (Astrom, 1991).
Achieving minimum variance for the process disturbance during a limited observation period does not necessarily result in acceptable performance for deterministic performance
measures.
Standard PID controllers are one degree of freedom controllers. It is well known that
these controllers cannot achieve the best disturbance rejection and setpoint tracking simultaneously; some trade-o between setpoint tracking and disturbance rejection is necessary.
Therefore, even though deterministic disturbances can be viewed as a special stochastic disturbance, achieving the best disturbance response can result in a poor setpoint response.
Care must be made in trading o the setpoint responses and load disturbance responses,
especially where setpoint change happens frequently.
The example in the previous subsection also depicts the con ict between deterministic
and stochastic performance requirements. In minimum variance control c~1 is chosen to cancel
out the e ect of unmeasured disturbance as much as possible. However, to achieve a desired
closed-loop response for setpoint changes, c~1 has to be chosen to match the desired closed
loop time constant. Therefore, trade-o is necessary between the two objectives.

3.4 Diagnosis of the Entire Systems


While poorly designed controllers are responsible for undesirable control performance, attention should also be paid to other elements in a control system which can signi cantly
deteriorate the performance, such as actuator and sensor malfunctions and faults. Problems
in this category include valve sticking, improper sizing, hysteresis, sensor malfunctioning
and degradation. It is well known that actuator hysteresis can cause limit cycle in the control loop which contributes additional variability to the controlled variable. These problems
18

cannot be overcome by re-tuning the controller.


It is interesting to note that diagnosis of these problems is practiced in some industries
routinely, for example, in pulp and papers (Bialkowski, 1990). Research work in the area
of fault detection for actuator faults may be integrated in control performance diagnosis
(Auburn, et al., 1995). A closely related topic that aims to detect and diagnose oscillation
in control loops can also help the diagnosis of the entire system (Thornhill and Hagglund,
1997; Hagglund, 1995). More interestingly, the development of intelligent sensors, valves,
and Fieldbus devices will provide much more information, such as device self-diagnosis, to
accomplish the task in a systematic manner. Intelligent devices that can diagnose their own
abnormal conditions will alleviate the diagnosis task of the entire system.
It should be noted that assessment and diagnosis of the entire control system performance
cannot be eciently achieved by using statistics based approaches alone. Pattern recognition, arti cial intelligence, and knowledge based approaches should also be used to e ectively
diagnose the root cause of poor performance. Hinde and Cooper (1993) provided an interesting approach for performance diagnosis using a vector quantizing neural network approach.
Early work in pattern recognition based performance evaluation can be found in Bristol
(1977). Further research in combining di erent approaches to performance monitoring and
diagnosis will signi cantly bene t the industry.

4 Future Directions
Recent research activities in the control performance monitoring area bring up the awareness
of evaluating control performance from a statistical point of view. Although stochastic control theory has been available for several decades, most control design strategies in practice
are only concerned with deterministic changes such as step changes. The work in control
performance monitoring may provide an opportunity for the process control community to
pay more attention to the stochastic nature of plant disturbances and consider them in control design. It should be pointed out that there are many theoretical issues and practical
problems that should be addressed before performance monitoring can have a signi cant
19

impact on industrial practice. Based on the previous discussions, we provide the following
suggestions for future directions in controller performance monitoring.
1. Diagnosis of root causes. It is fairly easy to know if a process is not performing
optimally from variance assessment and operation experience, but a challenging task is
how to identify the root cause. Further, the cause of poor performance should not be
limited to controller design and tuning; other elements in the control systems, such as
sensors and actuators, are often responsible for the poor performance. Information from
other sources than the controlled variables should be used simultaneously to identify
the root cause. The next generation of intelligent sensors and valves (e.g., Fieldbus
devices) will help provide additional information about the health of the each element
in the control loop and alleviate the diagnosis task of the entire control system. Work
by Stanfelj, et al. (1993) provided some diagnosis of the root cause. More research
e ort is needed to provide a systematic approach that makes use of all the information
for intelligent diagnosis. Statistical methods such as ANOVA, system identi cation,
and arti cial intelligence approaches should play their roles in such a diagnosis system.
2. Performance monitoring of MIMO systems. So far most work has focused on
SISO or MISO control performance monitoring. However, except for simple ow loops,
most industrial control loops are more or less interactive, making the performance assessment of SISO or MISO systems not applicable. Harris, et al. (1996) and Huang and
Shah (1997) discussed some methods for MIMO process control monitoring. While the
assessment of SISO systems requires only closed-loop operational data, the assessment
of MIMO systems may require additional process information. To make the performance monitoring for MIMO systems practical, it must have the following features
similar to the SISO control assessment.
(a) It requires minimum interference to process operation.
(b) It provides achievable performance bound for the existing control structure, e.g.,
multi-loop control.
20

(c) It should diagnose whether the existing control structure is adequate or not. For
example, it could recommend alternative control structure that will improve the
performance signi cantly.
The third item is more dicult to achieve but is more desirable in practice. It would be
of tremendous value to practitioners if, by assessing the performance, one can determine
whether alternative control strategies are needed to achieve the control objectives.
One could assess the impact of loop integration on process variability and determine
whether multivariable model-based control should be used rather than multi-loop or
SISO control. Traditional loop interaction measures such as relative gain arrays could
be useful in assessing the loop interaction. One could also assess the nonlinearity of
the process, determine that the source of variability is due to process nonlinearity, and
suggest the use of nonlinear control strategies.
Another issue in MIMO control design is to provide appropriate balance or trade-o
between variables that should be tightly control and those which should better be
loosely controlled. This brings up the next point of discussion.
3. Plant-wide variability analysis. Industrial practice indicates that one needs to
transfer variability from where it hurts to where it does not hurt as much. Therefore,
if one focuses only on reducing variability on a univariate basis, it can be misleading.
Methodologies are needed to analyze where the major variability comes from and where
variability can be increased in order to reduce the variability of more critical variables.
To accomplish this task, multivariable assessment theory is needed. Furthermore,
upper limits of a variable's variability should be used in performance assessment in
addition to lower limits.
4. Performance and robustness trade-o . Theoretical development is needed to
assess control performance and robustness in a stochastic framework. This trade-o has
been considered in control design, and should be considered in performance assessment
as well. It is not advisable to reduce the variance by sacri cing robustness and other
21

important factors in control design. One immediate approach to including trade-o


in performance assessment is to use linear quadratic regulators as a benchmark rather
than minimum variance control. Further research is needed in this area.
5. Performance diagnosis for model predictive control. Since model predictive
control is popular in industrial practice and inherently more complicated, it is desirable
to develop theory for monitoring and diagnosis of MPC performance. In this task the
assessment of variability is just one issue among other more important issues, including
model validity, process nonlinearity, constraint feasibility, bottleneck identi cation,
sensor and actuator integrity, and consistency between control and optimization results.

5 Conclusions
The recent interest in control performance monitoring due to Harris (1989) brings up the
importance and applicability of stochastic control theory. While traditional control design
practice has been concerned with deterministic control objectives such as settling time and
overshoot, the statistical view of process disturbances addresses an importance nature of
process disturbances and provides a systematic tool for control performance assessment and
design.
It should also be noted that the use of minimum variance as an assessment benchmark can
be neither achievable (by PID controllers) nor desirable to achieve. One should be cautious
about pushing the stochastic performance to the limit, which is only one of several factors
of practical concern. Trade-o s between robustness and performance, between stochastic
and deterministic performance, and between setpoint and load disturbance changes should
be considered; these trade-o s make the minimum variance undesirable to achieve. Future
research activities should address these trade-o s, assessment of MIMO or multi-loop control
systems, plant-wide variability assessment, root cause diagnosis, the inclusion of sensors
and actuators in the assessment task, assessment and diagnosis of model predictive control
strategies, and assessing the need for more advanced control strategies.
Industrial competitiveness in the coming century will highly depend on how one can
22

reduce variability in critical quality variables so that yield and throughput can be maximized.
Performance monitoring, including assessment and diagnosis, provides one e ective approach
to reducing variability. To meet this challenge of increased global competitiveness, one may
have to move more aggressively towards a broad sensor of collaboration between academia
and industry and between chemical engineers and chemists. The performance monitoring
research is highly relevant to industrial manufacturing. Industrial challenging problems that
address urgent industrial needs should be developed as test beds under the collaboration
between industry and academia. The development of new analytical sensors by chemists
could provide signi cant improvement to variability analysis and reduction, which calls for
an enhanced collaboration between chemists and chemical engineers.

Acknowledgment
Financial support from Air Products and Dupont is gratefully acknowledged. The author is
grateful to Air Products for providing process data for the analysis in this paper.

23

Nomenclature and Acronyms


Au ; B u
Aw ; B w
c1
c~1
d?1
di
F; G
Gc
~GIMC
c
Gff
i
Gp
Gu
Gv;i
Gw
Hi
i; j
k
li ? 1
m
ny ; nw
p; q; r
u
vi
w
y
yr

=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=

polynomials for the process transfer function


polynomials for the noise transfer function
model coecient in IMA(0,1,1)
exp( T )
process time delay
min(d; li)
polynomials in Diophantine equation
controller transfer function
IMC controller achieving closed-loop time constant 
the ith feedforward controller
process transfer function
process model transfer function
process transfer function for the ith feedforward path
unmeasured disturbance transfer function
the ith feedforward disturbance model
dummy indices
sample time instance
time delay for feedforward disturbance vi
number of feedforward variables
model order for the closed-loop ARMA model
ARIMA model order
manipulated variable
the ith feedforward disturbance variable
white noise for unmeasured disturbances
output or controlled variable
setpoint

24

Symbols

T = sampling time
 = desired closed-loop time constant
r = 1 ? z?1
y = auto-correlation function
j = impulse response coecients for (z ?1 )
ij = impulse response coecients for i (z ?1 )
?
1
(z ) = closed-loop transfer function for w(k)
i(z?1) = closed-loop transfer function for vi(k)
2
IMC
= IMC-achievable variance for a desired 
2
MV = MVC-achievable variance
2
PID
= PID-achievable variance
w2 = variance of unmeasured disturbance w
y2 = variance of the controlled variable y
(2)mv
w = minimum variance due to w(k )
(2)mv
= minimum variance due to vi(k)
v
2
fb
( )w = variance due to w(k) and non-optimal feedback control
(2)ff
v = variance due to vi (k ) and non-optimal feedforward control
2
ff=fb
( )v
= variance due to vi(k) and non-optimal feedback/feedforward control

Acronyms
ANOVA
ARIMA
ARMA
ARMAX
FOPDT
IMA
IMC
MIMO
MISO
MVC
PID
SISO
SOPDT
SP

=
=
=
=
=
=
=
=
=
=
=
=
=
=

analysis of variance
auto-regressive integrated moving average
auto-regressive moving average
auto-regressive moving average with exogenous inputs
rst order plus dead time
integrated moving average
internal model control
multiple-input-multiple-output
multiple-input-single-output
minimum variance control
proportional-integral-derivative
single-input-single-output
second order plus dead time
Smith predictor

25

References
[1] K. J. 
Astrom. Computer control of a paper machine { an application of linear stochastic
control theory. IBM J. Res. Dev., 11:389{405, 1967.
[2] K.J. Astrom. Assessment of achieviable performance of simple feedback loops. Int. J.
Adaptive Control and Signal Processing, 5(1):3{19, 1991.
[3] C. Auburn, M. Robert, and T. Cecchin. Fault detection in a control loop. Control Eng.
Practice, 3(10):1441{1446, 1995.
[4] Luis G. Bergh and John F. MacGregor. Constrained minimum variance controllers:
internal model structure and robustness properties. Ind. Eng. Chem. Res., 26:1558{
1564, 1987.
[5] W.L. Bialkowski. Process control for engineers. Control Notes of ENTEC Company,
Toronto, Canada, 1990.
[6] George E. P. Box and Gwilym M. Jenkins. Time Series Analysis: Forecasting and
Control. Holden-Day, Oakland, California, 1976.
[7] E.H. Bristol. Pattern recognition: an alternative to parameter identi cation in adaptive
control. Automatica, 13:197, 1977.
[8] E.B. Dahlin. Designing and tuning digital controllers. Instrum. and Control Systems,
41(6):77, 1968.
[9] L. Desborough and T. J. Harris. Performance assessment measures for univariate feedback control. Can. J. Chem. Eng., 70:1186, 1992.
[10] L. Desborough and T. J. Harris. Performance assessment measures for univariate feedforward/feedback control. Can. J. Chem. Eng., 71:605, 1993.
[11] W.R. DeVries and S.M. Wu. Evaluation of process control e ectiveness and diagnosis
of variation in paper basis weight via multivariate time-series analysis. IEEE Trans.
Auto. Cont., 23:702{708, 1978.
26

[12] A. Elnaggar. Variable regression estimation of unknown system delay. Univ. British
Columbia, 1990. Ph.D. Dissertation.
[13] Carlos E. Garca and Manfred Morari. Internal model control: 1. A unifying review and
some new results. Ind. Eng. Chem. Proc. Des. Dev., 21:308{323, 1982.
[14] T. Hagglund. A control-loop performance monitor. Control Eng. Practice, 3:1543{1551,
1995.
[15] T. Harris, F. Boudreau, and J.F. Macgregor. Performance assessment of multivariable
feedback controllers. Automatica, 32(11):1505{1518, 1996.
[16] T. J. Harris. Assessment of control loop performance. Can. J. Chem. Eng., 67(10):856{
861, 1989.
[17] R.F. Hinde and D.J. Cooper. Using pattern recognition in controller adaptation and
performance evaluation. In American Control Conf., pages 74{78, San Francisco, June
2-4 1993.
[18] B. Huang and S.L. Shah. Feedback control performance assessment of non-minimum
phase MIMO systems. In AIChE Annual Meeting, Los Angeles, Nov. 16-21 1997.
[19] B. Huang, S.L. Shah, and K.Y. Kwok. Good, bad or optimal? performance assessment
of MIMO processes. Automatica, 33(6):1175{1183, 1997.
[20] S.J. Kendra and A. Cinar. Controller performance assessment by frequency domain
techniques. J. Proc. Cont., 7(3):181{194, 1997.
[21] B. Ko and T.F. Edgar. Assessment of achievable PI control performance for linear
processes with dead time. In American Control Conf., Philadelphia, PA, June 1998.
[22] D. Kozub. Perfonal communication. February 1998.
[23] D.J. Kozub. Controller performance monitoring and diagnosis: experiences and challenges. In J.C. Kantor, C.E. Garcia, and B.C. Carnahan, editors, Fifth Int. Conf. on
Chemical Process Control, pages 83{96, Tahoe, CA, 1996. AIChE and CACHE.
27

[24] C.B. Lynch and G.A. Dumont. Control loop performance monitoring. IEEE Trans.
Cont. Sys. Tech., 4(2):185{192, 1996.
[25] J.F. Macgregor, T.J. Harris, and J.D. Wright. Duality between the control of processes subject to randomly occurring deterministic disturbances and ARIMA stochastic
disturbances. Technometrics, 26(4):389{397, 1984.
[26] Manfred Morari and Evanghelous Za riou. Robust Process Control. Prentice-Hall, Englewood Cli s, New Jersey, 1989.
[27] Z. Palmor and R. Shinnar. Ind. Eng. Chem. Process Des. Dev., 18:8, 1979.
[28] D. E. Seborg, T. F. Edgar, and D. A. Mellichamp. Process Dynamics and Control. John
Wiley and Sons, New York, 1989.
[29] F. G. Shinskey. Feedback Controllers for the Process Industries. McGraw-Hill, New
York, 1994.
[30] N. Stanfelj, T.E. Marlin, and J.F. MacGregor. Monitoring and diagnosis of process
control performance: The single-loop case. Ind. Eng. Chem. Res., 67(10):856{861, 1993.
[31] A. Swanda and D. Seborg. Evaluating the performance of PID-type feedback control
loops using normalized settling time. In ADCHEM 97, Ban , Canada, June 9 -11 1997.
[32] N.F. Thornhill and T. Hagglund. Detection and diagnosis of oscillation in control loops.
Control Eng. Practice, 5:1343{1354, 1997.
[33] M.L. Tyler and M. Morari. Performance monitoring of control systems using likelihood
ratio methods. In Proceedings of the 1995 American Control Conference, pages 1245{
1249, Seattle, Washington, June 1995.

28

Appendix A: Tutorial on Performance Assessment


2 from
The task of controller performance assessment is to estimate the minimum variance MV
the output data y(k) under closed-loop control (refer to Figure 1(a)). Then the estimated
2 is compared to the variance of output,  2, to see if the existing variance is close to
MV
y
minimum variance. If it is, then no improvement can be made with the existing process and
control design; if not, re-tuning of the controller can be performed to reduce the variance.
One needs only the following information about the process,

 appropriately collected closed-loop data for the controlled variable; and


 process time delay in whole periods of the sampling time.
The controlled variable y(k) is then modeled as a time series, often treated as an ARMA
process. There are two approaches one can use to examine if the current control is achieving
a variance that is close to minimum variance: one is to estimate the minimum variance,
2 , and compare it with the current variance of the output, the other is to check if the
MV
controlled variable y(k) has signi cant auto-correlation coecients beyond the time delay.
What follows is a detailed explanation of each approach. An industrial pH reactor process
is used to demonstrate the procedure involved in each approach.
2 and  2
Estimating MV
y
2 and  2 .
The following procedure can be followed to estimate MV
y

1. Collect the output data series fy(k); k = 1; 2;    ; N g. Make sure that the data are
free of outliers and missing points. The data should not be compressed and should be
long enough to cover all possible disturbances.
2. Estimate an ARMA model for the time series y(k),

y(k) =

n
n
X
X
a y(k ? i) + b w(k ? i) + w(k)
y

i=1

i=1

The model order (ny ; nw ) can be selected using a criterion (Box and Jenkins, 1976),
prior knowledge, or trial-and-error.
29

3. Expand the ARMA model into impulse response model and truncate at the rst d ? 1
coecients,
dX
?1
y(k) = w(k) +
i w(k ? i) + remainder
i=1

The rst d terms are invariant no matter what controller is used.


4. The minimum variance estimate is given by,

2

MV

=1+

d?1
i=1

5. The estimate of the output variance is given by

N
X
y2 = N 1? 1 (y(k) ? yr )2
k=1

The variances can vary with time in practice due to the time-varying nature of the disturbances and processes. Therefore, it is sensible to re-estimate the variances from time to
time.
Example. Industrial data from a wastewater treatment pH reactor are used to illustrate
2 and  2. The controlled variable is the pH of the reactor
the procedure for estimating MV
y
and the manipulated variable is acid ow. A major disturbance is the wastewater ow and
content which cause the reactor pH to vary. The reactor has about 45 minutes residence
time. The process data were sampled every minute and over 35,000 sample were collected,
as shown in Figure 4. From Figure 4 it is seen that several portions of the data contain
bad values. Therefore, we pick up three periods of the data which do not contain these bad
values. We also try to estimate each period separately to test the time-varying nature of the
process and disturbances.
A separate identi cation e ort discovered that the process time delay is one minute (i.e.,
d = 2). The data periods 1, 2, and 3 contain over 8000, 9000, and 5500 samples, respectively. A second order ARMA model is estimated from each data period using the System
Identi cation Toolbox of Matlab. The estimated ARMA models are:

z?1 + 0:090z?2 w(k)


Period 1: y(k) = 11??01::882
89z?1 + 0:918z?2
30

?1 + 0:170z ?2
1
?
0
:
803
z
Period 2: y(k) = 1 ? 1:71z?1 + 0:734z?2 w(k)
? 1:07z?1 + 0:251z?2 w(k)
Period 3: y(k) = 11 ?
1:91z?1 + 0:954z?2
After performing a long division, the rst d = 2 term are,

Period 1: y(k) = w(k) + 1:01w(k ? 1) + remainder


Period 2: y(k) = w(k) + 0:908w(k ? 1) + remainder
Period 3: y(k) = w(k) + 0:945w(k ? 1) + remainder
The estimated variances of the noise w(k) can also be obtained from the Matlab ARMA
function as follows: 0:0063, 0:0149, and 0:0085, respectively. Therefore, the corresponding
estimates of the minimum variances are,
Period 1:
Period 2:
Period 3:

2
MV
= (1 + 1:012)  0:0063 = 0:0127
2
MV
= (1 + 0:9082 )  0:0149 = 0:0232
2
= (1 + 0:9452 )  0:0045 = 0:0085
MV

The corresponding y2 for the periods are,


Period 1:
Period 2:
Period 3:

y2 = 0:0619
y2 = 0:1603
y2 = 0:0525

It can be seen that the minimum variance estimates vary somewhat for the three periods,
but it is clear that the actual variances are much larger than the minimum variances, which
suggests the need for controller re-tuning.

Estimating auto-correlation
For the time series fy(k); k = 1; 2;    ; N g, the auto-correlation function is estimated using
the following relation,
NX
?j
y (j ) = N1 (y(k) ? yr )(y(k + j ) ? yr )
k=1
31

where yr is the setpoint. In case that the control has an o set, the mean of y(k) should
replace yr in the calculation.
To check whether y (j ) is signi cantly di erent from zero, a con dence limit needs to be
calculated. Box and Jenkins (1976) shows that, if the coecient y (j ) is zero for j  d, then
the variance
dX
?1
1
var[y(j )]  N [1 + 2 2y (i)]; j  d
i=1

h q

Therefore, the 95% con dence interval for y (j ) is ?2 var[y (j )]; +2 var[y(j )] . If all
auto-correlation coecients y (j ) are inside this interval for j  d, the control is roughly
achieving minimum variance; otherwise, it is not.
Example. The same pH reactor example is used to illustrate the calculation of autocorrelation function. Figure 5 depicts the auto-correlation estimates for all three periods
of the data. Con dence limits of 95% con dence level are also plotted in Figure 5. It is
observed that the auto-correlation functions are well outside the con dence limits after the
time delay of one. Therefore, we conclude that the control is not achieving minimumvariance.
Furthermore, the auto-correlation functions are oscillatory, indicating that oscillation exists
in the original data y(k).

Appendix B: IMC-achievable variance


For a given noise model
and an IMC controller
the process output is given by,

Gw = 11??c1zz?1

?1

1 ? c~1 Au ;
G~ IMC
= 1?
c
c~ z?1 B
1

y(k) = Gw w(k) ? G~ IMC


c Gu Gw w(k )
?
d
= (F + Az rGq )w(k) ? G~IMC
c Gu Gw w(k )
w
= Fw(k) + 1 c~?1 ?c~ zc1?1 w(k ? d)
1

32

Therefore, the variance of y(k) is


2
2 =  2 + (~c1 ? c1)  2 :
IMC
MV
1 ? c~2 w
1

33

Table 1: Summary of the MVC forms for the IMA(0,1,1) disturbance model
Process
Gu
GMV
GSP
Notes
c
c
?
1
?
1
(1
?
c
b
z
1 )(1+a1 z )
1
First order
PI
1+a1 z?1
b1 (1?z?1 )
?
1
?
2
?
1
?
2
a1 z +a2 z )
Second order 1+b1az1z?+1 +b2az2z?2 (1(1??cz1?)(1+
PID
1 )(b1 z ?1 +b2 z ?2 )
?
1)
?
d
(1
?
c
)(1+
a
z
b1 z
1
1
FOPDT
PI + SP
1+a1 z?1
b1 (1?z?1 )
?
1
?
2)
?
d +b z ?d?1
(1
?
c
)(1+
a
z
+
a
z
b
z
1
1
2
1
2
SOPDT
PID + SP
1+a1 z?1 +a2 z?2
(1?z?1 )(b1?+1b2 z?1 )
(1?c1 )z Au
Au
Bu z ?d
1?c1
General
Dahlin's
Au
1?c1 z?1 ?(1?c1 )z?d Bu
1?z?1 Bu

34

Table 2: PID achievable variance for an FOPDT process


Disturbance model
PID-achievable var. Minimum variance
?
1
1?0:2z
1.11
1.11
(1?0:3z?1 )(1+0:4z??11 )(1?0:5z?1)
1+0
:6z
3.44
3.40
(1?0:6z?1 )(1?0:5z??11 )(1+0:7z?1)
1?0:2z
17.75
11.95
(1?0:3z?1 )(1+0:4z?1 )(1??1 0:5z?1 )(1?z?1 )
1+0
:6z
123.54
58.34
(1?0:6z?1 )(1?0:5z?1 )(1+0:7z?1 )(1?z?1 )

35

w(k)
yr(k)

MV

Gc

u(k)

Gw
+
Gp

y(k)

(a)
w(k)
yr(k)

IMC

Gc

u(k)

Gw
+
Gp

y(k)

+
+
-

Gu
(b)
w(k)
yr(k)

SP

Gc

u(k)

Gw
+
Gp

Gu*

y(k)

+
z-d+1

+
-

(c)

Figure 1: A block diagram of the minimum variance controller.

36

1.2

1.15

1.1

1.05

1
1
0.8

10
0.6

8
6

0.4

0.2
Phi_1

2
0

Timde delay

2 to  2 versus time delay and 1


Figure 2: The ratio of PID
MV

37

2.2

2
MVC
PI

Variances

1.8

1.6

1.4

1.2

1
0

5
6
Timde delay

2 and  2 versus time delay


Figure 3: The variances PID
MV

38

10

2.5

1.5

0.5

0.5

1.5
0

period 1
0.5

period 2
1

1.5

period 3
2

2.5

3.5

4
4

x 10

Figure 4: The closed-loop pH data.

39

Autocorrelation

period 1

1
0.5

0
0

period 1

2
0

5000

0.5
0

10000

0.5

2
0

5000

0.5
0

10000

period 1

50
Autocorrelation

100

50
Autocorrelation

100

50

100

1
0
0
1
0

2000

4000

1
0

6000

Figure 5: The estimated auto-correlation functions for three periods of data.

40

List of Figures
1
2
3
4
5

A block diagram of the minimum variance controller. . . . . . . . .


2 to  2 versus time delay and 1 . . . . . . . . .
The ratio of PID
MV
2 and  2 versus time delay . . . . . . . . . . .
The variances PID
MV
The closed-loop pH data. . . . . . . . . . . . . . . . . . . . . . . . .
The estimated auto-correlation functions for three periods of data. .

41

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

36
37
38
39
40

You might also like