Chemical Engineering Science 69 (2012) 287295

Contents lists available at SciVerse ScienceDirect

Chemical Engineering Science

journal homepage: [Link]/locate/ces

Monitoring and retuning of low-level PID control loops

Akradej Leosirikul a, David Chilin a, Jinfeng Liu a, James F. Davis a, Panagiotis D. Christodes a,b,n
a
b

Department of Chemical and Biomolecular Engineering, University of California, Los Angeles, CA 90095-1592, USA
Department of Electrical Engineering, University of California, Los Angeles, CA 90095-1592, USA

a r t i c l e i n f o

abstract

Article history:
Received 5 September 2011
Received in revised form
16 October 2011
Accepted 20 October 2011
Available online 29 October 2011

In this work, we focus on the problem of monitoring and retuning of low-level proportional-integralderivative (PID) control loops used to regulate control actuators to the values computed by advanced
model-based control systems like model predictive control (MPC). We consider the case where the realtime measurement of the actuation level is unavailable, and thus PID controller monitoring has to be
achieved on the basis of process state measurements. A fault detection and isolation (FDI) method
involving process models and real-time process measurements is used to monitor the PID control loops
and compute appropriate residuals. Once poor tuning is detected and isolated, a PID tuning method
based on the estimated transfer function of the control actuator is applied to the isolated, poorly
functioning PID controller. An example of a non-linear reactorseparator process operating under MPC
with low-level PID controllers regulating the control actuators is used to demonstrate the approach.
& 2011 Elsevier Ltd. All rights reserved.

Keywords:
Process control
PID control
Stability
Simulation
Process monitoring
Chemical processes

1. Introduction
Model predictive control (MPC) is widely used in industry
because of its ability to handle input/state constraints and to
incorporate optimization considerations in a single formulation
(Mayne et al., 2000; Qin and Badgwell, 2003). In general, in the
calculation of the optimal input trajectories for the manipulated
inputs via MPC, the dynamics of the corresponding control actuators
that will implement the control actions computed by the MPC are
neglected and the MPC-computed control actions are assumed to be
directly implemented by the control actuators. However, in practice,
these control actuators have their own specic dynamics. As a result
of this, there are always discrepancies (i.e., time lags, magnitude
differences, etc.) between the actual control actions applied to the
process by the control actuators and the control actions requested
by the MPC. To mitigate the inuence of these discrepancies in
closed-loop performance, PID controllers (typically called lowlevel PID controllers) are usually implemented on the control
actuators to regulate the outputs of the actuators at the values
requested by the MPC (Astrom et al., 1993). The representation of
this added extra layer of the PID controllers around the control
actuators is shown in Fig. 1. In this case, the tuning of the PID
controllers is critical for the overall control actuator and closed-loop
system performance. An actuator with a well-tuned PID controller

n
Corresponding author at: Department of Chemical and Biomolecular
Engineering, University of California, Los Angeles, CA 90095-1592, USA.
E-mail address: pdc@[Link] (P.D. Christodes).

0009-2509/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/[Link].2011.10.046

can effectively implement the actions requested by the MPC;

whereas, an actuator with a poorly tuned PID controller may reduce
the performance of the closed-loop system dramatically or may
even cause instability of the closed-loop system.
Monitoring the performance of low-level PID loops provides
the motivation for this work. With respect to previous works on
the subject, there is indeed a plethora of techniques discussed in
the literature on monitoring of the performance and tuning of PID
controller parameters. With respect to tuning, methods such as
ZieglerNichols (Ziegler and Nichols, 1942; Hang et al., 1991),
CohenCoon (Cohen and Coon, 1953), internal model control
(Skogestad, 2003; Veronesi and Visioli, 2009), pole placement
(Wang et al., 2009; Zayed et al., 2009), and others have been
widely used to tune PID controller parameters based on either the
estimated plants transfer function or experimentally obtained
step response and/or frequency response curves. Gain scheduling
(Rugh, 1991; Zhao et al., 1993) has also been developed to allow
PID controllers to be able to self-tune to accommodate changing
operating conditions. Multiple works have also been published on
automatic retuning of PID parameters based on the current
performance of the PID controller and online system identication (Sung et al., 1998; Veronesi and Visioli, 2009; Zhuang and
Atherton, 1993; Chand, 1992; Saito, 1990; Teng et al., 2008;
Anderson et al., 1988; Nishikawa et al., 1984). On the monitoring
front, Eriksson and Isaksson (1994) and Qin (1998) provide a
survey of available monitoring techniques. Specically, minimum
variance control (Harris, 1989) has been developed as a tool to
assess PID performance, while Tsung and Shi (1999), Tsung et al.
(1999), and Tsung (2000) utilize statistical process control (SPC)

288

A. Leosirikul et al. / Chemical Engineering Science 69 (2012) 287295

Fig. 1. Closed-loop system with MPC as advanced model-based controller and low-level PID controller implemented to regulate the control actuators.

to monitor and provide performance criteria to assess the performance of PID controllers. In another work (Shi and Tsung, 2003), a
monitoring scheme was proposed to determine poor tuning/faults
using principal component analysis (PCA) and neural networks.
One common feature in all of the works in the PID monitoring
eld mentioned above is the assumption that measurements of
the output of the PID-controlled loop are available.
Motivated by the above considerations, we address the problem
of real-time monitoring and retuning of low-level PID controllers in
the case where the measurement of the actual control action
implemented on the process is unavailable. Specically, we present
a method for monitoring the PID performance via a model-based FDI
method (Mhaskar et al., 2006, 2007) coupled with real-time process
measurements. Using an estimated transfer function model of the
control actuators, model-based FDI can be used to detect the
discrepancies between the expected actuation level and the actual
actuation level performed by the control actuators. Based on the
patterns of the residuals, a poorly tuned actuator can be isolated and
retuned accordingly. An example of a non-linear reactorseparator
process under MPC with low-level PID controllers around the
control actuators is used to demonstrate the approach.

model predictive controller (LMPC) (Mhaskar et al., 2006). One

assumption about the design of the model-based control system
used in this work is that it does not explicitly account for the
dynamics of the control actuators and the presence of the process
noise. Therefore, the model used for the design of the modelbased control system assumes the following dynamics for the
process:
~
~
x~_ t f xt
Gxtu
m t

where um is the computed actuation by the high-level MPC.

We make the following assumptions regarding the stability of
the closed-loop system. We assume that there exists a Lyapunov~ such that the origin of the nominal closedbased controller hx
loop system under this controller, i.e., system of Eq. (2) with
~ 8t, is asymptotically stable; this controller hx
~ can be
um t hx
designed using geometric control or Lyapunov-based control
techniques (e.g., Christodes and El-Farra, 2005). Using converse
Lyapunov theorems, this implies that there exist class K
functions1ai , i 1; 2,3; 4 and a continuously differentiable Lya~ for the nominal closed-loop system that
punov function Vx
satisfy the following inequalities:
~ r Vx
~ r a2 JxJ
~
a1 JxJ

2. Preliminaries
2.1. Class of non-linear systems
In this work, we consider non-linear process systems with
constraints on the inputs described by the following state-space
model:
_ f xt Gxtua t wt
xt

where xt A Rnx is an nx-element column vector representing nx

states of the system, ua t A U D Rmu is an mu-element column
vector representing mu inputs to the system, and wt A W D Rnx is
an nx-element column vector representing the process noise to
the system. U is a convex set, f  is a non-linear sufciently
smooth vector function, and G is a nx  mu matrix whose
elements are sufciently smooth functions that relate the jth
input to the ith state with 1r j r mu and 1r i rnx . Without loss
of generality, x 0 is assumed to be the equilibrium of the
_ 0 when x 0, ua 0, and w0. The
unforced system, i.e., xt
operator 9  9 is used to denote the absolute value of a scalar. The
operator J  J is used to denote Euclidean norm of a vector.
Since the central focus of this work is on the difference
between the requested actuation computed by the model-based
controller and the actual actuation level applied to the process by
the control actuators, we shall distinguish the two elements by
calling the requested actuation um(t) and the actual actuation ua(t).

~
@Vx
~ Gxh
~ x
~ r a3 JxJ
~
f x
@x


@Vx
~ 


~
 @x  r a4 JxJ

3a
3b

3b

for all x~ A D D Rnx where D is an open neighborhood of the origin.

We denote the region Or D D2 as the stability region of the
nominal closed-loop system, i.e., Eq. (2), under the control
~
um t hx.
~ allows us to formulate an
The existence of the controller hx
~ (Mhaskar et al.,
MPC that inherits the stability properties of hx
2006), and it is described by the following optimization problem:
Z Nc D
T
^ t uTc tRuc t dt
min
x^ tQ x
4a
uc A SD

^ t Gx
^ tuc t
x^_ t f x

4b

^
x0
xt k

4c

uc t A U

4d

@Vxt k
@Vxt k
Gxt k uc 0 r
Gxt k hxt k
@x
@x

4e

where SD is the family of piece-wise constant functions with

sampling period D, Q and R are strictly positive denite symmetric weighting matrices, xt k is the process state measurement

2.2. Lyapunov-based MPC

In this work, the model-based controller that is used to
determine the set-points for each actuator is a Lyapunov-based

1
A continuous function a : 0,a-0,1 is said to belong to class K if it is
strictly increasing and a0 0.
2
We use Or to denote the set Or : fx A Rnx 9Vx r rg.

A. Leosirikul et al. / Chemical Engineering Science 69 (2012) 287295

obtained at tk, x^ is the predicted trajectory of the system under

the MPC, Nc is the number of steps in the prediction horizon, and
~
V is the Lyapunov function corresponding to the controller hx.
The optimal solution to this optimization problem is denoted
by unc t9t k . The LMPC is implemented following a receding
horizon strategy; at each sampling time tk, a new state measurement xt k is received from the sensors and the optimization
problem of Eq. (4) is solved, and unc 09t k is sent to the actuators
and it is implemented for t A t k ,t k 1 . The constraint of Eq. (4e)
guarantees that the value of the time derivative of the Lyapunov
function at the initial evaluation time of the LMPC is less than or
equal to the value obtained if only the Lyapunov-based control
~ is implemented. This constraint allows the LMPC to
um hx
~
inherit the stability properties of the Lyapunov-based control hx
for sufciently small sampling period D; in particular, practical
stability of the closed-loop system can be proven for sufciently
small D. For detailed results on Lyapunov-based MPC, see
Mhaskar et al. (2006).
Remark 1. Note that in the design of the LMPC of Eq. (4) and its
closed-loop stability analysis, one assumption is that the
requested actuation um(t) is applied directly to the process by
the control actuators. In a practical setting, however, um(t) has to
go through the dynamics of the PID-controlled actuators before
the system is actuated with ua(t). The central focus of this work is
on how to bring ua(t) to be as close as possible to um(t). The
relationship between ua(t) and um(t) will be discussed in detail in
the next section.
Remark 2. Though a Lyapunov-based MPC is used in this paper
as the model-based control system to demonstrate how the
problem of low-level PID monitoring and retuning based on
process state measurements can be approached, the monitoring
and retuning methods presented here can be applied to any type
of model-based control system (i.e., geometric control or Lyapunov-based control, Christodes and El-Farra, 2005; distributed
MPC, Liu et al., 2009, 2010, etc.). Specically, as long as the
requested actuation level um(t) and the process state measurements are available to the monitoring and retuning system at all
times, the same method presented in this work can be applied to
detect the deviation of the actual actuation level ua(t) from the
requested actuation level um(t).
2.3. Low-level PID control loops
As depicted in Fig. 1, um(t) is sent from the model-based
controller as the set-point to the control actuators. PID controllers
are installed around these control actuators to help accelerate the
actuators response so that ua(t) can approach the value of um(t)
faster. Eq. (5) shows the relationship between um and ua in the
Laplace domain:
ua s

Gp Gc
um s
1 Gp Gc

where Gp is the actuators transfer function and Gc is the PID

controllers transfer function. Gc contains three parameters: Kc
(proportional gain), tI (integral time constant), and tD (derivative
time constant) and takes the following form:


1
tD s
Gc K c 1
6
tI s
The transfer function of the actuators dynamics, Gp, on the other
hand, can be approximated as a rst-order transfer function with
dead time G0p as follows:
G0p K p

etd s

tp s 1

289

where Kp is the actuators gain, td is the actuator dead time, and

tp is the actuators time constant.
The estimation of the actuators transfer function G0p will be
needed by the FDI algorithm below when the actuators expected
behavior is calculated and also at the retuning step when a new
set of PID parameters is calculated. The expected actuation level
(denoted by u0a t) will be used as the benchmark upper limit of
how well the control actuators can perform. We note that the
parameters of the PID controller should be tuned in such a way
that the low-level closed-loop response (i.e., actuator under the
PID controller) is fast relative to the sampling time of the MPC
such that the actual actuator output (control action implemented
on the process) is as close as possible to the control action
requested by the MPC at each sampling time. A rigorous analysis
of this problem can be done using singular perturbation techniques for two-time-scale processes.

3. Monitoring and retuning of low-level PID loops

We consider the case where there is no access by the
monitoring system to the measurements of the actual actuation
levels ua(t) implemented by the control actuators on the process.
Therefore, the detection of poor PID tunings must be performed
based on the measurements of the states of the process. To this
end, an FDI method is used as the main tool to extract actuator
behavior from the process state measurements (Chilin et al.,
2010). We use exponentially weighted moving average (EWMA)
residuals to detect and isolate poorly tuned PID loops. Once a
poorly tuned actuator is isolated, a model-based tuning rule such
as CohenCoon or internal model control is applied to the PID
controller that regulates the poorly tuned actuator.
The residuals are constructed from the difference between the
expected behavior and the actual behavior of the plant. This is
done by comparing the evolution of the actual system obtained
from the state measurements against the evolution of the ideal
ltered states based on the plant model. The actual closed-loop
system state (x(t)) evolves in the following manner:
_ f xt Gxtua t wt
xt
ua s

Gp Gc
um s
1 Gp Gc

where um(t) is the control action computed by the MPC and ua(t)
is the actual actuation performed by the actuators. The lter state

(xt),
on the other hand, evolves as follows:
x_ i t f i x^ i t Gi x^ i tu0a t
x^ i x1    xi1 , x i ,xi 1    xnx T
u0a s

G0p G0c
1 G0p G0c

um s

 Dm xNDm ,
xN

8N 0; 1,2, . . .

where Dm is the MPC sampling time, G0p is the estimated transfer

function matrix of the control actuators, and G0c is a well-tuned
PID controller transfer function matrix based on the estimated
model of the actuator G0p . This makes u0a t the expected actuation
level of ua(t).
Using Eqs. (8) and (9), the real-time measurements of x(t) can

be compared against the evolution of xt.
The residual, or the
difference between xi(t) and x i t denoted by ri(t), is expressed in
the following manner:
r i t 9x i txi t9

10

290

A. Leosirikul et al. / Chemical Engineering Science 69 (2012) 287295

In the absence of noise and if G0p Gp , whenever the jth

element of ua deviates from its expected behavior u0aj and the
ith-rowjth-column element of the G(x) matrix is non-zero, the
ith residual (ri) would instantaneously become non-zero. In other
words, ri is non-zero only when there is a problem with the
actuators that directly affect the ith state of the system (relative
degree of 1) (Mhaskar et al., 2006, 2007).
In practice, however, model mismatch, process noise, and
measurement noise are always present to some degree. Therefore,
in a practical setting, the residuals will be non-zero regardless of
the accuracy of the process model used in Eq. (9). Thus, before the
model-based FDI method can be used in practice, the effects of
process and measurement noise levels must rst be recorded
from fault-free closed-loop process operation data (with both the
PID controllers and the MPC being well-tuned). On the basis of
these noisy closed-loop system states, the mean and the standard
deviation of the residuals are calculated and the thresholds are
determined.
Occasional noise spikes can make the residuals exceed the
thresholds for a brief period of time even when the actuators are
functioning well; this can lead to the common problem of false
alarms. To reduce the incidence of false alarms, we dene a
modied residual r E,i ,i 1, . . . ,nx , for each residual ri, calculated at
discrete time instants t k with t k t 0 kDr , k 0; 1,2, . . . and Dr
being the interval between two consecutive state measurements.
The weighted residual is calculated using an exponentially
weighted moving average (EWMA) method as follows (Chilin
et al., 2010, in press):
r E,i t k lr i t k 1lr E,i t k1

11

with r E,i t 0 r i t 0 and the weighting factor l A 0; 1. The parameter l determines the rate at which past data enters into the
calculations of the weighted residual. When l 1, r E,i is equivalent to ri. The typical range of l is between 0.2 and 0.5 depending
on the desired level of sensitivity (Chilin et al., in press; Lucas and
Saccucci, 1990). Lower values of l make the rE(t) curve smoother
as potential noise spikes will have a smaller effect on the overall
shape of the curve; i.e., instances of false alarm will be reduced.
However, in the event where an actual poor tuning occurs, it may
be detected and isolated more slowly.
The threshold, denoted by OE,i , for fault detection is dened as
follows:
r

OE,i mi asi

2l

12

where a is a threshold parameter determining how sensitive the

FDI is; typical value of a is an integer value between 1 and 5. The
parameters mi and si are the mean and the standard deviation of
the ith residual during normal operation. Once r E,i exceeds the
threshold OE,i for a xed amount of time td (determined by the
user), then poor tuning is declared in the actuator(s) directly
affecting the ith state and the retuning algorithm is activated.
Fig. 2 shows the schematic of how the EWMA residuals are used
to activate the PID retuning algorithm at the end of waiting
time td.
Once a poorly tuned actuator is isolated, a PID tuning method
can be applied to the PID controller based on the estimated
transfer function of the actuator G0p . To help ensuring the stability
of the retuning algorithm, we employ a stability constraint.
Specically, whenever retuning is performed, the retuning algorithm makes sure that G0p Gc =1 G0p Gc contains only strictly
negative poles. In this work, we use CohenCoon and internal
model control method to retune the PID parameters to demonstrate the approach. If desired, other model-based tuning rules
may be used as well. See Skogestad (2003), Sung et al. (1998),

Fig. 2. Monitoring scheme of PID response behavior based on the EWMA residuals
of the process state. Poor tuning is declared after r E,i exceeds its threshold OE,i
continuously for t t d .

Veronesi and Visioli (2009), and Zhuang and Atherton (1993) for
other PID tuning methods.
Remark 3. One feature that should be noted is that the PID
retuning will be initiated if the magnitude of the residuals is
above a certain threshold. This means that even if the difference
between uaj(t) and u0aj t is appreciable but the difference between
x i t and xi(t) is smaller than the threshold, the retuner will do
nothing. This is a direct result of the fact that the real value of
ua(t) is unknown and has to be estimated from the trajectories of
the process states. A scenario like this can also happen when Gij 
is small.
Remark 4. The isolability structure of the system is also critical
to the use of the monitoring algorithm proposed here. If, from the
patterns of the residuals, a poorly performing actuator cannot be
isolated with high condence (i.e., two actuators have the same
signature because they directly affect the same system state),
then all control actuators that may be poorly tuned should be
retuned. In principle, it is also possible to use empirical models
from inputoutput data in the MPC design as well as in the
monitoring of the PID control loops. One potential problem of
using this approach is the difculty of isolating which specic PID
control loop is poorly performing because input/output empirical
models cannot account for the coupling between different process
variables the way state-space rst principles models do.
Remark 5. In the design of the lter of Eq. (9), a well-tuned PID
controller, G0c , is assumed to be known and is used to calculate the
benchmark performance of the overall control system. In the case
that G0c is not known, the control action computed by the MPC, um,
can be used directly in the lter design (i.e., replace u0a by um in Eq.
(9)) to obtain an estimate of the expected process state evolution.
Furthermore, once a poorly tuned actuator is isolated, retuning of
the parameters of PID controller used in this actuator should be
carried out to account for changes in operation conditions as well
as control actuator wear and tear over time.
4. Application to a non-linear chemical process network
4.1. Process description and modeling
We demonstrate the PID monitoring and retuning methodology presented in the previous section using a three-vessel
reactorseparator chemical process network. A schematic of the
process is shown in Fig. 3. The rst two vessels are assumed to be
ideal CSTRs, followed by a ash tank separator. There are two
fresh feed streams of pure reactant A of concentration C A10 to both
reactors (with ow rates F10 and F20 respectively) and a recycle
stream (Fr) from the ash tank to the rst reactor. Specically, the
overhead vapor from the ash tank is condensed and recycled to

A. Leosirikul et al. / Chemical Engineering Science 69 (2012) 287295

291

dC B2
F1
F 20

C B1 C B2 
C B2 k1 eE1 =RT 2 C A2
dt
V2
V2

13g

dC C2
F1
F 20

C C1 C C2 
C C2 k2 eE2 =RT 2 C A2
dt
V2
V2

13h

Hvap F r
dT 3
F2
Q3

T 2 T 3 

dt
V3
rC p V 3 rC p V 3

13i

dC A3
F2
Fr

C A2 C A3 
C Ar C A3
dt
V3
V3

13j

dC B3
F2
Fr

C B2 C B3 
C Br C B3
dt
V3
V3

13k

dC C3
F2
Fr

C C2 C C3 
C Cr C C3
dt
V3
V3

13l

Fig. 3. Schematic of the process. Two CSTRs and a ash tank with recycle stream.

Table 1
Process parameter values.
T10 300, T20 300
F10s 5, F20s 5, Fr 1.9
Q1s 0, Q2s 0, Q3s 0
V1 1.0, V2 0.5, V3 1.0
E1 5E4, E2 5.5E4
k1 3E6, k2 3E6
DH1 5E4, DH2 5:3E4
Hvap 5
Cp 0.231
R 8.314
r 1000
aA 2, aB 1, aC 1:5, aD 3
MWA 50, MWB 50, MWC 50

K
m3/h
kJ/h
m3
kJ/kmol
1/h
kJ/kmol
kJ/kmol
kJ/kg K
kJ/kmol K
kg/m3
Unitless
kg/kmol

the rst CSTR, and the bottom product stream is removed. The
efuent of vessel 1 is fed to vessel 2 and the efuent from vessel
2 is fed to the ash tank. Each vessel has an external heat input or
heat removal system (Q 1 , Q 2 and Q3). The steady-state ow rate
and heat input are denoted by F 10s , F 20s , Q 1s ,Q 2s and Q 3s and their
values are given in Table 1. There are two parallel chemical
reactions considered in this process; rst, reactant A is converted
to desired product B and second, A is converted to undesired
product C (referred to as reactions 1 and 2, respectively). Under
standard modeling assumptions, the dynamic energy and material balance equations that can describe this process take the
following form:
dT 1
F 10
Fr
Q1
DH1

T 10 T 1
T 3 T 1

k eE1 =RT 1 C A1
dt
V1
V1
rC p V 1
rC p 1
DH2

k eE2 =RT 1 C A1
rC p 2

C Br
C Cr

aA C A3

13c

dC C1
F 10
Fr

C C1
C Cr C C1 k2 eE2 =RT 1 C A1
dt
V1
V1

13d

dT 2
F1
F 20
Q2
DH1

T 1 T 2
T 20 T 2

k eE1 =RT 2 C A2
dt
V2
V2
rC p V 2
rC p 1
13e

dC A2
F1
F 20

C A1 C A2
C A20 C A2 k1 eE1 RT 2 C A2 k2 eE2 =RT 2 C A2
dt
V2
V2
13f

14a

aB C B3

14b

aC C C3

14c

K aA C A3

dC B1 F 10
Fr

C B1
C Br C B1 k1 eE1 =RT 1 C A1
dt
V1
V1

DH2
k eE2 =RT 2 C A2
rC p 2

C Ar

13a

dC A1
F 10
Fr

C A10 C A1 C Ar C A1 k1 eE1 =RT 1 C A1 k2 eE2 =RT 1 C A1

dt
V1
V1
13b

where T1, T2, and T3 are the temperatures of vessels 1, 2, and 3,

respectively, T10 and T20 are the temperatures of the feed streams
to vessels 1 and 2, respectively, F10 and F20 are the volumetric
feed ow rates into vessels 1 and 2, respectively, and F1 and F2 are
the volumetric ow rates out of vessels 1 and 2, respectively. Fr is
the recycle stream volumetric ow rate from vessel 3 to vessel 1.
V1, V2, and V3 are the volumes of the three vessels, Q1, Q2, and Q3
are the heat inputs into the vessels, CA1, CB1, CC1, CA2, CB2, CC2, CA3,
CB3, and CC3 are the concentrations of A, B, and C in the vessels 1, 2,
and 3, respectively, CAr, CBr, and CCr are the concentrations of A, B,
and C in the recycle stream. r is the mass density of the reacting
uid, Cp is the heat capacity of the reacting uid, k1 and k2 are the
pre-exponential reaction rate constants of reactions 1 and 2,
respectively, E1 and E2 are the activation energies of reactions
1 and 2, respectively, DH1 and DH2 are the enthalpies of reactions
1 and 2, respectively, and Hvap is the heat of vaporization for the
uid in vessel 3. Finally, R is the universal gas constant.
The composition of the ash tank recycle stream is described
by Eq. (14), which assumes constant relative volatility for each
species within the temperature operating range. This assumption
allows calculation of the composition in the recycle stream
relative to the composition of the liquid holdup in the ash tank.
Each tank is assumed to have static holdup and the reactions in
the ash tank are considered negligible. Specically, we have:

MW A

aB C B3

MW B

aC C C3

MW C

aD xD

14d

where aA , aB , aC , and aD are the relative volatility constants of the

three reacting species along with the inert species D. MWA, MWB,
and MWC, are the molecular weights of the three reacting species.
Finally, xD is the mass fraction of the inert species D in the liquid
phase of vessel 3. The values of the process parameters are given
in Table 1.
The system of Eq. (13) is solved numerically using explicit
Euler method with a time step of Dp 0:001 h. Process and sensor
measurement noise are also used in the process simulation. The
sensor measurement noise is generated using a zero-mean
normal distribution with a standard deviation of 2.5 K for the
three temperature state measurements and 1 kmol=m3 for the
nine concentration state measurements. The process noise is
generated similarly and it is included as an additive term in the

292

A. Leosirikul et al. / Chemical Engineering Science 69 (2012) 287295

right-hand-side of the ordinary differential equations of Eq. (13)

with a zero-mean normal distribution and the same standard
deviation values used for the measurement noise. In all three
vessels, the heat inputs are used as the manipulated variables for
controlling the process network at the operating steady-state.
Therefore, the corresponding relative degrees of these variables
with respect to the temperatures of the three vessels (reactor 1,
reactor 2 and separator) are all one, thereby allowing isolation of
poor-tuning in each one of these actuators from process measurements. In addition the second tanks inlet ow rate is chosen as
another manipulated variable. The system has one unstable and
two stable steady states. The operating steady-state is the
unstable steady-state shown in Table 2.
We focus on the problem of monitoring and retuning of the
PID controllers used to regulate the three heat input control
actuators to each of the vessels: Q1, Q2, Q3, at the values computed
by the MPC in each sampling time. In order to calculate the
benchmark performance for each actuator u0a s and a new set of
PID parameters when PID retuning is needed, a rst-order
approximation of the transfer function of the actuator G0p must
be computed. In this example, all actuator dynamics are modeled
with rst-order transfer functions with time delay. All actuators
have the same time constant tp of 2.82 s and time delay td of
3.60 s, resulting in the following transfer function:
e3:60s
Gactuator
2:82s 1

15

The control action computed by the MPC is sent to the control

actuators every Dm 0:01 h. Thus, at every sampling time
t NDm , N 0; 1,2 . . ., the low-level PID controllers take the
MPC command (um(t)) as the set-point and drive the actual
actuation level (ua(t)) to the set-point under the following
closed-loop dynamics:
ua s

Gp Gc
um s
1 Gp Gc

We choose the following parameters for PID monitoring and

retuning. We pick the EWMA parameter l to be 0.2. The EWMA
residual threshold parameter a is chosen to be 5. The waiting time
for fault isolation based on the EWMA residual is set to be
t d 0:01 h.
For the actuators with the transfer function presented in Eq.
(15), the PID parameters that give the best closed-loop response
were found to be the following:

to be all negative. This, in conjunction with the approximate

transfer function G0p of the actuators of Eq. (15), was then used to
approximate the ideal actuation performance u0a s of each
control actuator.
4.2. Simulation results
In the following two examples, we will illustrate how PID
monitoring and retuning are applied to the system.
4.2.1. Example 1
In this example, we start the process from the following initial
condition: x0 0:8xs where xs is the operating steady-state. All
the control actuators are properly tuned with the PID parameters
shown in Eq. (16). At time t 0:45 h, we apply poor tuning to the
PID controller for the actuator Q1 with the following parameters:
K c 0:00909

tI 11:9 s
tD 0:655 s

17

Fig. 4 shows the comparison between the requested actuation

level um(t) and the actual actuation level ua(t) for Q1 if the
monitoring and retuning system is inactive. The EWMA residuals
of the temperature in three vessels are shown in Fig. 5.
With the monitoring system active, Fig. 6 shows the evolution
of PID response ua(t) as it is retuned at t 0:475 h. As shown in
Fig. 7, at t 0:465 h, r E,T 1 starts exceeding its threshold OE,T 1 . At
this point, the value of r E,T 1 starts being monitored closely for
t d 0:01 h. By the time the system reaches t 0:475 h, the value
of r E,T 1 is found to have been above its threshold OE,T 1 for the
entire duration from t 0:465 h to t 0:475 h. Because the
process state T1 is the only state that is directly affected by the
control actuator Q1, given the model-based FDI lter design, any
anomaly detected in r E,T 1 is the result of a problem with the Q1
control actuator. Therefore, the actuator Q1 can be isolated with
high condence as the actuator with poor PID tuning. While other
residuals (r E,T 2 and r E,T 3 ) occasionally exceed their thresholds at
various time instances during the operation, they do not exceed
the thresholds for longer than t d 0:01 h. Thus, the monitoring
system concludes that their values exceed their thresholds simply
because of process and measurement noise.

K nc 0:648

tnI 5:94 s

Control Action for Q1 (105 KJ/hr)

10

16

These parameters were used to calculate G0c . The poles of

calculated with the parameters above are found

5
um(t)

tD 0:54 s
n

10

Table 2
Operating steady-state (xs).
370
3.32
0.17
0.04
435
2.75
0.45
0.11
435
2.88
0.50
0.12

K
kmol/m3
kmol/m3
kmol/m3
K
kmol/m3
kmol/m3
kmol/m3
K
kmol/m3
kmol/m3
kmol/m3

0.2

0.4

0.6

0.8

0.2

0.4
0.6
Time (hr)

0.8

10
5
ua(t)

T1
CA1
CB1
CC1
T2
CA2
CB2
CC2
T3
CA3
CB3
CC3

0
5

G0p G0c =1 G0p G0c

0
5
10

Fig. 4. Example 1: requested actuation level by the MPC (um(t)) and actual
actuation level (ua(t)) when PID retuning is not implemented.

4
3
2
1
0
4
3
2
1
0

rE,T1

0.2

0.4

0.6

0.8

rE,T2

4
3
2
1
0

0.2

0.4

0.6

0.8

rE,T3

rE,T3

rE,T2

rE,T1

A. Leosirikul et al. / Chemical Engineering Science 69 (2012) 287295

0.2

0.4
0.6
Time (hr)

0.8

Fig. 5. Example 1: temperature residuals for the three vessels computed via
EWMA when PID retuning is not implemented. The dashed lines represent the
EWMA residual thresholds OE,i .

Control Action for Q1 (105 KJ/hr)

10
um(t)

4
3
2
1
0

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

0.2

0.4
0.6
Time (hr)

0.8

Fig. 7. Example 1: temperature residuals for the three vessels computed via
EWMA when PID retuning is implemented. The dashed lines represent the EWMA
residual thresholds OE,i .

estimated transfer function of the actuator G0p presented in

Eq. (15), the resulting parameters for the PID of Q1 are as follows:

K c 1:29

tI 6:15 s
0

0.2

0.4

0.6

0.8

10

tD 1:06 s

19

After Q1 is retuned, no more problem can be detected from the

EWMA residuals of T1. In terms of the actual control actuator
performance, after being retuned with CohenCoon method, ua(t)
tracks um(t) quite well; see Fig. 6.

5
ua(t)

4
3
2
1
0

where Kp is the actuators gain, td is the actuator dead time, and

tp is the actuators time constant. With this tuning rule and the

10

4
3
2
1
0

293

0
5
10

0.2

0.4
0.6
Time (hr)

0.8

Fig. 6. Example 1: requested actuation level by the MPC (um(t)) and actual
actuation level (ua(t)) when PID retuning is implemented.

Once the Q1 control actuator is isolated as the poorly tuned

actuator, CohenCoon tuning method is applied to the controller
around Q1 based on the estimated transfer function of the control
actuator G0p . The CohenCoon tuning rule is based on the rstorder-plus-dead-time estimation of the transfer function of the
controlled process. Specically, the CohenCoon tuning rule is as
follows (Cohen and Coon, 1953):


tp 4 td

Kc
18a
K p td 3 4tp

td
tp
tI td
t
13 8 d
tp
32 6

tD td

td
tp

11 2

18b

18c

4.2.2. Example 2
In this example, we will use internal model control tuning rule
(Skogestad, 2003) to tune the PID parameters. We initialize the
process model from the following initial condition: x0 0:8xs
where xs is the operating steady-state. All PID controllers start out
being properly tuned with the parameters presented in Eq. (16).
At time t 0:1 h, a poor PID tuning with the following parameters:
K c 6:48

tI 0:594 s
tD 5:40 s

20

is applied to the PID controller for the control actuator Q3. Fig. 8 shows
that the tuning of the PID controller for Q3 causes ua(t) to oscillate
signicantly. Fig. 9 shows the EWMA residuals of the temperature of
the three vessels when PID retuning is not implemented.
With the monitoring system implemented, Fig. 11 shows that
r E,T 3 is found to start exceeding its threshold OE,T 3 at t 0:206 h.
After waiting for t d 0:01 h, r E,T 3 is found to have been continuously above its threshold until t 0:216 h. Because Q3 is the only
actuator that has relative degree 1 with the process state T3, at
t 0:216 h the monitoring system isolates Q3 and declares that Q3
is poorly tuned. As a result, at t 0:216 h, a set of PID parameters
is calculated via internal model control tuning method based on

294

A. Leosirikul et al. / Chemical Engineering Science 69 (2012) 287295

Control Action for Q3 (105 KJ/hr)

10

5
um(t)

um(t)

5
0
5
10

0.2

0.4

0.6

0.8

10

5
ua(t)

10
ua(t)

5
10

0.2

0.4
0.6
Time (hr)

0.8

4
3
2
1
0

rE,T1

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

0.2

0.4
0.6
Time (hr)

0.8

0.2

0.4

0.6

0.8

0.2

0.4
0.6
Time (hr)

0.8

Fig. 10. Example 2: requested actuation level by the MPC (um(t)) and actual
actuation level (ua(t)) when PID retuning is implemented.

rE,T2

4
3
2
1
0

10

rE,T3

rE,T1
rE,T2

10

Fig. 8. Example 2: requested actuation level by the MPC (um(t)) and actual
actuation level (ua(t)) when PID retuning is not implemented.

rE,T3

0
5

4
3
2
1
0

Control Action for Q3 (105 KJ/hr)

10

4
3
2
1
0
4
3
2
1
0
4
3
2
1
0

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

0.2

0.4
0.6
Time (hr)

0.8

Fig. 9. Example 2: temperature residuals for the three vessels computed via
EWMA when PID retuning is not implemented. The dashed lines represent the
EWMA residual thresholds OE,i .

Fig. 11. Example 2: temperature residuals for the three vessels computed via
EWMA when PID retuning is implemented. The dashed lines represent the EWMA
residual thresholds OE,i .

the estimated transfer function of the control actuator G0p . For a

tuning with fast PID step response, internal model control tuning
rule suggests the following PID parameters for processes that can
be approximated with rst-order-plus-dead-time transfer function (Skogestad, 2003):

Fig. 10 shows the resulting actual actuation level (ua(t)) of Q3.

Though poor PID tuning is applied at t 0:1 h, its effect in terms
of PID response of the control actuator is observed at t 0:185 h
when the step change happens. In terms of detecting this
oscillation pattern from the process sate measurements, this is
detected and isolated at t 0:216 h and the PID parameters of Q3
are retuned.
Notice in Fig. 9 that the magnitude of the residuals of the
directly affected process state (r E,T 3 in this case) is much lower
than r E,T 1 in Example 1 (shown in Fig. 5). This is because the poor
PID tuning problem in this example results in an actuator
oscillation (ua(t)) that oscillates with very high frequency around
the set-point (um(t)). In terms of the process states, this leads to a
smaller overall deviation of the actual process state (x(t)) from the

expected process state (xt).
This is why there is a slightly larger
time lag between the initial time when ua(t) starts deviating from
um(t) and the time when the poor tuning is isolated, compared to
Example 1.

Kc

tp
2K p td

21a

tI mintp ,8td

21b

tD 0

21c

where Kp is the actuators gain, td is the actuator dead time, and

tp is the actuators time constant. This results in the following PID
parameters:
K c 0:392

tI 2:82 s
tD 0 s

22

A. Leosirikul et al. / Chemical Engineering Science 69 (2012) 287295

Remark 6. While the mean and standard deviation of the

residuals are calculated in the presence of process noise under
normal operation at the desired steady-state, the applicability of
the proposed dynamic lter for computing the residuals together
with real-time state variable measurements is not limited to
steady-state operation; the reason is the design of the proposed
dynamic lter which can accurately predict normal evolution of
the process state variables away from the steady-state in the
closed-loop system, thereby leading to the computation of residual values that are valid for process operation away from the
steady-state (note that the initial condition in the example is not
chosen to be the steady-state).

5. Conclusion
In this work, we focused on the problem of monitoring and
retuning of low-level PID control loops used to regulate control
actuators to the values computed by advanced model-based
control systems like MPC. Focusing on the case where the realtime measurement of the actuation level is unavailable, we use
process state measurements and process models to carry out PID
controller monitoring and compute appropriate residuals. Once a
poorly tuned PID controller is detected and isolated, a PID tuning
method based on the estimated transfer function of the control
actuator was applied to retune this PID controller. The proposed
method was applied to a non-linear reactorseparator process
operating under MPC with low-level PID controllers regulating
the control actuators and its performance was successfully
evaluated via extensive simulations.
References
Anderson, K., Blankenship, G., Lebow, L., 1988. A rule-based adaptive PID
controller. In: Proceedings of IEEE Conference on Decision and Control, Austin,
Texas, pp. 564569.
Astrom, K., Hagglund, T., Hang, C., Ho, W., 1993. Automatic tuning and adaptation
for PID controllersa survey. Control Eng. Prac. 1, 699714.
Chand, S., 1992. Self-Monitoring Tuner for Feedback Controller.
Chilin, D., Liu, J., Davis, J.F., Christodes, P.D. Data-based monitoring and reconguration of a distributed model predictive control system. Int. J. Robust
Nonlinear Control, doi:10.1002/rnc.1759. In press.
oz de la Pena, D., Christodes, P.D., Davis, J.F., 2010.
Chilin, D., Liu, J., Mun
Detection, isolation and handling of actuator faults in distributed model
predictive control systems. J. Process Control 20, 10591075.
Christodes, P.D., El-Farra, N.H., 2005. Control of Nonlinear and Hybrid Process
Systems: Designs for Uncertainty, Constraints and Time-Delays. Springer, New
York, NY.
Cohen, G.H., Coon, G.A., 1953. Theoretical consideration of retarded control. ASME
75, 827834.

295

Eriksson, P.-G., Isaksson, A.J., 1994. Some aspects of control loop performance
monitoring. Control Appl. 2, 10291034.
Hang, C., Astrom, K., Ho, W., 1991. Renements of ZeiglerNichols tuning formula.
IEE Proc.D 138, 111118.
Harris, T.J., 1989. Assessment of control loop performance. Can. J. Chem. Eng. 67,
856861.
oz de la Pen
a, D., Christodes, P.D., 2009. Distributed model predictive
Liu, J., Mun
control of nonlinear process systems. AIChE J. 55, 11711184.
oz de la Pen
a, D., Christodes, P.D., 2010. Sequential and
Liu, J., Chen, X., Mun
iterative architectures for distributed model predictive control of nonlinear
process systems. AIChE J. 56, 21372149.
Lucas, J., Saccucci, M., 1990. Exponentially weighted moving average control
schemes. Technometrics 32, 112.
Mayne, D.Q., Rawlings, J.B., Rao, C.V., Scokaert, P.O.M., 2000. Constrained model
predictive control: stability and optimality. Automatica 36, 789814.
Mhaskar, P., Gani, A., El-Farra, N.H., McFall, C., Christodes, P.D., Davis, J.F., 2006.
Integrated fault-detection and fault-tolerant control of process systems. AIChE
J. 52, 21292148.
Mhaskar, P., El-Farra, N.H., Christodes, P.D., 2006. Stabilization of nonlinear
systems with state and control constraints using Lyapunov-based predictive
control. Syst. Control Lett. 55, 650659.
Mhaskar, P., McFall, C., Gani, A., Christodes, P.D., Davis, J.F., 2007. Isolation and
handling of actuator faults in nonlinear systems. Automatica 44, 5362.
Nishikawa, Y., Sannomiya, N., Ohta, T., Tanaka, H., 1984. A method for auto-tuning
of PID control parameters. Automatica 20, 321332.
Qin, S.J., 1998. Control performance monitoringa review and assessment.
Comput. Chem. Eng. 23, 173186.
Qin, S.J., Badgwell, T.A., 2003. A survey of industrial model predictive control
technology. Control Eng. Pract. 11, 733764.
Rugh, W.J., 1991. Analytical framework for gain scheduling. IEEE Control Syst. 11,
7984.
Saito, T., 1990. PID Controller System.
Shi, D., Tsung, F., 2003. Modeling and diagnosis of feedback-controlled process
using dynamic PCA and neural networks. Int. J. Prod. Res. 41, 365379.
Skogestad, S., 2003. Simple analytic rules for model reduction and PID controller
tuning. J. Process Control 13, 291309.
Sung, S.-W., Lee, I.-B., Lee, B.-K., 1998. On-line process identication and automatic
tuning method for PID controllers. Chem. Eng. Sci. 53, 18471859.
Teng, F., Lot, A., Tsoi, A., 2008. Novel fuzzy logic controllers with self-tuning
capability. J. Comput. 3, 916.
Tsung, F., 2000. Statistical monitoring and diagnosis of automatic control processes using dynamic PCA. Int. J. Prod. Res. 38, 625637.
Tsung, F., Shi, J., 1999. Integrated design of run-to-run PID controller and SPC
monitoring for process disturbance rejection. IIE Trans. 31, 517527.
Tsung, F., Shi, J., Wu, C., 1999. Joint monitoring of PID-controlled process. J. Qual.
Technol. 31, 275285.
Veronesi, M., Visioli, A., 2009. Performance assessment and retuning of PID
controllers. Ind. Eng. Chem. Res. 48, 26162623.
Wang, Q., Zhang, Z., Chek, L., Astrom, K., 2009. Guaranteed dominant pole
placement with PID controllers. J. Process Control 19, 349352.
Zayed, A., El-Fallah, A., El-Fandi, M., Hussain, A., 2009. A novel implicit adaptive
pole-placement PID controller. In: Proceedings of the IASTED International
Conference on Modeling and Simulation, vol. 12, Banff, Canada, pp. 296300.
Zhao, Z.-Y., Tomizuka, M., Isaka, S., 1993. Fuzzy gain scheduling of PID controllers.
IEEE Trans. Syst. Man Cybern. 23, 13921398.
Zhuang, M., Atherton, D., 1993. Automatic tuning of optimum PID controllers. IEE
Proc.D 140, 216224.
Ziegler, J.G., Nichols, N., 1942. Optimum settings for automatic controllers. ASME
64, 759768.