Journal of Process Control 51 (2017) 5567

Contents lists available at ScienceDirect

Journal of Process Control

journal homepage: www.elsevier.com/locate/jprocont

A robust control scheme for nonlinear non-isothermal uncertain

jacketed continuous stirred tank reactor
Fahad Wallam a, , Attaullah Y. Memon b
a
Karachi Institute of Power Engineering, Pakistan
b
National University of Science and Technology, Pakistan

a r t i c l e i n f o a b s t r a c t

Article history: A Non-isothermal Jacketed Continuous Stirred Tank Reactor (CSTR) is extensively used in chemical as well
Received 30 November 2015 as in other process industries to manufacture different products. The dynamics of non-isothermal CSTR
Received in revised form 6 August 2016 are highly nonlinear and open-loop unstable in nature. Moreover, it may have parametric uncertainties,
Accepted 7 November 2016
disturbances and un-modeled side reactions which may cause the reactor temperature to deviate from the
reference value. This deviation may degrade quality of the product because the chemical reaction inside
Keywords:
the CSTR depends on reactor temperature. For such a nonlinear, unstable and uncertain process, designing
Non-isothermal
a control scheme with the ability to reject the effects of disturbances along with a good reference tracking
Uncertain
Unstable
capability is a challenging control engineering problem. In this work, a novel robust sliding mode control
CSTR technique named as Improved Integral Sliding Mode Control (IISMC) has been presented for uncertain
Robust non-isothermal jacketed CSTR process. Moreover, a variety of recently developed sliding mode control
CISMC techniques such as Classical Integral Sliding Mode Control (CISMC) and Super Twisted Algorithm based
STA-SMC Sliding Mode Control (STA-SMC) have also been devised and compared with the proposed approach in
IISMC order to investigate the effectiveness of the proposed scheme. A Lyapunov based analysis has also been
HGO provided to assure the robust stability of the closed loop process. Furthermore, in order to extend the
EHGO
state feedback approach to the output feedback scheme, two robust observers; High Gain Observer (HGO)
State Feedback
and Extended High Gain Observer (EHGO), are also designed for the very process. They have also been
Output Feedback
compared with each other and have been investigated for robust stability using Lyapunov based approach.
Finally, an output feedback control scheme using IISMC and EHGO has been presented and its performance
has been examined and compared with the IISMC based state feedback approach. The simulation results
show that the proposed control scheme effectively rejects the uncertainties and disturbances without
leading the process to instability and offers good reference tracking capabilities.
2016 Elsevier Ltd. All rights reserved.

1. Introduction
Continuous Stirred Tank Reactor (CSTR) is one of the widely used
reactors in chemical plants [24,10] as well as in other process indus-
tries. It possesses highly nonlinear with unstable characteristics
due to which the design of high performance control scheme would
become a challenging problem. Moreover, the process model may
have parameters uncertainties, input disturbance and un-modeled
side reactions and nonlinearities due to poor knowledge of the
process, may lead the process to instability. Some industrial appli-
cations such as alkylation of benzene with ethylene process, etc.,
may require robust and fast response control scheme in order to
Corresponding author.
E-mail addresses: fwallam@gmail.com, fwallam@yahoo.com (F. Wallam),
attaullah@pnec.nust.edu.pk (A.Y. Memon).
maintain the quality of the product [10]. Therefore, the problem for
designing the control scheme for uncertain process would become
more interesting for the control engineers.
If the classical control techniques are used to design a control
scheme, it is very difcult to meet the performance requirements
for the control of highly nonlinear process. Indeed, conventional
linear control schemes involve the linearization of the process
model around an operating point, which poses two major limita-
tions: First, it can only predict the local behavior of the process
around an operating point; secondly, the dynamics of linear pro-
cesses is not much rich as compared to dynamics of nonlinear
processes [3].
During past few years, control engineers and researchers have
proposed a variety of nonlinear control schemes. Some of them
are based on differential geometric concepts [13], adaptive lin-
earization [6,8], robust linearization [7] and asymptotically exact

http://dx.doi.org/10.1016/j.jprocont.2016.11.001
0959-1524/ 2016 Elsevier Ltd. All rights reserved.
56 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567
linearization [9,5]. Others are based on sliding mode control (SMC),
adaptive SMC, lyapunov redesign and Nonlinear Model Predic-
tive Control (NMPC). Among them, SMC is one of the widely used
schemes due to its ability to reject disturbances and its insensitiv-
ity to parameter variations [18]. Many researchers have proposed
different combinations of sliding mode control schemes together
with other approaches. One such scheme is a differential geometric
approach with sliding mode scheme which has provided a poten-
tial way of designing a robust control scheme for the control of
uncertain nonlinear process [1,19]. A Dynamical Sliding Mode Con-
trol (DSMC) strategy is another scheme, which is presented in [21]
for the Fliesss generalized observability canonical form based sys-
tem. A robust SMC based feedback linearization technique and
deterministic approach is presented in [22]. A control scheme, by
combining the differential geometric feedback linearization tech-
nique together with SMC and adaptive state feedback techniques,
is presented in [19]. Some researchers combined SMC scheme
with fuzzy logic control scheme, which is known as Fuzzy Slid-
ing Mode Control (FSMC). One such scheme is employed in [23]
where authors have used the sliding surface and rate of change of
sliding surface as inputs to design a fuzzy logic control for chemical
processes.
Recently, an output feedback exact linearization control scheme
has been designed for non-isothermal jacketed CSTR [11]. In this
scheme, nonlinear observers have been designed to estimate the
states as well as disturbances in some parameters of the process.
The performance of such control scheme may degrade if distur-
bance occurs in the parameters that are not estimated by the
observer or if there is an input disturbance or un-modeled side
reaction. More recently, a Terminal Sliding Mode Controls (TSMC)
along with nite time stability observer have been proposed in [10]
for non-isothermal CSTR process neglecting the dynamics of the
cooling jacket.
Physical state variables are required in order to monitor the pro-
cess. Moreover, state feedback control techniques require all states
of the process to generate control signal(s). But, in real practice,
not all the state variables are measured because the instruments
required to measure the state variables are either too expensive
or does not exist. Therefore, an observer is required which esti-
mates the unmeasured states from the available measurement of
the process. The classical linear observer only provides local estima-
tion for state variables of the process having intrinsic nonlinearity.
But, a nite time stable nonlinear observer may provide global
estimation of state variables of highly nonlinear process and may
give the desired performance specication. In addition to compen-
sate for nonlinearity, the observer must converge the estimated
states to the unmeasured states of the process in nite time in
the presence of parametric uncertainties. Different types of non-
linear observers such as Luenberger like Nonlinear Observer (LNO),
Extended Kalman Filter (EKF) and Sliding mode observers (SMO)
are proposed in [11] for the CSTR process.
The aim of present work is to propose a high performance and
simple robust control scheme, Improved Integral Sliding Mode
Control (IISMC), for the regulation of nonlinear non-isothermal
uncertain jacketed continuous stirred tank reactor. In this work,
different existing control techniques which have proven to be
robust such as Classical Integral Sliding Mode Control (CISMC)1 and
recently proposed Super Twisted Algorithm (STA) are also devised
for the very process. These control techniques are compared with
the proposed approach for reference tracking, parameter uncer-
tainties and un-modeled side reaction and input disturbance. The
1
In order to differentiate Integral Sliding Mode Control Scheme (ISMC) presented
in [3] with the proposed ISMC, we named the preceding scheme as CISMC and the
later scheme as IISMC.
robust stability analysis of the closed loop system is also carried
out using Lyapunov stability technique. Moreover, for designing
an output feedback scheme, two nonlinear robust observers; High
Gain Observer (HGO) and Extended High Gain Observer (EHGO) are
also devised, compared and investigated. The stability analyses of
these observers are also presented.
The rest of the paper is organized as follows: In Section 2, a
brief introduction of the CSTR process is presented. In Section 3,
robust control techniques, CISMC, STA-SMC and IISMC are devised
for the uncertain CSTR process. The performance of the controllers
is also discussed in the same section. The robust observers, HGO
and EHGO, are designed and discussed in Section 4. In Section 5,
output feedback control scheme is developed by combining IISMC
and EHGO for the CSTR process. Performance comparison between
the IISMC based state feedback control technique and IISMC and
EHGO based output feedback control scheme is also carried out in
the same section. Finally, in Section 6, conclusion of the work is
presented.
2. Non-isothermal jacketed continuous stirred tank reactor
(CSTR)
We have considered a non-isothermal jacketed continuous
stirred tank reactor problem in which exothermic irreversible rst
order reaction takes place. This problem has been extensively used
in the area of control research due its highly nonlinear behavior.
Before discussing the process model dynamics, consider the fol-
lowing assumptions:
Assumption 1. The temperature inside the reactor is uniformly
distributed by assuming perfect mixing in the reactor.
Assumption 2. A constant volume inside the reactor is assumed.
By considering the above mentioned assumptions, following
dynamic equations, described the physical model of the reactor,
are derived [12,11].
x = f (x) + g (x) qc + f (x) (1)
where,
 
q x1f x1 x1  (x2 )
 
f (x) = q x1f x2 (x2 x3 ) + x1  (x2 ) (2)
1 2 (x2 x3 )
0
g (x) = 0 (3)
 
1 x3f x3
f1 (x)
f (x) = 0 (4)
0
where, f(x) and g(x) are locally Lipschitz in x, state variable x1 , x2
and x3 is the dimensionless reactant concentration, reactor tem-
perature and cooling jacket temperature, respectively, q denotes
the reactor feed ow rate, x1f , x2f and x3f is the dimensionless reac-
tor feed concentration, reactor feed temperature and cooling-jacket
feed temperature, respectively, is the dimensionless heat of reac-
tion, is the dimensionless heat transfer coefcient, 1 is the reactor
to cooling-jacket volume ratio, 2 is the reactor to cooling-jacket
density heat capacity ratio,  is the nominal Damkohler number
based on the reactant feed, the controlled variable qc denotes the
cooling jacket ow rate ranges from 0 to 1 [11], f1 denotes the un-
 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567 57

modeled rst order side reaction from x1 and therefore it is assumed where, r is the reference signal. The process dynamics (8)(10) may
that f1 is only dependent on x1 [20] such that be rewritten in error form as
 
f1 = x1 (5)  = p (, e1 ) = q x1f   (e1 )  (12)

e 1 = e2 (13)
where,  is an unknown positive constant.  is the dimensionless
Arrhenius reaction rate nonlinearity and is dened as     

 e 2 = q + z2 + 1 qc x3f x3 + 2 (z1 x3 )
x2  
1+ 1 x
 (x2 ) = e 2 (6) 
+ (e1 )  +   z2 r (14)
1 +  1 z1
where,  is the dimensionless activation energy. The Arrhenius
rate expression () makes the process dynamics nonlinear and where,  is dened as
unstable which poses operational and control problems. The CSTR  
e1 +r
considered here acts as an unstable open loop when the reactor
1+ 1 (e1 +r )
dimensionless temperature is between 1.5 and 3.0. Due to eco-  (e1 ) = e (15)
nomic constraints, the industries are interested to operate the The steady state value r for the internal dynamics  may be
reactor inside this region. It may be veried that following property computed as
holds for the CSTR process model.  
q x1f r r  (0) r = 0 (16a)
Property 1. The model described by (1)(4) has a relative degree
2. qx1f
r = (16b)
q +  (0) + 
Remark 1. The state variables x2 and x3 are both measureable
In order to achieve stable tracking, the internal dynamics 
[11].
should remain stable throughout the region of interest [19]. The
internal dynamics  remains stable if the zero dynamics
3. Design of robust nonlinear controllers
 = p (, 0) (17)
The design objective of robust nonlinear control scheme is to is stable over the region of interest [3]. In order to ensure the sta-
control the temperature inside the reactor to follow a desired bility of internal dynamics following lemma has been constructed.
trajectory as well as to reject the effects of input disturbances,
un-modeled side reaction and parametric uncertainties. Different Lemma 1. The internal dynamics  of the process (12)(14)
nonlinear robust controllers are reported in the literature. How- remains stable if the following holds
ever, we consider only two robust control algorithms; CISMC and
STA-SMC, which are robust, fast and efcient. Moreover, we pro- q +  (0) +  > 0
pose a novel nonlinear robust control scheme, Improved Integral
Sliding Mode Control (IISMC), for the nonlinear uncertain process Proof. Let,
and compare it with the control techniques mentioned earlier.
=  r
 (18)
In this section, rst, we develop existing nonlinear robust con-
trol schemes that is, CISMC and STA-SMC, for the CSTR process.
= p (, e1 ) = qx1f 
 [q +  (e1 ) + ]
Next, we propose a new robust control scheme for the same pro-
cess. Later on, we present performance analysis among different r [q +  (e1 ) + ] (19)
control schemes considered in this section.
To design nonlinear robust controllers, the process dened in
(1)(4) needs to be transformed in normal form. Choosing the coor- The zero dynamics of (19) is given by
dinate transformation as
= p (, 0) = 
 [q +  (0) + ] (20)
 T
T1 =  z1 z2 Let, the Lyapunov candidate be selected as
   (7)
T1 = x1 x2 q x1f x2 (x2 x3 ) + x1  1 2
=
V0 () 
(21)
2
We get,

  V 0 ()
=
 (22)
 = q x1f    (8)
By using (20), we get,
z 1 = z2 (9)  
V 0 ()
=

[q +  (0) + ] (23)
    
z 2 = q + z2 + 1 qc x3f x3 + 2 (z1 x3 ) V 0 () 2 [q +  (0) + ]
=  (24)
 

V 0  0q +  (0) +  0 (25)

+  +   z2 (10)
1 +  1 z1
The process model may contain parametric uncertainties and
This form decomposes the system into an internal part  and an input disturbance. Therefore, the whole process is now transformed
external part z [3]. Now, let, into perturbed parameter form in order to simplify the design of
control schemes. In this form, the parameters of the process model
e1 = z1 r (11a) is decomposed into its nominal term and uncertain term as
e2 = z2 r (11b) + 
=
58 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567

Table 1 where, F2 and G2 is dened in (28c) and (29), respectively and 

CSTR model parameters nominal values.
be dened as
Parameter Value

s
v = 1 sat (34)
 0.072
8.0 where, 1 is the gain, is a small positive constant and sat(.) is
0.3
dened as [3]

q
20
1.0

y, |y| 1
1 10 sat (y) = (35)
2 1.0
sign (y) , |y| > 1
x 1f 1.0
where,
x 2f 0.0

1.0
x 3f
1, y > 0
sign (y) = 0, y = 0 (36)

where, is the nominal parameter of the model and  is the 1, y < 0
uncertainty. Also, = q, x1f , , , , 1 , 2 , , x3f . The nominal val- Theorem 1. Suppose that the uncertain process dened in
ues of the parameters are given in Table 1 [12,11]. Similarly, the (26)(30), satisfy lemma 1 is subjected to the control designed in
input to the process may be written as (33)(36). The closed loop system is stable if the following condition
qc = u + u holds

where, u is the input disturbance term. Now, the process model 1 > 
(12)(14) may be written in perturbed parameter form as
  Proof. See Appendix A.
 + w1
 = q x 1f   (26)
Small values of 1 will slow down the controller response and
e = F (.) + G (.) u +  (27)
large values of 1 will make the controller more reactive which
where, results in chattering. For better accuracy, should be chosen as
  small as possible. But, too small value of will induce chattering
F1 because as tends to 0, sat(s/ ) tends to sign(s). So, there is a trade-
F (.) = (28a)
F2 off between the amount of chattering and control precision [3,19].

F1 = e2 (28b) 3.2. Super twisted algorithm-sliding mode control (STA-SMC)


F2 = q + z2 + 2 1 2 (z1 x3 ) +
  The STA-SMC retains the robustness of rst order SMC while
removes chattering introduced by the rst order SMC. Higher order
  SMC also removes chattering while preserving the robustness of
  
q x 1f  +
    z2 r (28c) rst order SMC but it requires more than one time-derivative of
1 +  1 z1 sliding surface for example, an rth order SMC requires up to r-
s , . . ., sr1 (where, s is
1 time-derivative of sliding surface, i.e. s,
    the sliding surface). But, STA-SMC does not require higher time-
G1 0
G (.) = =   (29) derivatives [13]. In order to design STA-SMC based control law, let
G2 1 x 3f x3 the sliding surface be
    s = a1 e1 + e2 (37)
1 0
 = = (30) where, a1 is a positive constant. The control law for the uncertain
2 
w3 + w
1
process may be devised as
where, w1 contains the parametric uncertainty and un-modeled 1  
side reaction and likewise, w3 contains the parametric uncertain- u= F2 a1 e2 + v1 (38)
G2
ties and input disturbances.
where, 1 be dened as [14]
3.1. Classical integral sliding mode control (CISMC) v1 = k1 |s1 |1/2 sign (s) + v2 (39)

In this section, we present the design of CISMC for the uncertain and 2 be dened as
process dened in (26)(30). The sliding surface for the design of v 2 = k2 sign (s) (40)
CISMC may be selected as [3]
Theorem 2. Suppose that the uncertain process dened in
s = eI + a1 e1 + e2 (31) (26)(30), satisfy lemma 1, is subjected to the control designed in
where, a1 is positive constant and eI is dened as: (38)(40). Assume the following bound on 

e I = a0 e1 (32) || |1 | + |2 |

where, a0 is another positive constant. The control law, with the where,
sliding surface (31), for the uncertain process (26)(30) may be
computed as |1 | 1 |s|1/2
1
u= [F2 a0 e1 a1 e2 + v] (33) 2 | 2
G2 |
 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567 59

Table 2
where, 1 and 2 are some positive constants. The closed loop
Controllers parameters.
system is stable if the following conditions hold
Controller Parameters
k1 > 21 CISMC a0 = 0.02, a1 = 5, 1 = 20 and = 0.1
STA-SMC a1 = 200, k1 = 5 and k2 = 0.05
 2 IISMC a1 = 200, 1 = 0.01, 2 = 0.01, k1 = 2, = 0.01 and k2 = 10
2k12 + 42 + k1 1 + 41 k13 + 8k12 2 4k14
k2 >
8 (k1 21 )

s
Proof. See Appendix B. k1 tanh  + k2 s2 > 

3.3. Improved integral sliding mode control (IISMC)
Proof. Let, the Lyapunov candidate be selected as
In this section, we propose a novel robust algorithm which is V = V0 + V1 (49)
fast, efcient and robust. The proposed algorithm is an improved
version of the CISMC presented in Section 3.1. The sliding surface where, V0 is dened in (21) and V1 is dened as
for the IISMC is dened as 1 2
V1 = s (50)
s = eI + a1 e1 + e2 (41) 2
By taking the time derivative of the Lyapunov candidate dened
where, a1 is a positive constant and eI is dened as in (50), we get

e  
1
e I = 1 tanh + 2 |e1 |sign (e1 ) (42) V 1 = ss (51)

Take the time derivative of (41) and put into (51), we get,
where, 1 , 2 and are also some positive constants. The proposed
control law is V 1 = s [e I + a1 e2 + F2 + G2 u + ] (52)
1  
u= F2 e I a1 e2 + v (43) By combining (42), (43), (44) and (52), we get,
G2
s
V 1 = k1 s tanh k2 s3 sign (s) + s (53)
where,  be dened as

s 
s 
v = k1 tanh k2 s2 sign (s) (44) V 1 s k1  tanh  + k2 s2  (54)


s
k1 and k2 are another positive constants. The Eqs. (42) and (44)
V 1 < 0k1  tanh  + k2 s2 >  (55)
may respectively be written as

e I = k3 sign (e1 ) (45) By combining (25) and (55), we conclude that

v = k4 sign (s) (46) V < 0

where, k3 and k4 are dened as

e  
1
k3 = 1 | tanh | + 2 |e1 | (47) 3.4. Performance analysis


s
In this section, performance analysis of the controllers pre-
k4 = k1 | tanh | + k2 s2 (48)
sented in Sections 3.13.3 is carried out. For now, we assume that
all the states (x1 , x2 and x3 ) of the process are available to con-
The gain terms k3 and k4 changes according to the magnitude of
trollers.The states z1 , z2 and are found by using (7) where nominal
error and sliding surface, respectively. If the error is large, value of
parameters value (dened in Table 1) have been used in coordinate
the gain k3 will be high and if the error is small, value of the gain k3
transformation. Also, following parametric uncertainties and input
will be low. Similarly, if value of the sliding surface is large, value of
disturbance are assumed in the process model for analyzing the
the gain k4 will be high and if value of the sliding surface is small,
performance of controllers.
value of the gain k4 will be low. The Eqs. (45) to (46) show that IISMC
has advantage of both pure switching function and gain adaptation 0 0.25 (56)
in comparison with CISMC and STA-SMC. Due to gain adaptation,
the stability of IISMC is guaranteed. In contrast, CISMC and STA- 0.4  0.4 (57a)
SMC require a pre-specied gain for assuring stability of the closed 2  2 (57b)
loop system. Also, IISMC uses signum function, which enhances the
robustness of SMC in comparison with CISMC. It may conclude that 0.05 q 0.05 (52c)
IISMC is an adaptive gain rst order switching (signum) function 2 1 2 (57d)
based SMC. If sign(e1 ) and adaptive gain term k3 in (45) is replaced
respectively with e1 and constant gain a0 and sign(s) and adaptive 0.05 2 0.05 (57e)
gain k4 in (46) is replaced respectively with sat(s) and constant gain
0.05 x3f 0.05 (57f)
1 , the IISMC will simply become CISMC.
0 u 0.3 (58)
Theorem 3. Suppose that the uncertain process dened in
(26)(30), satisfy lemma 1, is subjected to the control law pro- For the uncertainties assumed in (56)(58), parameters of the
posed in (42)(44). The closed loop system is stable if the following controllers are given in Table 2. These parameters are selected such
condition holds that the closed loop system with the respective controllers remains
60 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567
stable for the uncertainties dened in (56)(58) as well as satis-
es other performance indicators (such as chattering, undershoots,
etc.).
Remark 2. It may be veried that lemma 1 holds for the uncer-
tainties and disturbances assumed in (56)(58).
Following four cases are evaluated for comparing the perfor-
mance among the controllers.
3.4.1. Reference tracking
In this case, the simulation is performed for changing the set-
point in step manner. Initially, it is assumed that the reactor is
operating at the dimensionless temperature (x2 ) of 1.5. At dimen-
sionless time of 20, the demand is set to new value of 2.0 as shown
in Fig. 1(a). All the controllers considered in this work, sets the
reactor dimensionless temperature to new value without being
instable as shown in Fig. 1(a). It may note that the IISMC response
is fast as compare to other controllers but it produces a minor over-
shoot while settling the reactor temperature to a new value. It may
also note that the STA-SMC produces almost no overshoot but its
settling time is less than IISMC.
Fig. 1(d). All the controllers reject the input disturbance as reveal
from Fig. 1(d). The reactor temperature goes to a minimum value
of 1.4986 (change of 0.093%) in the case of CISMC. The STA-SMC
and IISMC rejects the disturbance signal with a minimum shoot of
1.4997 (change of 0.02%) and 1.49996 (change of 0.0027%), respec-
tively.
This case study shows that the proposed scheme IISMC has good
capability of rejecting the input disturbance signal.
4. Design of robust observers
This section is concerned with the design of robust nonlin-
ear observers for the uncertain CSTR process. In particular, two
nonlinear robust observers have been proposed which can han-
dle the process uncertainty in decent manner. Moreover, their
performance analysis has been provided in order to show the effec-
tiveness of the presented estimation techniques.
We rst develop the preliminary work for designing the robust
observers for uncertain CSTR process. In order to take the advantage
of utilizing both measureable state variables x2 and x3 , we select the
coordinate transformation variables as
The above mentioned scenario shows that the proposed control 
T
technique IISMC efciently tracks the reference signal. However, a T2 = z1 z2
negligible over-shoot is observed in chasing the reference signal but
the settling time is shortest as compare to other control schemes   

T2 = x2 q x1f x2 (x2 x3 ) + x1  (59)
considered in this work.
The system (1) in new coordinates may be written as,
3.4.2. Parametric uncertainties
In this scenario, robustness of the controllers is tested by z 1 = z2 (60)(61)
introducing the uncertainties in the process parameters. For simu-
  

x 3 = 1 qc x3f x3 + 2 (x2 x3 ) (62)
lation, it is assumed that  = 0.4,  = 2.0, q = 0.05, 1 = 2.0,
2 = 0.05 and x3f = 0.05. All the controllers considered here where, (.) is locally Lipschitz in its argument and x1 is given by
reject the parameters disturbances as shown in Fig. 1(b). It may note  
that, in the case of STA-SMC, the reactor temperature goes to a min- z2 q x2f z1 + (z1 x3 )
x1 = (63)
imum of 1.3 (change of 13.33%) before settling down to a reference 
value of 1.5. It may also be observed that the IISMC takes the reactor
temperature to a minimum of 1.48 (change of 1.33%) while CISMC 4.1. High gain observer (HGO)
takes the reactor temperature to a minimum of 1.465 (change of
2.33%) before settling down to 1.5. It is also worth mentioning that The HGO is a robust nonlinear observer that has ability to dimin-
the IISMC settles the reactor temperature to its reference value in ish the perturbation effects in estimating the states of the uncertain
the shortest time as compare to other controllers. process by introducing sufciently high gain terms. The HGO for the
This scenario shows that the proposed scheme IISMC has para- process dened in (60)(62) may be designed as
1  
metric uncertainties rejection capabilities better than other robust

control schemes presented in this work. Thus the proposed scheme z 1 = z2 + z1 z1 (64)

is not only fast but it is robust also.
  2  
z 2 = 0 z1 , z2 , x3 , qc + 2 z1 z1 (65)
3.4.3. Un-modeled side reaction
In this case, the performance of controllers is analyzed for un- where, 1 and 2 are the gain terms, is a small positive constant
modeled dynamics of the process. It is assumed that there is an and  0 (.) is dened as
un-modeled side reaction which is neglected during modeling the
process. For simulation study, suppose that  = 0.25. Fig. 1(c) reveals       

0 (z1 , z2 , x3 , qc ) = q + z2 + 1 qc x 3f x3 + 2 z1 x3
that all the controllers reject the effects of un-modeled side reaction
and settles down to a reference value after a transient. The CISMC  
  x 1
bring the reactor temperature back to a reference value of 1.5 from 
+ k q x 1f x 1 x1 k +
   z2 (66)
the 1.4996 (change of 0.027%) while the STA-SMC and IISMC settles 1 +  1 z1
the reactor temperature back to 1.5 after a peak transient of 1.49996
(change of 0.0027%) and 1.499985 (change of 0.001%), respectively. where,  0 (.) is locally Lipschitz and x 1 be dened as
A small chattering is observed in the case of STA-SMC.  
z2 q x 2f z1 + (z1 x3 )
3.4.4. Input disturbance x 1 = (67)


This last case study, further investigate the robustness of the
controllers. In this scenario, performance of the controllers is tested Let,
for disturbance appears at the input to the process. For simula- z1 z1
tion study, it is assumed that a disturbance of the magnitude of 1 = (68)

0.3 (u = 0.3) is added to the input signal at time t = 20. The tran-
sient behavior of reactor dimensionless temperature is shown in 2 = z2 z2 (69)
 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567 61

(a) 2.1
Reference
CISMC
STA-SMC
2 IISMC

1.9
2.015

2.01
1.8
2.005

2
2

x
1.7 1.995

1.99
21.2 21.4 21.6 21.8 22 22.2 22.4 22.6 22.8

1.6

1.5

1.4
0 5 10 15 20 25 30 35 40 45 50
Time

(b) 2
Reference
1.51 CISMC
1.9 STA-SMC
1.5
IISMC
1.49
1.8
1.48

1.7 1.47

1.46
1.6

-2 -1 0 1
10 10 10 10
1.5
2
x

1.4

1.3

1.2

1.1

1
0 100 200 300 400 500 600 700 800 900 1000
Time

(c) Reference
1.5002 CISMC
STA-SMC
IISMC

1.5

1.4998

1.50001
x2

1.5
1.4996
1.49999

1.49998

1.4994
1.49997

1.49996

1.49995
1.4992 -2 -1 0 1
10 10 10 10

1.499
0 100 200 300 400 500 600 700 800 900 1000
Time

(d) 1.5015
Reference
CISMC
1.5 STA-SMC
1.501 1.49999 IISMC

1.49998

1.49997
1.5005
1.49996

1.49995
18 20 22 24 26 28

1.5
2
x

1.4995

1.5
1.499
1.4999

1.4998
1.4985
1.4997

16 18 20 22 24 26 28 30
1.498
0 50 100 150 200 250 300 350 400 450 500
Time

Fig. 1. (a) Reference tracking. (b) Parametric uncertainties. (c) Un-modeled side reaction. (d) Input disturbance.
62 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567

By differentiating (68) and (69), we get, uncertainties and disturbances [16,17]. The EHGO for the process
dened in (60)(62) may be designed as
 1 = 1 1 + 2 (70)
1  
 2 = 2 1 + ( 0 ) = 2 1 +  (71) z 1 = z2 + z1 z1 (80)

It may be written as   2  
z 2 = 0 z1 , z2 , x3 , qc + 2 z1 z1 +  (81)

 =  + B (72)
3  
 = 3 z1 z1 (82)
where,
 
1 1 where,  0 is dened in (66) and 3 is also a gain term. Only dif-
= (73a) ference between the HGO and EHGO is having an additional state
2 0  in EHGO. Theorem 4 is also valid in the case of EHGO with the
  following bound on uncertain term.
0
B= (73b) L

1
 is dened here as
It may be noted that reducing diminishes the effects of  in

(72). Assume ([4]), z1 z1
1 2
 L + M (73c)

 = 2 =
z2 z2

(83)
where, where, L and M are some positive constants. Let, the Lya-
punov candidate be selected as 3

V =  T P (74)
Differentiating (83) with respect to time, we get the same Eq.
where, P is a positive denite symmetric matrix and is computed that is,
(72) except  is replaced by ,
by using
 =  + B
(84)
P + T P = I (75)
where,
where, I is an identity matrix of order 2. Differentiating (74) with
1 1 0
respect to time and by using (72), we get,


T 
 = 2 0 1 (85a)
T
V =  + B P +  P  + B (76)
3 0 0
V =  T  + 2BP (77)
0
By using (73c), we get, B = 0 (85b)

1
V  (1 2LBP)  2MBP (78)
2MBP
V 0
(79)
1 2LBP
The steady state error decreases if reduces. So, should be
selected small for achieving higher accuracy. The above result may
be summarized by using the following theorem.
Theorem 4. Suppose that (0, ), every trajectory of the closed
loop system started inside a given compact set is bounded if the
following condition holds
2MBP
 >
1 2LBP
A major limitation of HGO is an occurring of peaking phe-
nomenon for small values of which may lead the process to
instability. During peaking phenomenon, a peak of the order of
O(1/) is observed which may cause the feedback controller to
become unstable. One of the solutions for this problem is to saturate
the controller outside the compact set of interest [15,17].
4.2. Extended high gain observer (EHGO)
EHGO estimates the unmeasured states along with the model
uncertainties and disturbances. As the name implies, EHGO is an
extension of the HGO. EHGO is primarily an HGO with an additional
state augmented to it. This additional state improves the estima-
tion performance by providing a lumped parameter that contains
the missing information of the process dynamics that is, the model
Just like HGO, EHGO has the same limitation of peaking phe-
nomenon which may be resolved by saturating the controller.
4.3. Performance analysis
In this section, the performance of the observers designed in Sec-
tions 4.1 and 4.2 is analyzed. The observers presented in (64)(65)
and (80)(82) are used to predict the un-measureable state of
the process. In order to test the robustness of the observers, sup-
pose that  = 0.4,  = 2.0, q = 0.05, 1 = 2.0, 2 = 0.05 and
x3f = 0.05. The HGO parameters are set as 1 = 13.7, 2 = 33.09
and = 0.1 and EHGO gains are set as 1 = 10.7, 2 = 33.99, 3 = 60.0
and = 0.1. The error dynamics  1 (1 = z1 z1 ) and  2 (2 = z2 z2 )
of the observers are plotted in Fig. 2(a)(b). As shown in Fig. 2(a),  1
of HGO initially attains a value of approximately 5.5 105 before
setting down to a value of approximately 5.62 105 whereas  1
of EHGO approaches to a peak value of 4 105 and then it set-
tles down to value of 0. It may be interesting to observe that there
remains a steady state error in the case of HGO but in the case
of EHGO, no steady state error is observed. Similarly, as shown
in Fig. 2(b), there remains a steady state error of approximately
7.6 103 in  2 for HGO but no steady error is observed in the case
of EHGO. This shows that EHGO offers more robustness than HGO.
5. IISMC and EHGO based output feedback control
From Section 3, it may conclude that the performance of the pro-
posed control technique, IISMC, is better than CISMC and STA-SMC
 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567 63

(a)

(b)

Fig. 2. (a) HGO and EHGO error 1. (b) HGO and EHGO error 2 .

 
and from Section 4, we conclude that EHGO is better in performance s 
comparison with HGO. In this section, robust output feedback con- v = k1 tanh k2 s2 sign s (88)

trol technique is proposed by utilizing IISMC and EHGO as both are
the best in comparison with other recently developed techniques where,
discussed in this article.
s = e I + a1 e 1 + e 2 (89a)
For the CSTR process, the EHGO is designed in (80)(82). Based
on EHGO, the IISMC dened in (42)(44) may now be modied as e 1 = z1 r (89b)

   e 2 = z2 r (89c)
1
u = sat (u)
= sat F2 e I a1 e 2 + v  (86)
  

G2 F2 = q + z2 + 2 1 2 z1 x3 +
 

where, sat(.) is a bounded continuously differentiable function such

 
  
= u in the domain of interest and |sat (u)|
that sat(u) 1 everywhere  +
q x 1f     z2 r (89d)
[16]. The other variables in (86) are dened below. 1 +  1 z1

  
e 1   Now, the closed loop system can be shown by
e I = 1 tanh + 2 |e1 |sign e 1 (87)
z 1 = z2 (90a)
64 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567

z 2 =  (z1 , z2 , x3 , qc ) (90b) 5.1.4. Input disturbance

  
In this case study, an input disturbance of magnitude same
x 3 = 1 qc x3f x3 + 2 (x2 x3 )

qc = u z1 , z2 , 3 , x3 , v , r, r,
r + u
 =  + B
where, (z1 ,z2 ,x3 ,qc ) is dened in (61), u(.) is dened in (86),
is dened in (83) and  and B is
dened in (85a)  and
(85b), respectively. We can write u z1 , z2 , z3 , x3 , v , r, r,
r =

u z1 2 1 , z2 2 , 3 , x3 , v , r, r,
r .
Note that when = 0 in (90e),  = 0. The closed loop system
(90a)(90e) is a simple state feedback for = 0. Thus, by reduc-
ing , the output feedback system gets closer to the state feedback
system [25]. Furthermore, it may be shown that the closed loop
system formed by (90a)(90e) is stable by following the procedure
mentioned in [16,17].
5.1. Performance analysis
In this section, performance analysis of the output feedback
controller is carried out. Moreover, the performance of the output
feedback is compared to that of the state feedback technique pre-
sented in Section 3.3. The same four scenarios, which are evaluated
in Section 3.3, are again considered here.
5.1.1. Reference tracking
In this case, it is assumed that reactor is operating at the dimen-
sionless temperature (x2 ) of 1.5. At time t = 20, the demand is set
to new value of 2.0 as shown in Fig. 3(a). It may be interesting to
observe that the performance of the State Feedback (SFB) control
and the Output Feedback (OFB) control techniques is almost simi-
lar. This shows that the performance of the state feedback control
is recovered by the output feedback control scheme.
5.1.2. Parametric uncertainties
In this scenario, the behavior of output feedback controller in
contrast to state feedback controller is evaluated. It is assumed that
reactor is operating at the temperature of 1.5. At time t = 20, step
uncertainties are introduced in the parameters of the process. The
magnitude of uncertainties in the parameters is considered here
same as that in Section 3.4.2. The transient behavior is plotted in
Fig. 3(b). It may note that both the schemes (output feedback and
state feedback) approaches to a peak value of about 1.48 before
settling down to 1.5. It may be remarkable to state that the out-
put feedback scheme settles down the reactor temperature very
promptly as compare to the state feedback technique. This is due
to the fact that the high gain terms and additional state () aids the
EHGO to rapidly reject parametric disturbances.
5.1.3. Un-modeled side reaction
In this scenario, the performance of the output feedback and the
state feedback controllers are compared for un-modeled dynamics
of the process. We suppose the same un-modeled dynamics of the
process as considered in Section 3.4.3. The transient plots of the
controllers are shown in Fig. 3(c). The output feedback scheme and
state feedback technique, respectively reaches to a peak value of
approximately 1.4997 (change of 0.02%) and 1.499985 (change of
0.001%). It may be notable in Fig. 3(c) that the output feedback
control scheme settles quickly as compare to state feedback scheme
but has some minor overshoot. It shows that the EHGO improves
the settling time of the IISMC control technique.
(90c) as mentioned in Section 3.4.4 is injected at time t = 20 to test
and compare performance of the output feedback scheme with
(90d) state feedback scheme. It is worth mentioning that performance
of both controllers is similar for input disturbance as shown in
(90e) Fig. 3(d). Again, the simulation study reveals that the output
feedback scheme recovers the performance of the state feedback
scheme.
6. Conclusions
The uncertain non-isothermal jacketed CSTR possesses expo-
nential nonlinearity and unstable dynamics. In CSTR, the chemical
reaction depends on the temperature of the reactor. If the temper-
ature of the reactor deviates from the reference temperature, the
quality of the product may degrade. The process may have paramet-
ric uncertainties, disturbances and un-modeled dynamics which
may cause the reactors temperature to change from the reference
value if the control scheme does not have the robustness feature.
Although, the linear control schemes is simple and easy to design
but they offer only local stability and may not provide the desired
performance characteristics. In this work, a simple, novel, robust
and easy to implementable nonlinear control technique, IISMC,
is proposed for the CSTR process. Moreover, two more nonlinear
robust control schemes, CISMC and STA-SMC, are also presented.
The performance of these control schemes are evaluated by sim-
ulating with four different transient scenarios. The results show
that
IISMC and CISMC track the reference signal with minor over-
shoot whereas STA-SMC tracks the reference with almost no over-
shoot. But, IISMC quickly settle to a new value as compare to CISMC
and STA-SMC.
1 The IISMC offers smallest under-shoot and shortest settling time
while largest undershoot and longest settling time is observed
in the case of STA-SMC for process parametric uncertainties sce-
nario.
2 In the scenario of un-modeled side reaction, the performance of
all controllers is almost same but a slight chattering is observed
in the case of STA-SMC.
3 The IISMC offers excellent settling time as compare to CISMC
and STA-SMC when the controllers are subjected to input dis-
turbance. The CISMC has worst settling time.
These results show that the performance of IISMC is consis-
tent in all the cases considered in this work. Moreover, two robust
observers, HGO and EHGO are also devised in order to estimate the
unmeasured state of the process. The performance analysis of these
two observers shows that EHGO has more robust capabilities than
HGO as no steady state error has been observed in the case of EHGO
where as in the case of HGO, a small steady state error remains.
By inspiring the capabilities of IISMC and EHGO, an output
feedback control scheme is also suggested for the CSTR process.
Simulation study shows that the performance of the output feed-
back control scheme is similar to that of the state feedback control
scheme for reference tracking and input disturbance scenarios. In
the case of parametric uncertainties and un-modeled side reaction,
the output feedback control scheme quickly settles the output in
contrast to state feedback technique and hence, the improvement
in the settling time is observed in the case of output feedback
scheme. This improvement is due to the fact that EHGO assists
IISMC to reject model intrinsic disturbances quickly by using high
gain terms and additional lumped disturbance estimation state.
 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567 65

(a) 2.1
Reference
SFB-IISMC
OFB-IISMC
2

1.9 2.01

2.005

2
1.8

1.995

2
x
1.99
1.7
1.985

1.98
1.6 20.5 21 21.5 22 22.5 23 23.5 24 24.5 25 25.5

1.5

1.4
0 10 20 30 40 50 60 70 80 90 100
Time

(b) 1.52
Reference
SFB-IISMC
OFB-IISMC
1.51

1.5

1.49

1.5
2
x

1.48 1.495

1.49
1.47
1.485

1.46 18 18.5 19 19.5 20 20.5 21 21.5 22

1.45
0 50 100 150 200 250 300 350 400 450 500
Time

(c) Reference
1.5001 SFB-IISMC
OFB-IISMC

1.5

1.500010
1.500005
1.4999 1.500000

1.499995
1.499990
2
x

1.4998 1.499985

0 0.5 1 1.5

1.5
1.4997
1.4999

1.4999

1.4998
1.4996
1.4998

1.4997
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
1.4995
0 5 10 15 20 25 30 35 40 45 50
Time

(d) Reference
1.5001 SFB-IISMC
OFB-IISMC

1.5

1.4999

1.50000
2
x

1.4998 1.49998

1.49996

1.4997
19 19.5 20 20.5 21 21.5 22 22.5

1.4996

1.4995
0 50 100 150
Time

Fig. 3. (a) Reference tracking. (b) Parametric uncertainties. (c) Un-modeled side reaction. (d) Input disturbance.
66 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567

Appendix A. . Proof of Theorem 1 Let the Lyapunov candidate be selected as mentioned in (49)
and V0 dened in (21). Let chooses V1 be dened as [14,13]
Let the Lyapunov candidate be selected as mentioned in (49).  2
Let also select V0 as dened in (21) and V1 as mentioned in (50). 1 2 1 s
V1 = 2k2 |s| + v + k1 v3 (B10)
Take the time derivative of (31) and put into (51), we get, 2 3 2 |s|1/2
V 1 = s [a0 e1 + a1 e2 + F2 + G2 u + ] (A1) Take the time derivative of (B10) and by using (B9), we get,

1 2
 s
Put (33) and (34) into (A1) we get,

s  V 1 = 2k2 + k s + 2v3 v 3
2 1 |s|
V 1 = s 1 sat

+  (A2)  
s2 v3 v3 sv 3
s k1 + s k1 (B11)
Outside the boundary layer {|s| }, sat will be simply 2|s|5/2 |s|1/2 |s|1/2
become sign(s) [3]. Therefore, we may write (A2) as
 By using (B5), (B6) and (B9), we may write (B11) as
V 1 = s 1 +  (A3)


1 2 1 1 1
V 1 k1 k2 + k |s| 2 + k12 |v3 | + 2k2 + k12 1 |s| 2
V 1 0, 1  (A4) 2 1 2
By combining (25) and (A4), we conclude that k1 |v3 |2 k1 1 |v3 | 1
+22 |v3 | 1
+ + k1 2 |s| 2 (B12)
V < 0 2|s| 2
2

Inside the boundary layer, the system may be represented by [3] By rearranging (B12), we get
 = A0  + B0 s (A5) 1

1 2

1 2
 
V 1 k1 k2 + k 2k2 + k 1 k1 2 |s|
where, |s|
1
2
2 1 2 1
 
 
0 a0 k1 1 k1
A0 = (A6) k12 + 1 + 22 |v3 ||s| 2 + |v3 |2 (B13)
2 2
1 a1
  It may be written as
0
B0 = (A7)
1 1
V 1  T Q (B14)
|s|1/2
By taking V1 = T P0 , where P0 is the solution of the equation
P0 A + A0 P0 T = I. It may be veried that where,

V < 0  T = |s|1/2 |v3 |

Appendix B. . Proof of Theorem 2

1 2

1 2

k1 k2 + k 2k2 + k 1 k1 2
2 1 2 1
The control law mentioned in (39)(40) may be written as Q =

1 k1 k1
k1 s k12 + 1 + 22
v1 = + v2 (B1) 2 2 2
|s|1/2
s V 1 0k1
21 ,
v 2 = k2 (B2)
|s|
By taking time derivative of (37), we get,
 2
2k12 + 42 + k1 1 + 41 k13 + 8k12 2 4k14
k2 > (B15)
s = a1 e2 + F2 + G2 u +  (B3) 8 (k1 21 )

Assume the following bound on  , By combining (25) and (B15), we conclude that

|| |1 + |2 | (B4) V < 0

where,
|1 | 1 |s|1/2 (B5)
References
2 | 2
| (B6)
[1] A. Isidori, Nonlinear Control Systems, Springer-Verlag, New York, 1999.
1 and 2 are some positive constants or functions. Put (38) and [2] A.K. Jana, A.N. Samanta, S. Ganguly, Globally linearized control on diabatic
(B1) in (B3), we get continuous stirred tank reactor: a case study, ISA Trans. 44 (3) (2005)
423444.
k1 s [3] H.K. Khalil, Nonlinear Systems, 3rd edition, Prentice Hall, 2002.
s = + v2 +  (B7) [4] H.K. Khalil, Nonlinear Controls, 1st edition, Pearson Education, 2015.
|s|1/2 [5] M. Bouhamida, B. Daaou, A. Mansouri, M. Chenafa, Observer-based
input-output linearization of a multivariable continuous chemical reactor, J.
Let,
Korean Math Soc. 49 (3) (2012) 641658.
[6] S.S. Sastry, A. Isidori, Adaptive control of linearizable systems, IEEE Trans.
v3 = v2 + 2 (B8)
Autom. Control 34 (11) (1989) 11231131.
[7] J.-J.E. Slotine, J.K. Hedrick, Robust input-output feedback linearization, Int. J.
By using (B4), (B7) and (B8) we get,
Control 57 (5) (1993) 11331139.
k1 s [8] D. Tyner, M. Soroush, C.G. Grady, Adaptive temperature control of
s = + v3 + 1 (B9) multiproduct jacketed reactors, Ind. Eng. Chem. Res. 38 (11) (1999)
|s|1/2 43374344.
 F. Wallam, A.Y. Memon / Journal of Process Control 51 (2017) 5567
[9] M. Groebel, F. Allgower, M. Storz, E.D. Gilles, Asymptotically exact
I/O-linearization of an industrial distillation column, Am. Control Conf. IEEE 4
(1995) 26482652.
[10] D. Zhao, Q. Zhu, J. Dubbeldam, Terminal sliding mode control for continuous
stirred tank reactor, Chem. Eng. Res. Des. 94 (2015) 266274.
[11] Silvina I. Biagola, J.L. Figueroa, A high gain observer: application to the control
of an unstable nonlinear process, Comput. Chem. Eng. J. 28 (2004) 18811898.
[12] D. Nagrath, V. Parasad, B.W. Bequette, A model predictive formulation for
control of open-loop unstable cascade systems, Chem. Eng. Sci. 57 (2002)
365378.
[13] I. Nagesh, C. Edwards, A multivariable super twisted sliding mode approach,
Automatica 50 (2014) 984988.
[14] J.A. Moreno, M. Osorio, A. Lyapunov, Approach to Second Order Sliding Mode
Controllers and Observers, Conference on Decision and Control, IEEE, 2008,
pp. 28562861.
[15] H.K. Khalil, L. Praly, High-gain observers in nonlinear feedback control, Int. J.
Robust Nonlinear Control 24 (2014) 9931015.
67
[16] L.B. Freidovich, H.K. Khalil, Robust Feedback Linearization Using Extended
High Gain Observers, Conference on Decision and Control, IEEE, 2006, pp.
983988.
[17] L.B. Freidovich, H.K. Khalil, Performance recovery of feedback linearization
based designs, IEEE Trans. Autom. Control 53 (2008) 23242334.
[18] Utkin, Sliding Modes in Control and Optimization, Springer, 1992.
[19] C.T. Chen, C.S. Dai, Robust controller design for a class of nonlinear uncertain
chemical processes, J. Process Control 11 (2001) 469482.
[20] C.T. Chen, S.T. Peng, A sliding mode control scheme for non-minimum phase
non-linear uncertain input-delay chemical processes, J. Process Control 16
(2006) 3751.
[21] H. Sira-Ramirez, Dynamical sliding mode control strategies in the regulation
of nonlinear chemical processes, Int. J. Control 56 (1) (1992) 121.
[22] Z.H. Li, T.Y. Chai, C. Wen, Systematic design of robust controllers for nonlinear
uncertain systems, Int. J. Control 62 (4) (1995) 871892.
[23] A. Shahraz, R. Bozorghmehry, A fuzzy sliding mode control approach for
nonlinear chemical processes, Control Eng. Pract. 17 (2009) 541550.
[24] B.W. Bequette, Process Dynamics: Modeling, Analysis and Simulation, 1st
edition, Prentice Hall, 1998.
[25] A.N. Attasi, H.K. Khalil, A separation principle for the stabilization of a class of
nonlinear systems, IEEE Trans. Autom. Control 44 (9) (1999) 16721687.