ISA Transactions 47 (2008) 15–24

www.elsevier.com/locate/isatrans

Observer-based monitoring of heat exchangers

Carlos-Manuel Astorga-Zaragoza a,∗ , Vı́ctor-Manuel Alvarado-Martı́nez a , Arturo Zavala-Rı́o b ,
Rafael-Maxim Méndez-Ocaña a , Gerardo-Vicente Guerrero-Ramı́rez a
a CENIDET, Interior Internado Palmira S/N, Col. Palmira, C.P. 62490, Cuernavaca, Morelos, Mexico
b IPICYT, A.P. 2-66, C.P. 78216 San Luis Potosı́, S.L.P., Mexico

Received 4 November 2005; accepted 2 May 2007

Available online 15 August 2007

Abstract
The goal of this work is to provide a method for monitoring performance degradation in counter-flow double-pipe heat exchangers. The overall
heat transfer coefficient is estimated by an adaptive observer and monitored in order to infer when the heat exchanger needs preventive or corrective
maintenance. A simplified mathematical model is used to synthesize the adaptive observer and a more complex model is used for simulation. The
reliability of the proposed method was demonstrated via numerical simulations and laboratory experiments with a bench-scale pilot plant.
c 2007, ISA. Published by Elsevier Ltd. All rights reserved.

Keywords: Performance monitoring; Heat exchanger; Adaptive observer

1. Introduction to develop observers is based on Kalman filters or extended

Kalman filters estimators (EKF) [10,11]. A different type of
Heat exchangers are devices used for efficient heat transfer
state estimators has been developed by other authors and
from one fluid to another through an intermediate solid surface,
implies the use of Luenberger observers or some extensions
and is widely used in engineering processes. Some typical
examples are intercoolers, preheaters, boilers, and condensers of them [12,13]. All these observers are used only for state
in power plants. estimation and assume constant model parameters. However,
One of the major problems with heat exchangers is it may not be reasonable for many practical applications. It is
the deterioration of the heat transfer surface due to the often the case that some process parameters are unmodelled,
accumulation of a fouling film. This is, in fact, an unavoidable physically not available for measurement or they are time-
difficulty. However many efforts are continuously made to varying. Indeed, parameters such as the overall heat transfer
detect this problem, and to mitigate or handle the consequent coefficient may change during the operation of thermal
effects [1–3]. Fouling reduces the heat transfer effectiveness, processes. For these possible scenarios, a more accurate
frequently leading to an increased energy consumption, estimation of the state can be accomplished by allowing at least
as well as performance degradation of the equipments one model parameter to vary. The parameters are then estimated
and controllers [4]. As a solution, this work proposes a along with the state.
straightforward procedure to detect performance degradation An adaptive observer is basically characterized by the
using an adaptive observer. simultaneous estimation of the parameters and state variables.
Observers, usually referred to as software sensors [5,6], For linear systems, they have been studied since the 1970s
have become a suitable alternative to overcome the lack of
[14]. Recent works related to adaptive observers commonly use
relevant on-line sensors. They are used to infer unknown
dynamical transformations to represent the original system in
parameters or unmeasured state variables from on-line and/or
some canonical form, in which the presence of the unknown
off-line measurements; see e.g. [7–9]. Much of the work done
parameters is simplified to some extent [15,16]. In a further
∗ Corresponding author. Tel.: +52 777 3 62 77 70x213; fax: +52 777 3 62 77 work [17], a simple constant-gain observer was proposed. This
95. approach requires two tuning parameters whatever the number
E-mail address: [email protected] (C.-M. Astorga-Zaragoza). of differential equations composing the model be. In addition,

c 2007, ISA. Published by Elsevier Ltd. All rights reserved.

0019-0578/$ - see front matter
doi:10.1016/j.isatra.2007.05.003
16 C.-M. Astorga-Zaragoza et al. / ISA Transactions 47 (2008) 15–24

Nomenclature
A Heat transfer surface area, m2
cpc Specific heat in the cold side, J/(kg K)
cph Specific heat in the hot side, J/(kg K)
Tci , Thi Inlet temperatures in the cold and hot sides,
respectively, K
Tco , Tho Outlet temperatures in the cold and hot sides,
respectively, K Fig. 1. One-cell representation of a countercurrent heat exchanger.
U Heat transfer coefficient, J/(m2 K s)
vc Flow rate in the cold side, m3 /s A1 constant volume in the tubes is considered,
vh Flow rate in the hot side, m3 /s A2 linear driving forces, this means that the rate at which
Vc Volume in the cold side, m3 heat is transferred depends linearly on the temperature
Vh Volume in the hot side, m3 difference between the two fluids,
ρc Density of the cold fluid, kg/m3 A3 the heat transfer coefficient is related to the temperatures of
ρh Density of the hot fluid, kg/m3 the fluids,
A4 there is no heat transference between the cold side and the
environment,
the proposed observer is proved not only to be stable but also to A5 the thermophysical properties of the fluids are constant,
yield an estimation error close to zero. A6 there is no energy storage in the walls.
The aim of the current work is to propose a method based
on adaptive observers for monitoring the performance of a The system dynamics is obtained through an energy balance
heat exchanger, by tracking the global heat transfer coefficient rule applied to every element of a lumped model [18].
U . Hence, having the knowledge of the variation path of this Since the lumping procedure generally assumes that every
coefficient, it is really feasible to determine when the equipment element behaves like a perfectly stirred tank [19], the fluid
needs a preventive or corrective maintenance. temperature at each of such elements is considered to be
This paper is organized as follows. In Section 2, a simplified uniformly distributed. This is why the outlet temperatures are
model of reasonable accuracy is presented. In Section 3, taken to represent the bulk temperatures (at every considered
two different observers are designed on the basis of the element) to get the accumulation terms, and the driving force
mathematical model described in Section 2 and subsequently, producing the heat conduction from the hot fluid to the cold
they are experimentally evaluated on a bench-scale pilot plant one is modelled through the outlet temperatures difference.
and by means of numerical simulations. Both, the experimental Thus, the application of the energy balance rule considering a
and simulation results are discussed in Section 4. Finally, single element per fluid (covering the whole tube length) to a
conclusions are stated in Section 5. countercurrent heat exchanger gives rise to the following 2nd
order dynamics:
2. Simplified model of a heat exchanger
vc

dTco UA
This work will focus on recuperators since such kind of heat
 = (Tci − Tco ) + (Tho − Tco )
cpc ρc Vc

dt Vc

exchangers are the most commonly found heat exchangers in dTho vh UA (1)
practice. These heat exchangers may be designed according
 = (Thi − Tho ) + (Tco − Tho ) .
cph ρh Vh

dt Vh

to one of the following configurations: (i) parallel-flow, in
which fluids flow in the same direction, (ii) counter-flow, in This model can be illustrated as a single cell which consists
which fluids flow in opposite directions and (iii) cross-flow, of two perfectly stirred tanks with inflows and outflows [20].
in which the direction of fluids are perpendicular to each The two tanks are connected by a heat transfer area between
other. Among them, the counter-flow heat exchanger design them (see Fig. 1). Despite its simplicity, the model in Eq.
is the most efficient when comparing heat transfer rate per (1) keeps the main features of the qualitative behavior of
unit surface area. Moreover, the counter-flow heat exchanger heat exchangers under the stated assumptions. Moreover, it
has three significant advantages over the parallel-flow design. turns out to be a suitable model for specific types of heat
First, its more uniform temperature difference between the two exchangers or certain types of applications as pointed out
fluids minimizes the thermal stresses throughout the exchanger. in [21, Section 15.4], [22, Section 3], and [20, Section 2.1.2].
Second, the outlet temperature of the cold fluid can approach Furthermore, it is not always possible to install sensors to
the highest temperature of the hot fluid (that is, the inlet measure the fluid temperatures at intermediate positions of
temperature). Third, the more uniform temperature difference the exchanger. For this reason, simple models, like (1), that
produces a more uniform rate of heat transfer throughout the approach the behavior of the outlet temperatures exclusively
heat exchanger. in terms of the inlet ones and the system properties, prove to
In this work a counter-flow double-pipe heat exchanger is be useful for certain types of applications that involve state
considered. The mathematical model presented here, takes into measurements, like feedback control [23] or state/disturbance
account the following assumptions: estimation [24]. It is worth mentioning that if the heat transfer
 C.-M. Astorga-Zaragoza et al. / ISA Transactions 47 (2008) 15–24 17

forms of the considered systems, which can be called for this

reason “adaptive observer forms” [17].
These systems have the form:

dy(t)


 = α(y(t), z(t), u(t)) + β(y(t), z(t), u(t))θ(t)
dt (3)
 dz(t) = γ (y(t), z(t), u(t))

dt
where y(t) ∈ R p is the output vector of the system (the
Fig. 2. Three-cell representation of a countercurrent heat exchanger. measurable variables), z(t) ∈ Rq is the vector of the
unmeasurable variables, u(t) ∈ Rm is the vector of the
coefficient, U , is assumed to be related to the temperatures of measurable bounded inputs and θ(t) ∈ Rr is the vector of the
the fluids (see Assumption A3 above), model (1) is actually unknown parameters. α(y(t), z(t), u(t)) and β(y(t), z(t), u(t))
nonlinear, and such a variable value of U may be seen as a are two globally Lipschitz functions with respect to z(t).
compensation that shortens the quantitative inaccuracy of such An adaptive observer for a system having the form (3) is
a simple model. On the other hand, it has been shown [25] that
given by [17]:
the consideration of several cascading cells, as shown in Fig. 2,
increases the precision of the model, keeping the main features 
dŷ(t)
of the qualitative behavior of heat exchangers. Note that the

 = α(y(t), ẑ(t), u(t)) + β(y(t), ẑ(t), u(t))θ̂(t)

 dt
outflow of one cell is the inflow of the next one. A three-cell  − k y (ŷ(t) − y(t))



model of a counter-flow heat exchanger is dẑ(t) (4)
= γ (y(t), ẑ(t), u(t))
dt

dTco 3vc UA 
(T2c − Tco ) + (T1h − Tco )

= 
cpc ρc Vc  dθ̂(t) = −k β T (y(t), ẑ(t), u(t))(ŷ(t) − y(t))

dt Vc


θ
dT1h 3vh UA dt
= (Thi − T1h ) + (Tco − T1h )
dt Vh cph ρh Vh such that for any initial conditions y(0), z(0), ŷ(0), ẑ(0) (the
dT2c 3vc UA hat (ˆ) represents the estimated value of the variable wearing
= (T3c − T2c ) + (T2h − T2c ) it, i.e. ŷ(t) represents the estimated value of y(t)), and any
dt Vc cpc ρc Vc
(2) measurable bounded input u(t), the Euclidian norm of the
dT2h 3vh UA
= (T1h − T2h ) + (T2c − T2h ) estimation error
dt Vh cph ρh Vh
dT3c 3vc UA e y = kŷ(t) − y(t)k (5)
= (Tci − T3c ) + (Tho − T3c )
dt Vc cpc ρc Vc asymptotically converges to zero as t tends to infinity, while
dTho 3vh UA
= (T2h − Tho ) + (T3c − Tho ).
dt Vh cph ρh Vh eθ = kθ̂(t) − θ(t)k (6)
A low-order model like (1) proves to be useful for (observer remains bounded. Also, if β T (y, z, u, t) is persistently exciting
or control) synthesis purposes, while a higher-order dynamics ((·)T denotes the transpose of a matrix), and its time derivative
like (2) is convenient for the implementation of numerical is bounded, then kθ̂ (t) − θ(t)kt→∞
−−→0. Constants k y > 0 and
simulations oriented to test the (observer or control) designed kθ > 0 are the tuning parameters of the observer.
algorithms.
3. Adaptive observer design 3.1.1. The simplified adaptive observer
An adaptive observer is a recursive algorithm to estimate If there are no unmeasurable states (which is the case in the
the state of a system containing unknown parameters or to application described below), the observer given by (4) can be
jointly estimate both its state and its unknown parameters. In simplified as follows:
this section two different approaches are presented. The main
dŷ(t)

difference among them lies in the way the observer gains are
 dt = α(y(t), u(t)) + β(y(t), u(t))θ̂(t)


computed. These are used to adjust the rate of convergence 
towards the “real” state. A comparison of the performance − k y (ŷ(t) − y(t)) (7)
 (t) = −k β T (y(t), u(t))(ŷ(t) − y(t)).
d θ̂

of these observers is presented. From these results, process


θ
engineers can determine which approach is more appropriate dt
for their needs. This observer assumes that measurements are available
3.1. Adaptive observer for systems having the adaptive continuously. This is not always the case in industrial
observer form applications. If measurements are available digitally, with a
sampling time Ts , then a discrete-time version of the observer in
Most designs of adaptive observers are based on specific (7) is obtained by approximating the derivatives using a simple
18 C.-M. Astorga-Zaragoza et al. / ISA Transactions 47 (2008) 15–24

T
Euler approximation as follows: where ẑ(t) = ẑ1 (t) ẑ2 (t) ∈ R2n , ẑ1 (t) ∈ Rn , ẑ2 (t) ∈ Rn ,

ŷ(k + 1) = ŷ(k) + Ts [α(y(k), u(k)) + β(y(k),

  
I 0
3(y(t), s(t)) = d ,
u(k))θ̂ (k) − k y (ŷ(k) − y(k))] (8) 0 F1 (y(t), s(t))
θ̂(k + 1) = θ̂(k) − Ts kθ β T (y(k), u(k)) (ŷ(k) − y(k))

Sθ is a symmetric positive-definite matrix which satisfies the
where y(k + 1) and y(k) are in a simplified form to write algebraic Lyapunov equation
y(kTs + Ts ) and y(kTs ) respectively.
θSθ + AT Sθ A + Sθ A − CT C = 0 (12)
3.2. High-gain observer for a class of nonlinear systems θ > 0 is the tuning parameter of the observer and the matrix A
is given by
High-gain observers have been proposed by Gauthier et al.  
[26]. In contrast to the adaptive observer presented above, the 0 Id
A= .
gain of the observer is computed dynamically. Farza et al. [27] 0 0
proposed a modified high-gain observer which can be used
like an adaptive observer for the simultaneous estimation of It has been demonstrated that the observer given by (11)
states and parameters. One of the main characteristics of these converges asymptotically to zero as t tends to infinity, i.e.
observers lies in the simplicity of their implementation. Due −−→0
ez = kẑ(t) − z(t)kt→∞ (13)
to the particular gain structure, the tuning of these observers is
reduced to the calibration of a single parameter. for any initial conditions z(0), ẑ(0) and any measurable
The nonlinear systems in consideration have the form: bounded input u(t). For further details about the synthesis and
 dz (t) the proof of convergence of this observer, see [26,27].
1
 = f (z1 (t), s(t))F1 (z1 (t), s(t))z2 (t)
dt


3.2.1. The discrete version
+ b1 (u(t), z1 (t), s(t))


(9)
 dz2 (t) = b (u(t), z (t), s(t)) + ε(t) The discrete version of the observer in (11) is obtained by

 dt

 2 1 using a forward Euler discretization as follows:
y(t) = z1 (t)
ẑ(k + 1) = ẑ(k) + Ts [ f (y(k), s(k))F(y(k), s(k))ẑ(k)



T  + B(u(k), y(k), s(k)) − f (y(k), s(k))

where z(t) = z1 (t) z2 (t) ∈ R2n , z1 (t) ∈ Rn contains the (14)
measured states and z2 (t) ∈ Rn contains the unmeasured states. 
 × 3−1 (y(k), s(k))S−1θ C (ŷ(k) − y(k))]
T

u(t) ∈ Rm , y(t) ∈ Rn and s(t) ∈ R p are respectively, the ŷ(k) = Cẑ(k).


input, the output and an arbitrary signal of the system (they are
The main objective in this work is to confirm by means of
assumed to be known). F1 (·) is an n × n matrix; f (·) is a scalar
numerical simulations and through experimental results that
real-valued function. ε(t) is an unknown but bounded function
the discrete adaptive observers given by Eqs. (8) and (14)
which may depend on u(t), s(t), z(t), noise, etc.
can be synthesized using the simplified one-cell model in Eq.
System (9) can be written in the condensed form:
(1) in order to design software sensors for monitoring the

dz(t) performance of heat exchangers. Further, the results obtained

 = f (z1 (t), s(t))F(z1 (t), s(t))z(t) through the implementation of both techniques are compared in
dt (10)
 + B(u(t), z1 (t), s(t)) + εv (t) order to conclude which is the one that best fits for applications

y(t) = Cz(t) involving double-pipe heat exchangers.
where
4. Application to a heat exchanger
0 F1 (z1 (t), s(t))
   
0
F(z1 (t), s(t)) = , εv (t) =
0 0 ε(t)
According to Assumption A5, A, cpc , cph , ρc , ρh , Vh and Vc
b1 (u(t), z1 (t), s(t))
 
B(u(t), z1 (t), s(t)) = , are constant. Then, the model given by Eq. (1) can be rewritten
b2 (u(t), z1 (t), s(t)) in the following form:


C = Id 0
dTco vc

 = (Tci − Tco ) + kc U (Tho − Tco )
Id is an identity matrix of adequate dimensions. It has been

dt Vc (15)
demonstrated in [27] that if certain assumptions are satisfied dT v
 ho = h (Thi − Tho ) + kh U (Tco − Tho )

(the heat exchanger model satisfies these assumptions which dt Vh
are related with boundedness of the states, functions, signals,
etc.), an observer for the system given by (10) is: where kc = A/(cpc ρc Vc ) and kh = A/(cph ρh Vh ) are constant.
Three supplementary assumptions are introduced:
dẑ(t)


 = f (y(t), s(t))F(y(t), s(t))ẑ(t) + B(u(t), y(t), s(t)) A7 the flows in the cold and the hot sides (vc and vh ) are
dt (11) measured and they are the inputs of the system (u 1 (t) and
 − f (y(t), s(t))3 (y(t), s(t))Sθ C (ŷ(t) − y(t))
−1 −1 T
u 2 (t) respectively),

ŷ(t) = Cẑ(t)
 C.-M. Astorga-Zaragoza et al. / ISA Transactions 47 (2008) 15–24 19

A8 the cold and hot outlet temperatures (Tco and Tho ) are  u 1 (t) 
(Tci − Tco )
measured and they are the outputs of the system (y1 (t) and
   
Tco  Vc  0
y2 (t) respectively), Tho   u 2 (t) 0
 
 
×  + (Thi − Tho ) +   (19)
ε
  
A9 the cold and hot inlet temperatures (Tci and Thi ) are U  Vh 
measured and they are constant. It is worth noting that the U 
0
 ε
exclusive consideration of such measurements is the usual 0 | {z }
} εv (t)
case in an industrial environment.
| {z
B(u(t),z1 (t),s(t))


Tco
4.1. Design of observer 1
y1 (t)
     
Tco 1 0 0 0 
Tho  .

= = (20)
y2 (t) Tho 0 1 0 0 U 
Under Assumptions A7–A9, the model in (15) may be | {z } | {z } U
y(t) C=[Id 0]
rewritten in the following matrix representation:
 dy (t)   u 1 (t)  It can be seen that the system given by Eqs. (19) and (20)
1
(Tci − y1 (t)) has the same form of the system given by (10), where
 dt   Vc 
 dy (t)  =  u 2 (t)
 z1 (t) = (Tco Tho )T , z2 (t) = (U U )T = (U1 U2 )T ,
2
(Thi − y2 (t))
 
f (z1 (t), s(t)) = 1, s(t) = 0. It should be noted that the original
dt Vh
system (15) has been augmented with the differential equation
kc (y2 (t) − y1 (t))
 
+ U (t) (16) dU (t)/dt = ε(t), i.e. the unknown parameter U (t) is treated
kh (y1 (t) − y2 (t)) like an unmeasured state. This differential equation is written
which has the same form of System (3) without nonmeasurable twice in Eq. (19) in order to satisfy z1 (t) ∈ R2 , z2 (t) ∈ R2 .
states z(t). In this model, U is considered as a parameter that Then, it is possible to design a discrete observer of the form
may undergo continuous-time variations during the transient (14) for the system given by (19) and (20) as follows:
period of the system responses. Then, a discrete adaptive
(i) The matrix 3−1 (y(k), s(k)) is:
observer of the form (8) for System (16) is given by Eqs. (17)
 
and (18): I2 0
3 (y(k), s(k)) =
−1

u 1 (k)
 0 F1 (y(k), s(k))
(T ci − y1 (k))
ŷ1 (k + 1) ŷ (k)
   
 Vc where
= 1

+ Ts 
ŷ2 (k + 1) ŷ2 (k) u 2 (k)
 
(Thi − y2 (k))
  
F1 (y(k), s(k))
Vh
k (T (k) − Tco (k))
 
0
 = c ho .
0 kc (Tho (k) − Tco (k))
kc (y2 (k) − y1 (k)) ŷ (k) − y1 (k) 
   
+ Û (k) − k y 1  (17)
kh (y1 (k) − y2 (k)) ŷ2 (k) − y2 (k)  (ii) The solution to Eq. (12) is:
 1 1 
I2 − 2 I2
Sθ =  θ θ .
T
k (y (k) − y1 (k))
 
Û (k + 1) = Û (k) − Ts kθ c 2 1 2 
kh (y1 (k) − y2 (k)) − 2 I2 I2

ŷ1 (k) − y1 (k)
 θ θ3
× . (18)
ŷ2 (k) − y2 (k) (iii) The gain of the observer in (14) is:
f (y(k), s(k))3−1 (y(k), s(k))S−1
θ C
T
4.2. Design of observer 2  
2θI2
= 2 .
θ F1 (y(k), s(k))
Under Assumptions A7–A9, the model given by (15) can be
rewritten as (iv) Finally, the discrete observer is given by:
 dTco  
T̂co (k + 1) = T̂co (k) + Ts kc (Tho (k) − Tco (k))Û1 (k)
 dt 
kc (Tho − Tco )
 
 dTho  0 0 0
u 1 (k)

kh (Tco − Tho )
 
 dt  0 0 0 + (Tci (k) − Tco (k)) − 2θ (T̂co (k) − Tco (k))
 dU  = 0 0 Vc
   
  0 0 
 (21)
 dt  0 0 0 0
  T̂ho (k + 1) = T̂ho (k) + Ts kh (Tco (k) − Tho (k))Û2 (k)
dU | {z }
F(z1 (t),s(t))
| dt u 2 (k)

{z } + (Thi (k) − Tho (k)) − 2θ (T̂ho (k) − Tho (k))
dz(t)/dt Vh
20 C.-M. Astorga-Zaragoza et al. / ISA Transactions 47 (2008) 15–24

Fig. 3. Block diagram for simulations 1 and 2.

Fig. 5. The inlet flow rate in the hot side vh (t).
Table 1
Physical data used in the simulations

Constant Value Units

A 0.633 m2
ρh 1000 kg/m3
ρc 1000 kg/m3
Vc 6.05 × 10−3 m3
Vh 3.2 × 10−3 m3
cpc 1910 J/( K kg)
cph 1590 J/( K kg)

convergence of the observers in the common situation in which

Fig. 4. Block diagram for simulation 3. U remains constant over a long time. Considering the process
model given by (1), Observers 1 and 2 were simulated using
Ts θ 2 (T̂co (k) − Tco (k)) the values given in Table 1. These constants, corresponding to
Û1 (k + 1) = Û1 (k) −
kc (Tho (k) − Tco (k)) a heat exchanger having fast dynamics, were taken from [20].
Ts θ 2 (T̂ho (k) − Tho (k)) A value of U = 180 J/(m2 K s) was considered. The inlet
Û2 (k + 1) = Û2 (k) − . temperatures Tci and Thi were 303 K and 320 K respectively.
kh (Tco (k) − Tho (k))
The cold fluid flow rate vc was 3.15 × 10−4 m3 /s. The hot
It should be remarked that although Û1 (k) and Û2 (k) undergo fluid flow rate vh was varied between 1.9 × 10−4 m3 /s and
different initial trajectories (each), they will converge to the 2.09 × 10−4 m3 /s as shown in Fig. 5. The initial conditions of
same “true value” of U as k tends to infinity. the process model were: Tco 0 = 303 K, T 0 = 310 K, while
ho
those of the observers were: T̂co 0 = 300 K, T̂ 0 = 306 K.
4.3. Numerical simulations ho
These small discrepancies between the initial conditions of the
In order to test the performance of the proposed observers, process and those of the observers are reasonable and realistic
three numerical simulations were carried out. They consist in considering that temperature is a process variable that can
estimating the heat transfer coefficient U from measurements be easily measured. The sampling time of the observers was
of the inlet flows and the inlet–outlet temperatures of a counter- Ts = 5 s. In order to evaluate the observer performance against
flow heat exchanger. uncertainties on the knowledge of U , the initial value of the
In Case 1, a constant heat transfer coefficient is considered in estimate was Û 0 = 225 J/(m2 K s) (this represents 25%
the process model, i.e. U (t) ≡ U . In Case 2, U (t) is considered of deviation from the process value). The tuning parameters
as time-varying. In these simulations, both the process and the were k y = 0.12, kθ = 150 (for Observer 1) and θ = 0.005
adaptive observers consider a one-cell heat exchanger model (for Observer 2). The results of the estimation of temperatures
(Fig. 3). In Case 3 the three-cell model given by (2) was Tco , Tho are reported in Figs. 6 and 7 respectively. The dashed
considered for the process, and the observers were designed curve 1 corresponds to the estimates obtained using Observer 1,
using the one-cell model given by (1) (Fig. 4). Subsequently, while the dashed curve 2 corresponds to the estimates obtained
the discrete observer given by Eqs. (17) and (18) and the one using Observer 2. It can be seen that the convergence of the
given by (21) will be denoted Observer 1 and Observer 2 estimates T̂co , T̂ho (dashed lines) proves to be fast (in about
respectively. twelve minutes). This fact is not surprising because, actually
Tco , Tho are the measured outputs of the process.
4.3.1. Case 1: One-cell model, one-cell observers, U (t) ≡ U Fig. 8 shows the estimation of U using Observer 1. Fig. 9
The objective of this series of simulations is to prove the shows the estimation of U using Observer 2; as expected,
 C.-M. Astorga-Zaragoza et al. / ISA Transactions 47 (2008) 15–24 21

Fig. 6. The outlet temperature in the cold side Tco (solid line) and its estimated Fig. 8. Simulated U (solid line) and its estimated value Û (dashed line) using
values T̂co using Observer 1 and Observer 2 (dashed lines). Observer 1.

Fig. 7. The outlet temperature in the hot side Tho (solid line) and its estimated Fig. 9. Simulated U (solid line) and its estimated values Û1 , Û2 (dashed lines)
values T̂ho using Observer 1 and Observer 2 (dashed lines). using Observer 2.

Û1 (k) and Û2 (k) follow different trajectories before they degradation problem). The degradation can be attributed to
converge towards the “true value”. In both cases, convergence fouling. Fouling is the accumulation of unwanted materials on
is guaranteed, and if adequate values of the tuning parameters the heat transfer surface. This increases the resistance to heat
k y , kθ and θ are selected, no matter the degree of deviation transfer and reduces the effectiveness of these equipments. In
of the initial value of Û 0 from the simulated value in the this simulation the fouling effects were taken into account in the
process model. Larger values of these tuning parameters ensure following way. Firstly, an initial value of U = 180 J/(m2 K s)
a smaller convergence time while smaller values have the was considered, followed by an abrupt change from 180 to
opposite effect. However, large tuning values shall be avoided 175 J/(m2 K s) at time t = 70 min. Several factors can cause
since the observer may become too sensitive to measurement this kind of variation, for instance, a fault in a sensor or abrupt
noise in real-time applications. sediments of impurities coming from previous stages of the
As was stated in Section 3 (Eqs. (5), (6) and (13)), the process. Secondly, at t = 130 min, U begins to deteriorate
Euclidian norm slowly due to a gradual deposition of a fouling layer. These
q variations are illustrated in Fig. 10.
kŷ(k) − y(k)k, i.e. (T̂co (k) − Tco (k))2 + (T̂ho (k) − Tho (k))2 , The simulation of the process model was carried out using
the same constants used in the previous simulation (Table 1) and
decays to zero as k tends to infinity while kθ̂ (k) − θ (k)k,
the same values of Tci , Thi , vc , Tco
0 , T 0 . The flow rate v was
h
i.e. kÛ (k) − U (k)k, remains bounded. ho
considered constant at 1.9 × 10−4 m3 /s. The initial conditions
of the two observers were slightly modified at T̂co 0 = 303 K,
4.3.2. Case 2: One-cell model, one-cell observers, time-
0
varying U (t) T̂ho = 310 K (no error was assumed in measuring Tco and
The present computations are executed to get an accurate Tho ); Û 0 = Û10 = Û20 = 198 J/(m2 K s) (this represents 10%
screening of the variation of the observer estimate, by of deviation from the “true value” U 0 ). These modifications in
corroborating if it is in agreement with the simulated overall the initial values of the observers were made because the main
heat transfer coefficient, which undergoes either an abrupt purpose of this simulation is to show the tracking capabilities
change (prompted in essence by an unexpected fault in of the estimators. The observer parameters (k y , kθ , θ) were the
the process) or a gradual variation (essentially due to a same as the previous ones.
22 C.-M. Astorga-Zaragoza et al. / ISA Transactions 47 (2008) 15–24

Fig. 10. U is considered time-varying. Solid line: simulated U ; dashed lines: Fig. 11. Estimation of U for a three-cell model using adaptive observers based
its estimated values Û using Observer 1 (curve 1) and Observer 2 (curve 2). on a single-cell model. Solid line: simulated U ; dashed lines: its estimated
values Û using Observer 1 (curve 1) and Observer 2 (curve 2).
Fig. 10 shows the simulation result of the estimation of
the heat transfer coefficient. The dashed curve 1 corresponds Table 2
to Û (k) generated by Observer 1, while the dashed curve 2 Physical data used in the experiments
corresponds to Û1 (k) generated by Observer 2 (for the sake of Constant Value Units
simplicity, in the following figures Û2 (k) will not be plotted, but
A 14.0 × 10−3 m2
it is worth noting that it converges to the same value of Û1 (k)). cpc 4174 J/( K kg)
It can be seen that once the observers converge, they track well cph 4179 J/( K kg)
U in spite of the time-varying nature of this parameter. Vc 135.0 × 10−6 m3
Vh 15.51 × 10−6 m3
4.3.3. Case 3: Three-cell model, one-cell observer, time- ρh 983.3 kg/m3
varying U (t) ρc 991.8 kg/m3
This study is similar to the previous case, but a more
complex model was implemented to simulate the heat 4.4.1. Instrumentation
exchanger performance. Observers 1 and 2 were designed based The pilot plant is equipped with the following instruments:
on the single-cell model in Eq. (1). Then, they were used two temperature transmitters RTD Pt-100 (to measure Thi
to monitor the value of U on a three-celled heat exchanger and Tco ); two glass thermometers (to measure Tci and Tho );
model (Fig. 4). The parameters of the process model, the initial two rotameters (to measure vc and vh ). Needle valves in the
conditions and the tuning parameters of the observers, were rotameters could be used to adjust the flow rates vc and vh .
the same as used in the previous case. The value of U was Additionally, the inlet temperature in the hot side Thi can be
the same for each cell, beginning at 180 J/(m2 K s) and manipulated via an ON–OFF thermoregulator; the inlet flow
having the same profile used in the previous simulation (see rate in the hot side vh can be regulated via a PID controller
Fig. 11). It can be seen that the estimated values Û (k) and and a pneumatic valve.
Û1 (k) do not coincide with the simulated process value of
U . This is an expected outcome provided that the adaptive
4.4.2. Operational conditions
observers do not have a precise knowledge about the dynamics
The inlet temperatures were Tci = 299.7 K and Thi =
of the process model. However, the observers detect well the
343.3 K. The inlet flow rate in the cold side was vc = 6.67 ×
variations of U . It can be seen that a change 1U1 (at instant
10−6 m3 /s; vh was time-varying between 16.7 × 10−6 m3 /s
t1 ) gives rise to a change 1Û1 . In the same way, a gradual
and 20.0 × 10−6 m3 /s as shown in Fig. 12.
change 1U2 occurring from t2 until t3 gives rise to a change
1Û2 . From this, we conclude that a single-cell observer can
be used for monitoring the overall heat transfer coefficient in 4.4.3. Observer implementation
order to detect a performance degradation of the heat exchanger. The adaptive observers were implemented in a personal
Nevertheless, the place where the fault takes place cannot be computer, using MATLAB/Simulink R
. Temperatures were
located physically. sampled at ts = 1 min intervals. These temperatures were
used off-line in order to estimate U . The sampling period of
4.4. Experimental results the adaptive observers Ts was 2.5 s, then the measures were
interpolated in order to have a measure available at these time
The experimental pilot plant is a counter-flow double-pipe periods. Larger sampling periods are often desirable, but it
heat exchanger where the hot water flows through the inner has been demonstrated that discrete observers could become
tube and the cooling water flows through the external tube. The unstable when the sampling period exceeds a critical value [27].
constants and physical data used for the internal model of the An alternative to overcome this difficulty is to use continuous-
observer are given in Table 2. discrete observers (e.g. [7]). The tuning parameters of the
 C.-M. Astorga-Zaragoza et al. / ISA Transactions 47 (2008) 15–24 23

Fig. 14. The outlet temperatures Tho , Tco and their estimates using Observer 2.
Fig. 12. The inlet flow rate in the hot side vh .

Fig. 15. The experimentally estimated value Û in a bench-scale pilot plant

Fig. 13. The outlet temperatures Tho , Tco and their estimates using Observer 1. using Observer 1 (solid line) and Observer 2 (dashed line).

observers were k y = 0.5, kθ = 1000 (for Observer 1); θ = 0.08 approach is based on adaptive observers, which imply simple
(for Observer 2). It is worth noting that these tuning parameters procedures for implementation and calibration, since the tuning
should not be chosen too high since this renders the observers parameters k y , kθ (for Observer 1) and θ (for Observer 2)
very sensitive to noise. The results are given in Figs. 13–15. are constant and consequently it is not necessary to solve any
Figs. 13 and 14 show the outlet temperatures Tco , Tho and their dynamical system. In contrast, when Kalman filters are used, a
estimates T̂co , T̂ho respectively. They correspond to the results Riccati equation must be solved and numerical instability can
obtained using Observers 1 and 2 respectively. It can be seen easily arise from accumulated errors [28].
that the estimates converge well towards the measured values. The performance evaluation of the proposed observers
Concerning the heat transfer coefficient U , Fig. 15 shows integrates data reduction with numerical simulations and
that the two observers estimate the same value. As expected, experimental data from a bench-scale heat exchanger. At
the variations of U reflect the variations on the operating the first stage, the numerical simulations confirmed the
conditions. Thus, the value of U must remain unchanged for all convergence of the observers. Indeed, the results prove that
the experiments developed under the same operating conditions the estimated values of T̂co (Fig. 6), T̂ho (Fig. 7) and
of the process. Û (t) (Figs. 8–11) compare quite well with the simulated
variables Tco , Tho and U (t), though no relation exists between
5. Conclusions
the U (t)-simulation model and the observers. After that,
During the course of this work, an improved method the real-time experiments with a bench-scale pilot plant
was developed for monitoring performance degradation in demonstrated good agreement in temperature measurements
counter-flow double-pipe heat exchangers. It essentially aims and simulations (Figs. 13 and 14). It is important to note that
at estimating the overall heat transfer coefficient, U (t), whose even if U (t) cannot be measured directly, the estimation of
variation analysis conveys a full description of heat transfer rate this parameter in terms of experimental temperatures exhibit
evolution. Thus, any reduction observed in heat transfer rates a consistent and characteristic behavior in several respects, as
may be directly related to heating performance degradation. already noted in Fig. 15: (a) As expected, at short times, the
Accordingly, the information obtained from the monitoring of estimated overall heat transfer coefficient remains the same for
U (t) can readily provide a strong support base to determine a set of operation conditions. In practice, any variation of this
when a preventive or a corrective maintenance is necessary for parameter due to performance degradation can take some weeks
preserving or restoring the heat transfer rates in processes. This or several months to appear. (b) Then, flow perturbations were
24 C.-M. Astorga-Zaragoza et al. / ISA Transactions 47 (2008) 15–24

programmed to study the sensibility of U (t) as a function of [17] Besançon G. Remarks on nonlinear adaptive observer design. Systems &
operation variables. Each time, the corresponding response of Control Letters 2000;41:271–80.
[18] Fazlur-Rahman MHR, Devanathan R. Modelling and dynamic feedback
U and the corresponding steady state values were observed.
linearisation of a heat exchanger model. In: Proceedings of the third IEEE
Because observers have been derived from a simplified conference on control applications. 1994. p. 1801–6.
lumped-model, it may be argued that the present analysis does [19] Varga EI, Hangos KM, Szigeti F. Controllability and observability of
not relate the variation of the overall heat transfer coefficient heat exchanger networks in the time-varying parameter case. Control
to the axial position in the heat exchanger (1). The reason for Engineering Practice 1995;3:1409–19.
[20] Weyer E, Szederkényi G, Hangos K. Grey box fault detection of heat
this is that the model does not take care of spatial information.
exchangers. Control Engineering Practice 2000;8:121–31.
In spite of this disadvantage, the effectiveness of the simple [21] Walas SM. Modeling with differential equations in chemical engineering.
lumped-parameter model to synthesize properly the adaptive Stoneham (MA, USA): Butterworth-Heinemann; 1991.
observer has been verified. [22] Wolff EA, Mathisen KW, Skogestad S. Dynamics and controllability of
A further point to be discussed as a potential advantage is heat exchanger networks. In: Proceedings of computer oriented process
engineering. 1991. p. 117–22.
the fact that minimal information is demanded for achieving the [23] Abdelghani-Idrissi MA, Arbaoui MA, Estel L, Richalet J. Predictive
observer computations. Summarizing, only limited knowledge functional control of a counter current heat exchanger using convexity
on the system behavior is needed, no assumptions are necessary property. Chemical Engineering and Processing 2001;40:449–57.
to initialize the computation of U (t) and no assumptions [24] Parrish JR, Brosilow CB. Nonlinear inferential control. AIChE Journal
regarding the system dynamics are required. 1988;34:633–44.
[25] Mathisen KW, Morari M, Skogestad S. Dynamic models for heat
exchangers and heat exchangers networks. Computers in Chemical
References Engineering 1993;(Suppl.):S459–63.
[26] Gauthier JP, Hammouri H, Othman S. A simple observer for nonlinear
[1] Sawyer A, Ruggles AE, Upadhyaya BR, Eryurek E, Kavaklioglu K, Miller systems, applications to bioreactors. IEEE Transactions on Automatic
J. Development of a data-based method for performance monitoring of Control 1992;37:875–80.
heat exchangers. In: Proc. of the 2003 ECI conf. on heat exchanger fouling [27] Farza M, Hammouri H, Othman S, Busawon K. Nonlinear observers
and cleaning: Fundamentals and applications. 2003. p. 255–8. for parameter estimation in bioprocesses. Chemical Engineering Science
[2] Isogai S, Nakamura M. Analysis and steps to mitigate heat exchanger 1997;52:4251–67.
fouling in an aromatics plant. In: Proc. of the 2003 ECI conf. on heat [28] Verhaegen M, Van-Dooren P. Numerical aspects of different Kalman
exchanger fouling and cleaning: Fundamentals and applications. 2003. filter implementations. IEEE Transactions of Automatic Control 1986;31:
p. 221–5. 907–17.
[3] Joshi HM, Brons G. Analysis and steps to mitigate heat exchanger fouling
in an aromatics plant. In: Proc. of the 2003 ECI conf. on heat exchanger
fouling and cleaning: Fundamentals and applications. 2003. p. 221–5. Carlos-Manuel Astorga-Zaragoza was born in Mi-
[4] Hoo KA, Piovoso MJ, Schnelle PD, Rowan DA. Process and controller natitlán, Veracruz, Mexico, in 1967. He received the
performance monitoring: Overview with industrial applications. Inter- M.Sc. degree in electronic engineering from the Cen-
tro Nacional de Investigación y Desarrollo Tecnológico
national Journal of Adaptive Control and Signal Processing 2003;17:
(CENIDET), Cuernavaca, Morelos, Mexico, in 1993,
635–62. and the Ph.D. degree in process engineering from the
[5] Bogaerts P, Wouwer AV. Software sensors for bioprocesses. ISA Université Claude Bernard Lyon 1, France, in 2001.
Transactions 2003;42:547–58. Since 1993, he has held teaching and research positions
[6] Sotomayor OA, Park SW, Garcı́a C. Software sensor for on-line at the CENIDET, where he mainly works on automatic
estimation of the microbial activity in activated sludge systems. ISA control of energy conversion processes. His research
Transactions 2002;41:127–43. interests include the nonlinear observers, fault diagnosis, and control systems
[7] Astorga C-M, Othman N, Othman S, Hammouri H, McKenna TF. with applications, in particular to process engineering.
Nonlinear continuous-discrete observers: Application to emulsion
polymerization reactors. Control Engineering Practice 2002;10:3–13. Vı́ctor-Manuel Alvarado-Martı́nez was born in Mex-
[8] Dochain D. State observers for processes with uncertain kinetics. ico in 1968. He received the Ph.D. degree in auto-
International Journal of Control 2003;76:1483–92. matic control from the Laboratoire d’Automatique de
Grenoble, Institut National Polytechnique de Grenoble,
[9] Nadri M, Hammouri H, Astorga C. Observer design for continuous-
France in 2001. Since 1993, he works as a research en-
discrete time state affine systems up to output injection. European Journal gineer at the Centro Nacional de Investigación y Desar-
of Control 2004;10:252–63. rollo Tecnológico (CENIDET), Cuernavaca, Morelos,
[10] Bogaerts P. A hybrid asymptotic-Kalman observer for bioprocesses. Mexico, where he mainly works on energy conversion
Bioprocess Engineering 1999;20:249–55. processes control and modelling. His research interests
[11] Boutayeb M, Aubry D. A strong tracking extended Kalman observer for encompass identification methods, in particular, model
nonlinear discrete-time systems. IEEE Transactions on Automatic Control validation and robust digital control.
1999;44:1550–6.
[12] Luenberger DG. An introduction to observers. IEEE Transactions on Arturo Zavala-Rı́o was born in Merida, Yucatan,
Automatic Control 1971;16:596–602. Mexico, in 1967. He received the B.Sc. degree in
[13] Quintero-Mármol E, Luyben WL, Georgakis C. Application of an electronic systems engineering and the M.Sc. degree
extended Luenberger observer to the control of multicomponent batch in control engineering from the ITESM, Mexico,
in 1989 and 1992, respectively, and the Advanced
distillation. Industrial & Engineering Chemistry Research 1991;30:
Studies Diploma and Ph.D. degrees in automatic
1870–80. control from the Institut National Politechnique de
[14] Lüders G, Narendra KS. An adaptive observer and identifier for a linear Grenoble, France, in 1994 and 1997, respectively. He
system. IEEE Transactions on Automatic Control 1973;18:496–9. was a visiting researcher at the Mechanical Engineering
[15] Bastin G, Gevers M. Stable adaptive observers for nonlinear time varying Laboratory, Japan, from January to December 1998.
systems. IEEE Transactions on Automatic Control 1988;33:650–8. Since August 2001, he has held a Lecturer–Researcher position at the Instituto
[16] Marino R, Tomei P. Nonlinear control design. In: Geometric, adaptive and Potosino de Investigacion Cientifica y Tecnologica, Mexico. His research
robust. London: Prentice Hall; 1995 [chapter 5]. interests include the modelling, analysis, and control of nonlinear systems.