Copyright © IFAC Algorithms and Architectures for

Real-Time Control, Ostend, Belgium, 1995

ESTIMATION AND CONTROL OF FOULING IN HEAT EXCHANGERS

Peel, D., Wbeeldon, P., Virdee. G.S.

School o/Science and Technology

University o.fTeesside
Middlesbrough
Cleveland
UK
TS13EA

Abstract: The development of non-invasive sensors to indicate the presence and the degree
of fouling is described. This application relates primarily to fouling in the crude oil
refining industry although the approaches are generally applicable. Data driven models
derived using the standard linear least squares technique and artificial feed-forward neural
networks are used to predict the degree of fouling either directly or indirectly.

Keywords: Least-squares estimation; neural network models; chemical industry:

modelling; efficiency enhancement.
1.0 IN1RODUCTION availability of such a sensor can be useful in
providing evidence for fault diagnosis
Heat exchangers are designed, like most unit
operations of a chemical plant, to run continuously 2. THE PROCESS
at steady state. Like all unit operations. they do not.
In the case of heat exchangers. one of the more A small scale rig has been constructed to investigate
common problems is that of fouling. Fouling is a fouling of crude oil at conditions which are related
tenn which is given to the deposition of inert to the operating flows and temperatures experienced
material on the heat exchange surfaces. The fouling by the full scale industrial process. The rig which
material depends on the application to which the has been used to produce the experimental results
heat exchanger is put, however, the consequences demonstrated throughout this paper is shown below:
are generally consistent: with significantly reduced
Temp-4
heat transfer efficiency: reduced effectiveness of the
associated control systems: and increased down time
and difficulty in cleaning and maintenance.

The industrial interest in the development of a T~mp · 3

fonling estimator (<;ensor) is the rl~re to monitor

the perfonnance of the system as it fouls and thereby
judge when. where and how much anti-fouling agent
needs to be added to the process stream to maintain
a satisfactory le"el of perfonnance. The development
of a non-intrusi\'c fouling scnsor is a fundamental
requircmcnt for pcrfonnance monitoring.

This paper describes the de\'e!cpment and

comparison of a number of on-line fouling sensors
Tamp.' Flowrate
using methods based on linear least-squares
parameter estimation and 'standard' artificial feed-
forward neural networks. Additionally, the hg 1. f:.xpcT1m~ntal lUg showmg Instrunl~ntauon
Locations.

311
Items indicated Temp-l through to Temp-J are a heat transfer efficiency which proportioned the
temperatures which are recorded every thirty sensible energy used in raising steam to the total
seconds. as is the flowrate. HTX-I to HTX-3 energy entering the system. Unfortunately this
represent the heat exchanger units as shown in approach does not have a simple comparison in this
greater detail in figure 2. work due to the constant power input in the system
under consideration.
Inlet temperature ANNULAR GAP oullct tern)
Equation (I) introduces a fouling factor, Rr.
(faborek et al 1972) which is used in this study to
describe the amount of fouling.

(1)

where h is the film coefficient between the fouled

surface and the oil, "i arc the thicknesses of thc
layers in the composite wall between the heater and
the fouling surface in contact with the oil. le; are the
Oil t10wrate thermal conductivities of each of the layers. The only
Surface temperature
\-ariabks in this expression are the thickness of
fouling. "f. and kr- the thermal conductivity of the
Fig 2. Schematic of a Heat Exchanger clement fouling. Both of which are time, temperature and
operating condition dependent leading to a complex
time dependent, non-linear relationship.
During normal operation of this rig the effects of
fouling can be demonstrated over a period of 6-8 A typical set of data collected over a 12 hour period
hours whereas on the real process the equivalent will result in a fouling factor vs Time graph as
time can vary from weeks to months. This shown below.
'acceleration' was a design requirement of the rig as
was the low volume of oil neceSS,1ry to operate it.
The operating conditions of the experimental rig are
recognised as being significant by the collaborating
company.

2. 1 Process Operation

Tn the experimental rig each of the three heat

exchange units has a solid heater in the central core
with oil flowing in the annular space. The heaters
provide a power input of between 2-3 kW which is
then transferred to the oil through the inner tube
wall of the annulus and the fouling deposited on the
outer surface of this wall. The oil temperature
therefore rises as it passes through each of the heat
exchange units. In order to maintain the input oil Fig. 3. A Typical Variation of Fouling Factor
temperature constant (as measured at Temp-I), the
oil is cooled in the reservoir prior to its recycling It is possible to take 2 distinct approaches to this
through the system. Once these models have been problem. the first being to provide an estimate of the
established. their use will be directed towards the 'normal' Ran-fouled situation and to regard fouling
situation where anti-foulants are introduced to the as a process fault and commence with a 'fault
system. The estimators which are developed must detection' approach. or secondly. to attempt to
therefore take this proposed use into account. predict the effects of fouling directly. Additionally.
it is important that the model also provides evidence
with regard to the fouling mechanism. i.e. does
3.0 MODELLING APPROACHES increased fouling occur with prolonged operation at
high temperatures and low flowrates. This is a more
A simple and related study on the fouling of the constrained requirement in comparison with the
stream raising pipes in a coal-fired boiler. development of a fouling sensor.
(McMichael. 1988) based the measure of fouling on

312
3. I Fouling Estimation 4.0 RESULTS

The estimation of the fouling in the system has been Due to the relatively narrow range of temperatures
undertaken by anal~'sis of the data collected during which were used, the non-linearities of the system
operation. From this data. a range of linear and non- were not as pronounced as they would be on the
linear techniques have been used to produce fouling industrial units. Consequently, the linear methods
estimators. resulted in similar performance as the non-linear
results. However, as the operating ranges increase.
The overall heat transfer coefficient U. and the the non-linear estimators outperform the linear
thermal conductivity, k. are dependent on time, estimators.
flowrate and temperature(s). Consequently, "f. the
thickness of fouling is also a similarly dependent A range of fouling estimates using ANNs are sho\\n
variable. Although the system has been described as below:
time varying. the use of time as an input has obvious
problems as it is not clear how time itself can be
scaled in such a way as to be easily generalised.

Linear [,east Squares. Least squares methods for the F

estimation of coefficients of linear equations is a o
u
well understood practice. In this case equation (1)
1
describing the fouling factor is non-linear thereby
reducing the general applicability of this approach to n
this problem. g

Non-linear E...timation. The particular non-linear F

approach which was used was that of the Feed a
c
forward Artificial Neural Network. Although many t
structures of 'Neural Network' have been suggested, o
the F ANN has been chosen due to its relative r 0
simplicity in providing a non-linear structure
relating inputs to outputs. As the aim of the
investigation was to provide an estimation of degree
of fouling the increased burden of a 'relatively slow'
training method or a non exhaustive approach to
achieving the optimal architecture are issues which
FigA. Estimation ofRrfrom TC02 and Swface
are not discussed further in this paper.
Temperatw·e.
The mechanism which was used was to fix the
number of inputs and outputs to the network; vary
the number of hidden neurons in the one hidden
layer which was implemented and to look for a
minimum in the performance measure which would ~~
then indicate a satisfactory architecture. This
approach was supplemented by repeatedly trying the
~ J
(apparently) converged network on unseen data. The
final strucnlre which was used di!iplayed a minimum
; 01
in the performance measure for both the training F 04j
data and the unseen data . This added weight to the IQ3I
argument for deciding upon this stnlctl.1fC. t ~
o
r 0
Perfnrmance .\fe(lslIres. As both least square and
Rackpropagation embody thc principle of 00----------------------------------
minimising the sum of the squares of the prediction
errors then their performances can be directly
compared given that the error scales are equivalent
(ic. eithcr both scaled or both unsealed). Fig. 5. Estimation of Rrusing TeO:! and TCO j .

313
 displayed in fig . 7 can also be used to train a
network.

F
o
u
~
F~,
o
n um
g I I

I
F
a n
c g
1
o
r 0

Fig. 6. Estimation of Rrusing Te02 and Oil

Flowrate.

The results shown in fig. 4-6 highlight the extent to ... BiBIl ... Mud
which the fouling of the system can be estimated. Figure 8. Estimator Trained using TeO I , Te02 ,
These results form the basis for further analysis Element Temperature and Oil Flowrate.
relating to the sensitivity of the operating condition
to produce fouling . For example, the oil flowrate was Eliminating the tlowrate allows a similar estimation
approximately constant for the above results with a to be produced but these estimates present a better
slight variation due to the inherent system noise. average trend than that shown in figure 8.
However, a second series of experiments were
carried out where the flowrate was varied
considerably. A representative result is shown in
figure 7.

, ... ISmi ... Mad

Ii--- ~-~~-~~- Figure 9. Eslimalor bast:d on TeO I' Te0 2 alld
10 Element Temperature.
Figure 7. Variation in Oil Flowrate.
Figures 8 and 9 require one further question to be
answered. That is what is the effect of flowrate itself
/\5 an example of how the models C'.m be used to
on the fouling factor. The result shown in figure 10
:ield information regarding the underlying process
indicates that the answer is that the effect is
mechanisms. the ci.'lta which has been generated :md
negligible.

314
~ ~"
U 'ut!
"
I ;
I
1 \ 1
. /\
' ", I . \ . ... , ,I, ,.'.') ",\r
~
i •
I I • : 'I •
. ~. • - . . . . . .-.... I ' 1' t- • • .- t-+-............-.-·............~ .-. •
I j'- .....

n
I
I
9 , J

F 1 ,. ...
~~/
roi
J" " """ I

... Bhid ... taJej

Figure 10. Estimator Based on Oil Flowrate.

Further model generation is also underway using

physical measures of the amount of fouling in place
of the fouling factor. However the proposed use is
the same, that is to use these models to investigate
and quantify the effect of anti-foulants on the
operation of the process.

5.0 CONCLUSION

The results shown in this paper demonstrate the

usefulness of non-linear estimation techniques in the
development of an on-line estimator for heat
exchanger fouling. This approach has been
developed for use in evaluating the contribution of
anti-fouling agents during operation. The models
which have been produced are currently being used
as part of an ongoing study into crude oil fouling
alongside the development of a mechanistic model.

6.0 REFERENCES

McMichael. D.W.. (1988), "On-Line Fault

Detection: a system nonspecific approach" .
OeEL J 729/'88. Oxford University.

Taborek. 1.. Aoki. T.. Ritter. RB. . Palen, 1.W ..

Knudscn. 1. G.. (1972) "Predictive Methods for
Fouling Behaviour". Chemical En.~ineering
Progress. 68. 7.

315