Slovak University of Technology in Bratislava

Institute of Information Engineering, Automation, and Mathematics

PROCEEDINGS
of the 18th International Conference on Process Control
Hotel Titris, Tatranská Lomnica, Slovakia, June 14 – 17, 2011
ISBN 978-80-227-3517-9
http://www.kirp.chtf.stuba.sk/pc11

Editors: M. Fikar and M. Kvasnica

Friebel, T., Haber, R., Schmitz, U.: Lifetime Estimation of Heat Exchangers with Consideration of On-line Cleaning,
Editors: Fikar, M., Kvasnica, M., In Proceedings of the 18th International Conference on Process Control, Tatranská
Lomnica, Slovakia, 434–439, 2011.

Full paper online: http://www.kirp.chtf.stuba.sk/pc11/data/abstracts/062.html

 18th International Conference on Process Control
June 14–17, 2011, Tatranská Lomnica, Slovakia Le-Th-3, 062.pdf

Lifetime Estimation of Heat Exchangers with Consideration of On-line Cleaning

T. Friebel* R. Haber* U. Schmitz**

* Cologne University of Applied Science

Institute of Process Engineering and Plant Design, Laboratory for Process Control,
Köln, Germany {thomas.friebel, robert.haber}@FH-Koeln.de

** Shell Deutschland Oil GmbH

Rheinland Raffinerie Godorf, Germany
ulrich.schmitz@shell.com

Abstract: In the presented paper two quality parameters are used to represent the state of a heat exchanger.
The remaining lifetime can be estimated by trend regression. Also of interest is the uncertainty of the
predicted lifetime which is determined by the confidence interval of the parameter estimation. These
algorithms developed are used in this paper in an off-line evaluation of the measurements on a heat
exchanger in a refinery. It is shown that the time point of the heat exchanger cleaning can be predicted. So
the presented method can be used for planning the cleaning time point in advance and saving money in
maintenance.

observation can start from any state of the heat exchanger. A

1. INTRODUCTION
great and important target of the conditioning monitoring is
In practice there is always a risk that the pipes of a heat to predict the time interval until a detected disturbance
exchanger become clogged with solid particles due to strong reaches a tolerance level and becomes a fault. The dwell time
temperature differences. During the operation it is not of the fluids in the heat exchanger is very small against a
possible to look into the interior of the heat exchanger. So the normal observation period. In addition, if there are complex
state of heat exchanger has to be monitored based on chemical and physical reactions it is probably not possible to
measurable quantities. Such measurements are temperature, predict the future with an exact model. As explained before
flow and pressure drop. These measures allow not only the the two quality parameters ε and UAF are observed and
description of the current state but the planning of the remaining lifetime is estimated by trend regression. The
maintenance in due time. Most methods, which are known in quality of the regression can be observed by using statistical
the literature, are based on models. The models can be tests like a t-test. Also of interest is the uncertainty of the
separated into two groups. In the first case, the models are predicted lifetime which is determined by the confidence
based on multivariate regression, PCA, neural networks and interval of the parameter estimation. The presented method is
so on. An advantage of all these methods is that they can be used with real measurements from an oil refinery. The target
used without detail knowledge about the inner states and is to predict the time points of the cleaning on-line. Thereby
chemical reactions in the heat exchanger. All these models the measurements are evaluated in on-line mode. Also it is
commonly need fault free training data to generate e.g. the shown that the quality parameter UAF allows a better
regression model. Also there are disadvantages. If the heat prediction than with the classical degree of efficiency ε .
exchanger leaves the normal working point, probably a neural
network becomes bad, or a fault with no effect on the used 2. METHODS AND THEORY
principal components will be not detected, because not all
possible conditions can be realized with a real plant. In the In the following a counter current heat exchanger will be
other case physical models based on inlet and outlet are used. dealt with. In the actual application the inlet and the outlet
There are several methods, which try to observe the inner flow are coupled on thermal side, see Fig. 1. The cold
state of the heat exchanger. Using complex physical models reactant with temperature T1E enters the heat exchanger on
can cause problems with the generalization far from the the cold side. It is preheated by the product flow and it leaves
working point. However, if there is only one question: “Is the
the heat exchanger with temperature T2 E on the hot side. The
heat exchanger in a normal condition?” a simple model can
be used. In this paper, two quality parameters both based on hot product enters the heat exchange with temperature T2 P
an easy physical model are compared. The first one is ε the on the hot side. The fluent is cooled down by the reactants
degree of efficiency and the second one is UAF a and it leaves the heat exchanger with temperature T1P on the
combination of the heat transmission coefficient with the cold side. In the following chapters both quality parameters
inner surface and the flow. Big advantages of these methods
are that they can be used without any training data and the

434
 18th International Conference on Process Control
June 14–17, 2011, Tatranská Lomnica, Slovakia Le-Th-3, 062.pdf

for the monitoring of the heat exchanger are presented. After In (2) Q stands for the heat flow, U for the coefficient of
that the method for rest live time prediction is shown. heat transmission, A for the surface of the heat exchanger and
ΔTlog for the logarithmic mean temperature difference. The
logarithmic mean temperature difference is defined in (3).
Increasing fouling leads to a decreasing coefficient of heat
transfer due to additional heat resistance; see (4).
ΔT1 − ΔT2
ΔTlog = (3)
 ΔT 
Fig. 1. Simplified flow diagram of the plant ln 1 
 ΔT2 
1
U= (4)
1 s wall sdep 1
+ + +
αi λ λdep αa
In (4) α denotes the heat transfer coefficient, s the coat
thickness and λ the heat conductance coefficient. The index
i stands for the inner and a for the outer side of the pipe
Fig. 2. Temperature profile of a current flow heat exchanger and dep denotes biomass coat or solid deposition. The heat
with tube length L
flow Q can be calculated from the measured process para-
meters by (5) see Friebel et al. (2009) and Wagner (2005).
2.1 Quality parameter – degree of efficiency
Q = FE ⋅ c pE ⋅ ΔTE (5)
Fig. 2 shows a typical temperature profile over a counter
current heat exchanger. The state of the heat exchanger can In (5) FE stands for mass flow of the reactant, c pE :for the
be described by the ratio of the actually transferred energy heat capacity of the reactant and ΔTE for the temperature
and the maximum transferable energy. The actual amount of
transferred energy is proportional to the temperature difference from outlet to inlet of the reactant, see Fig. 2. By
difference between inlet and outlet of the reactant, combining (2) and (5) the quality parameter UA can be
ΔTE = T2 E − T1E . The maximum transferable energy is defined as shown in (6) see Friebel et al. (2009).
proportional to the temperature difference between the fluid ΔTE
UA = FE ⋅ c pE ⋅ (6)
inlet temperatures, T2 P − T1E . For this definition the ΔTlog
assumptions ΔTE > ΔTP and ΔT1 > ΔT2 are used. The Simulations presented in Friebel et al. (2009) show that the
ratio of actual and maximal energy transfer is given by (1), quality parameter UA is sensitive for fouling. This is a big
where ε is the degree of efficiency see Wagner (2005). advantage against the degree of efficiency because fouling
T2 E − T1E can be distinguished from model input drift. This means the
ε= (1) degree of efficiency is sensitive for
T2 P − T1E
• a drift in one or both of the inlet temperature,
The efficiency depends on the set point and the inner state of • a drift in one or both of the fluid flows and
a heat exchanger. The set point is defined by the inlet • a drift in the model parameter UA i.e. in the heat transfer.
temperatures and the rate of fluid flow. The inner state
By analyzing the above listing it is clear that it is nearly
depends on the fouling. Therefore in the present paper nearly
impossible to differ among all possible combinations of
steady-state conditions are assumed for the inlet temperatures
drifts. Therefore, it is more practical to estimate the model
and the amount of fluid flow. In this case a change in the
degree of efficiency is caused by fouling. parameter UA at every steady-state sampling point k
according to (6). In the special case where the flow of the
reactant FE cannot be measured the flow and c pE are
2.2 Quality parameter – heat transfer coefficient
assumed constant. Then (6) can be transformed to (7)
As discussed above the quality parameter (degree of
efficiency) depends on the set point and the inner state of the ΔTE
UAF = (7)
heat exchanger. Therefore an additional formulation will be ΔTlog
used. The model equation for heat transfer can be written as
(2) see Wagner (2005). It is important to see that the parameter UAF is sensitive to
fouling and disturbances in the flow.
Q = U ⋅ A ⋅ ΔTlog (2)

435
 18th International Conference on Process Control
June 14–17, 2011, Tatranská Lomnica, Slovakia Le-Th-3, 062.pdf

2.3 Extrapolation of a regression model The principle of the calculation is shown in Fig. 3. The
calculated uncertainties in time sm and sp have different
Above a proper quality parameter for the heat exchanger is
defined. The last question is: how long is the rest lifetime values. For an easier interpretation in the practical use a
until a critical level is reached. Therefore a linear discrete- middle uncertainty s is defined in (10)
time regression model is used, see (8). Here uk is the sm + sp
s= with sm < sp (10)
independent variable and yk the dependent variable, see 2
Montgomery et al. (2001). In the following application the
actual measurement is the best representation of the inner 3. APPLICATION
state of the heat exchanger. Therefore the point of origin is
equal to the actual measurement. The regression model 3.1 Problem description
contains only the unknown parameter ĉ1 .
As it is seen in Fig. 1 the reactant is preheated by the product
y k = c1u k = yˆ k + ek = cˆ1u k + ek (8) and the product is cooled by the reactant. In Fig. 4a the tubes
of a cleaned heat exchanger are shown. In Fig. 4b the
With a simple extrapolation the rest lifetime upred can be problem with fouling, adhesion on the surface inside and
outside the tubes is shown. On the right side in Fig. 4c the
calculated, while the regression model ŷ pred is equal to the totally blocked tubes can be seen. Blocked tubes cause the
tolerance limit ytol . As it was shown in Friebel et al. (2009) following problems:
the quality of the regression can be proven by a statistical t- • pressure drop over the heat exchanger increase
test. The uncertainty of a regression can be shown by his • the maximal cooling power of the heat exchanger
confidential intervals. With the assumption, that the decreases
uncertainty at the actual measurement is equal to zero the • product have to be cooled additional before entering the
confidential limits for future measurements can be calculated storage
by (9) see Montgomery et al. (2001) • reactant have to be heated additional before entering the
production unit
σ2 Fouling costs some money. Normally a heat exchanger is
yˆ k ± sreg = yˆ k ± uk ⋅ tα ⋅ (9) observed by the degree of efficiency ε . To prevent the above

N −1 2
2
; N −1
u
k =0 k listed problems the quality parameters UA and UAF are
Here σ is the residual error, N the amount of used data used.
with the regression and t the value of a t-distribution for a
given significance level α and N − 1 degrees of freedom. 3.2 Problem solution
By setting the left term in (9) equal to ytol and solving the Typical measurements (temperature) are shown exemplary in
equation the predicted rest lifetime uk = upred can be Fig. 5. In this analysis, several years are taken into account,
but only some examples are shown in this paper. Therefore,
calculated with its uncertainty s reg . the discrete time k does not start at one. The period starts and
ends with a cleaning of the heat exchanger, all temperatures
drift y are low. The cleaning was performed if the degree of
efficiency reached a value of e.g. 90%. Now the target is to
predict these cleaning time point in order to plan a cleaning in
uncertainty of the ypred advance. The temperatures of the production process are in a
predicted lifetime range between 350 and 450 °C see Friebel et al. (2010). The
tolerance sreg reactant is drawn with solid lines and the product in dashed
limit ytol sreg lines.
uakt actual
measurment extrapolated
sm linear
sp regression

upred
a) b) c)

used data rest lifetime time u

Fig. 4. Problems with fouling in a tube bundle heat exchanger
a) clean surface and tubes, b) surface with fouling and
c) blocked tubes after fouling see Friebel et al. (2010)
Fig. 3. Estimation of the rest lifetime

436
 18th International Conference on Process Control
June 14–17, 2011, Tatranská Lomnica, Slovakia Le-Th-3, 062.pdf

Fig. 5. Temperature measurements of the heat exchanger for

a time period between the plant revisions

It can be seen, that the product temperatures (dashed lines)

increase over the time. Also the temperatures are noisy
especially the product outlet T1P and the reactant outlet T3 E .
The reason is a periodical sinusoidal disturbance with a
period of approximately 11 days, which is caused by the plant
management. The problem is that the amplitude is not
constant and also there are some stepwise phase shifts in the
periodical signal. In Fig. 6 the degree of efficiency ε and the
model parameter UAF calculated by (1) and (7) are shown
for an interesting part of this time period.

Fig. 7. Regression based lifetime estimation

a) 19, b) 17, c) 15, d) 13 e) 11 and f) 9 weeks before
shutdown.

It can be seen, that the calculated degree of efficiency ε is

very noisy. The sinusoidal disturbance is clearly visible. But
by observing the model parameter UAF these disturbances
are eliminated. In the model parameter some additional
information can be detected. There are two steps which are
marked with arrows. These steps are caused by not recorded
technological handlings. It is not possible to detect both steps
in the degree of efficiency. Because this signal is caused by
periodical disturbances and the parameter is not so sensitive
for this case. Therefore it is a good idea to use the model
Fig. 6. Example for calculated
parameter UAF instead the degree of efficiency ε .
a) degree of efficiency ε and b) model parameter UAF

437
 18th International Conference on Process Control
June 14–17, 2011, Tatranská Lomnica, Slovakia Le-Th-3, 062.pdf

Fig. 9. Model parameter UAF with a step caused by online

cleaning in a) and the corresponding rest lifetime
estimation in b)

Fig. 8. Two observation periods; a) and c) degree of

efficiency (solid line) and filtered signal (dashed line),
b) and d) predicted rest lifetime tpred versus real time treal
until shutdown

In Fig. 7 the calculated efficiency for another example is

presented. The regression was carried out with 50
measurements in the on-line mode. Fig. 7a shows the
regression 19 weeks before the shutdown. The following
figures show in turn the regression always two weeks later.
The rest lifetime is predicted with a small uncertainty,
because the interval limits on the tolerance level are close
together. As a conclusion of Fig. 7 the following points can
be marked out, see also Friebel et al. (2010).

• The rest lifetime t pred could be predicted nearly exact Fig. 10. Model parameter UAF for the time period from Fig.
several weeks before the shutdown. 9 with considered step in a) and the corresponding rest
• With a t-test it could be shown, that the used regression lifetime estimation in b)
model is always significant.
• The corresponding significance values α are nearly 0 %. prediction of rest lifetime in Fig. 8d. The trend is clearly
• The confidence limits lie near to the predicted visible along the diagonal line. Because of the noisy signal
rest lifetime t pred . and the technological handling there are changes in sign of
the slope.
In Fig. 8 two additional periods are analysed. The time until
the next realized shutdown t real is shown on the horizontal The calculated model parameter UAF is sensitive for any
not recorded technological handlings. During such a
axis. In Fig. 8a the filtered and calculated heat exchanger procedure some cleaning solution is added to the reactant.
efficiency is shown for a further period between two cleaning The result is a shortly better heat exchanger condition (higher
cycles. It can be seen, that the signal is not very noisy. In Fig. value for the model parameter), because the amount of solid
8b the predicted rest lifetime t pred is plotted on the vertical depositions decreases and the heat transfer increases.
axis. In the ideal case the times t pred and t real are equal, Fig. 9a shows the last section of the data form Fig. 6b.
which is marked by the diagonal line. Two month before the Around day 2840 a step in the model parameter UAF was
shutdown an acceptable prediction is possible. In the second detected, see the arrow. The estimated rest lifetime is shown
period in Fig. 8c it can be seen that calculated degree of in Fig. 9b. It is easy to see that around the step in the model
efficiency temporally increases because of a not recorded parameter the estimated rest lifetime becomes infinite large.
technological handling. This has a direct impact on the In the period marked by the arrows no practical prediction is

438
 18th International Conference on Process Control
June 14–17, 2011, Tatranská Lomnica, Slovakia Le-Th-3, 062.pdf

possible. 100 and 80 days before the shut down the prediction
were disturbed by additional changes in the inlet
temperatures.
In the new approach all regressions before the stepwise
disturbance are calculated with the data from Fig. 9a. For all
regressions after the stepwise regression the values before the
disturbance are shifted upwards that the step in the model
parameter disappeared. The new course of the model
parameter is shown in Fig. 10a and the corresponding
estimation of the rest lifetime is shown in Fig. 10b. It is easy
to see that in this case in every time point a practical
prediction is possible. It is important to know that Fig. 10a
shows the model parameter for the view after day 2840. Fig.
11 shows all data from Fig. 6b and also at the first step at day
2720 a proper prediction is possible. The values in Fig. 11d
are not infinite high; they are smaller than 1900 days.

4. CONCLUSION
In Friebel et al. (2009 and 2010) a simple method for lifetime
estimation was presented and analyzed in some case studies.
In the presented paper an additional representative parameter
for the state of the heat exchanger is used with the explained
method. The assumptions and simulations were tested by
analyzing measurements of a real plant. By comparing the
results and simulations the following results can be
summarized.
• The classical parameter for the observation of the state of
the heat exchanger is the degree of efficiency ε . This
quality parameter is sensitive for drifts in the inlet
temperatures, the fluid flows and the heat transfer
coefficient.
• The model parameter UA , a combination of the heat Fig. 11. Model parameter UAF for the time period from Fig.
transfer coefficient and the inner surface of the heat 6b with considered step in a) and the corresponding rest
exchanger, is not sensitive for a drift in inlet temperature lifetime estimation in b)
and fluid flows.
• If the flow cannot be measured then only the observed REFERENCES
parameter UAF can be calculated. Hereby a constant flow
is assumed, otherwise a drift in flow and in the heat Friebel, T., Haber, R., Schmitz, U. (2010). Cleaning time
transfer cannot be differed. point prediction with heat exchanger fouling in a
• Using the model parameter UAF is better than using the refenery. Proceedings of 9th Int. Conf. Process Control,
degree of efficiency ε . Kouty nad Desnou, Czech Republic
• It would be better to use a quality parameter which is Friebel, T., Stockmann, M., Haber, R. (2009). Heat
independent of the working point. Therefore the universal exchanger fouling detection and lifetime estimation by
model parameter UA should be preferred if possible. regression. Proceedings ICONS 2009 - 2nd IFAC Int.
• The prediction of the rest lifetime can be made by a simple Conf. on Intelligent Control Systems and Signal
linear regression. The quality of the regression can be Processing, Istanbul - Turkey
proven with a statistical t-test. Also a statistical based Montgomery, D., Peck, E.A., Vining, G.G. (2001).
uncertainty of the predicted rest lifetime can be Introduction to linear regression analysis. Wiley and
formulated. Sons, 2001, 3th edition, 44ff
• It is easier to detect and compensate the on-line cleaning in
Pearson, R.K. (2002). Outliers in process modeling and
the model parameter UAF than in the degree of identification. IEEE Trans. on Control Systems
efficiency ε . Technology, 10, 1, 55-63.
• By considering the steps caused by the on-line cleaning,
Wagner, W. (2005). Wärmetauscher - Grundlagen, Aufbau
the predicted rest lifetime becomes more practical with
realistic predictions. und Funktion thermischer Apparate (Heat exchangers -
Fundamentals, construction and function of thermal
Further research work is planned in order to detect and to apparatus). Vogel Buchverlag, 3rd edition, 194ff
consider the change in flow during the data recording.

439