2.4 Case study: the countercurrent heat exchanger
2.4.1 Models on various levels of detail
Oversimplied model
On the top levelthe dynamics ismodeledbya singlepair of perfectlystirred lumps
forming a so called heat exchanger cell. Each cell consists of two perfectly stirred
tanks with in- and outows. The two tanks are connected by a heat transfer area
between them.
Themodelequationsforasinglecellconsistofthetotalmassandenergybalances
for the hot and cold sides respectively. In order to connect the top level model to
theotheronesthecellindexj (j =1)isusedforthesinglecell. Constantmass (and
thusconstantvolumeV
jh
with constantdensities
jc ,
jh
)isassumed inthehotside
thereforethe totalmass balanceispresent onlyforthe cold side. The equationsare
asfollows:
dT
where the subscripts h and c denote the hot and cold sides respectively. T
hi
are the inlet and outlet temperatures, v
h
and v
c
are owrates, v
jl is
the leaking owrate, V
jc
and V
jh
are the volumes,A the heat transfer surface area,
c
pc
and c
ph
the specic heats,
c
and
h
the densities and U
j
is the heat transfer
coecient (abbreviated as HTC). Note that all of the variables and parameters
above are positive by denition.
With only a single cell in g. 2.1.a. describing the heat exchanger the outlet
temperatures T
ho and T
co
of the heat exchangerare the same asthat of the cell. In
this case only two state variables fT
co
;T
ho
g are considered with a single triplet of
equations(2.1)-(2.3).
In case of the simplest top level model all the model parameters including the
owrates v
h
and v
c
are constants and the inlet temperatures T
hi
and T
ci
are the
input ordisturbance variablesonly.
T ci
Figure 2.1: a. Single heat exchanger cell; b. Cascade model of a countercurrent
heat exchanger
Linear cascade model
Thecountercurrentheatexchangerismodeledasasequence ofcascadedcellsabove.
Athreecellmodelisused(seeing. 2.1.b)asamiddlelevelcascade modelbecause
it can describe the dynamics of the heat exchanger well enough for fault diagnosis
purposes for majority of the industrialequipments.
With the three-cell model three set of modelequations (2.1)-(2.3)are usedwith
the followingvariablerelationsinstead of Eqs. (2.4):
j =1;2;3; T
The modelabovecan beregarded asalumped version ofthe distributed
param-eter model of the heat exchanger, therefore the cells are associated with particular
spatiallocationsalongtheequipmentlength. Thestatevariablesoftheabovemodel
arethe temperatures ofthe coldand hot sidesand the coldside volumeinthe cells,
i.e.fT
g. Ifanequidistantinspacelumpingis
per-formed, i.e. the parameters A
j
and V
jh
are the same for allcells then heat transfer
coecients U
1
;U
2
;U
3
carry spatialinformation.
Here again, allthe modelparameters includingthe owrates v
h and v
c
are
con-stantsand the inlettemperatures T
hi and T
ci
are the inputordisturbance variables
only.
Bilinear cascade model (Nonlinear model)
The nonlinearity of process models is caused either by the convective ow when
the owrate is alsoa dynamic variable (bilineartype nonlinearity) or by the
chem-ical reactions and other mechanisms appearing in the source term of the balance
equations. In our case only the bilinear type nonlinearity appears caused by the
convective ow.
Inthebottomlevelthree-cellbilinearcascademodelthreesetofmodelequations
(2.1)-(2.3)canbeusedwiththe variablerelationsEqs. (2.5). Butnowthe owrates
v
h andv
c
are time varying variablesand areinput ordisturbancevariablestogether
with the inlet temperatures T
hi
and T
ci .
There are two types of faults considered: the deterioration of the heat exchanger
surfaceand leakingintheoutercontainer. Empiricalgrey-boxmodelisusedforthe
deteriorationand physical white-box modelis set up for the leakage.
Deterioration of the heat transfer area
Under normal operating conditions the HTC is constant or slowly decreasing due
toalayerof settledmaterialbuildingup on theheat transfer surface (see g.2.2.a).
In oldheat exchangers large pieces of the settled materialcan break away fromthe
surface, causing damage in the equipment. When the settled material breaks o,
the HTCs will undergo abrupt positive jumps. These jumps can be grouped into
two qualitativelydierent types.
1. Inthe beginningwhenthe jumpsstart tooccur,they aresmall andinfrequent
becauseof the smallpiecesof CaCO
3
that leave theheat transfer surface(see
g.2.2.b).
2. Then in the later stages the jumps become large and frequent when bigger
'stones'of CaCO
3
comeo (see g.2.2.c).
The usual industrial practice is to wash heat exchangers with some kind of acid
between certain time intervals. However, this washing procedure requires the heat
exchangerandpossibly otheroperatingunitstobestoppedand thusitmay become
a rather expensive operation. On the other hand, if some information is available
abouttheprocessofthechangingoftheHTC,thenweareabletoavoidunnecessary
stoppings and damage caused by the settled material coming o the heat transfer
surface, too. Thus, one of the main ideas in this work is to track the HTC in the
heat exchanger and detect abruptpositivejumps in it.
Leakage
When the outer container is leaking, there is an unmeasurable outow v
cl
on the
cold side from the heat exchanger (see Eq.2.3 and g.2.3) which may result in a
slowdecrease of the mass inthe outerphase, or- if the levelsare controlled, inthe
decrease of the outow from the equipment. Here it is assumed that the volume
ow rates v
h and v
c
are known and constant or measured, leaking can therefore be
detected via detection of a slow decrease in the cold side volume V
c
. Note that we
want to detect smallchanges in the cold side volume compared tothe initialvalue
ofthe volumeinthe heatexchanger. Thisisareasonableaimbecausethedetection
oflargeandquick changes(i.e. bigleaks)can beasimplertaskwithothermethods.
Moreover, there is no possibility to get spatial informationabout the leakage even
with our cascade model, since the liquid level on the cold side decreases uniformly
inthe whole heat exchanger. Thatis why itis importanttodevelop such a leakage
detectionalgorithmthatrequiresasfewtemperaturemeasurementdataaspossible.
H T C
t
H T C
t
H T C
t
a b
c
Figure 2.2: Three consecutive stages of the change of the heat transfer coecient:
a.Constant or slowly decreasing HTC; b.Smaller and infrequent jumps; c.Large
and frequent jumps
T hi T h o
T ci T c o
v h v h
v c v c
v cl T c o
Figure2.3: Leakage modelledas anunmeasurable outowfrom the cold side
Filteringis needed toavoidnumericaldierentiationof the noisytemperature
mea-surements, because the forthcoming parameter estimation methodsrequire the
ap-proximationof the rst and secondderivatives of certainmeasured signals.
Numerical dierentiation is sensitive to noise and round-o errors (particularly
whenthesamplingintervalisshort),andlargeerrorsinthedierentiationcanoccur,
whichmayleadtonon-robustidenticationmethods. Thatis,themethodsmayfail
or have a very slow convergence in the presence of noise. The question is how can
we avoidtaking derivatives.
For adiscrete time signaly(k) the so-called Æ operatoris dened as[55]
Æy(k)=
y(k+1) y(k)
t
s
(2.6)
where t
s
is the samplingtime.
Assume we have an nth order discrete time system in the delta-operator
(ob-tained fromacontinuous time system), i.e.
A(Æ)y(k)=B(Æ)u(k) (2.7)
where the highest order occurring in the A and B polynomials is n. This means
thatthe nthorderderivativeofthesignalsisinvolved. Oneobviouswayof avoiding
taking this derivative is to integrate both sides of the equation n times, that is we
lter the signals through the lter 1
Æ n
. The system equation(2.7) then reads
1
Æ n
A(Æ)y(k)= 1
Æ n
B(Æ)u(k) (2.8)
and we see that 1
Æ n
A(Æ) and 1
Æ n
B(Æ) are now polynomials in 1
Æ
, i.e. we are now
integratingthesignalsinsteadoftakingderivatives. Thelter 1
Æ n
isjustoneexample
of a possible lter, but we have to integrate at least n times in order to avoid the
derivatives which means that the lter must be at least of order n (see [55]). This
means that if a particular parameter estimation algorithmrequires the rst or the
second derivative of certain signals then a rst or second order lter is applied on
the measurement data respectively.
The ideaofthe InniteImpulseResponse (IIR) lterdesigninour case isthata
continuous time lter is designed rst using the continuous time model of the heat
exchanger andthen the obtained lter istransformed into discretetime usingthe Æ
operator.
The generalmethodforconstructingsucha lterisasfollows[55]. Assumethat
acontinuous time system isgiven by the following input-outputmodel
Y(s)
U(s)
= B(s)
A(s)
(2.9)
the orderof B(s)). Thenthe lter thatshould beused is 1
A(s)
. After applyingthis
lter onthe system weget
Y(s)= B(s)
A(s)
U(s) (2.10)
To obtainthe discretetime lter, we substitute s in(2.10) for the Æ operator.
On the basis of the guidelines above, the lters that will be used in the fault
detection methods are the following.
First order lter. In this case we want to use a rst order lter with cuto
frequency !
c
=e
0
. The transfer function describing sucha lter is given by
F
(s)are theLaplacetransformsofthe signaltobelteredandthe
lteredsignalitselfrespectively. The discretetimeversionof thislter iswrittenas
f
fromwhichweget
Æf
Tocalculatetheappropriateltercoecientinthiscase,letusselectfromEqs.(2.1)
and(2.2)theonethatdeterminesthedominanttimeconstantofoneheatexchanger
cell. Let us assume that this equation is Eq.(2.1), and that the cold side liquid
volume and the HTC in the heat exchanger are constant (i.e. we assume normal
operation). Let uswrite the Laplace transformof Eq.(2.1) inthe followingform.
T
According to the previously discussed lter design method, the rst order lter
coecient isas follows.
e
where V
jc
and U
j
denote the nominal values of the cold side liquidvolume and the
HTC in the jth cellof the heat exchanger respectively.
Second order lter. Again we start from the transfer function of a continuous
time second orderlinear lter whichreads
F
Converting it intodiscrete time gives
f
Æ
Thecalculationof thesecondorderlter'sparametersfromthe heatexchanger's
parameters willbe shown inSection 2.6.2.
2.6 Parameter estimation
2.6.1 Linear cascade model - all measurements available
In this ideal and not too realistic case we assume that we have access to both the
hot andcold side temperatures ineachcellof the heatexchangerinadditiontothe
inlet temperatures. In other words, we know the temperature distribution along
the heat exchanger on both the cold and the hot sides. Furthermore, itis assumed
that A, c
are known. It must be admitted that it is
not often possible to measure the cold and especially the hot side temperature on
several points alongthe heat exchanger. On the other hand,we have agoodreason
toexpect thatweobtainthe mosteasilyimplementablealgorithmsinthis casethat
produce the most reliableestimates forthe parameters.
Estimating the HTC
An additionalassumption inthis case is that the cold side volume is constant and
known in the heat exchanger since the purpose of the algorithm discussed here is
onlyto give anestimate forthe HTC.
Using the gradient method with a forgetting factor 0 < < 1, it is possible to
obtain estimates directly for both U
jc
and U
jh
using the discrete time modelof the
heat exchanger. The forgetting factor means that old measurements are weighted
at an exponentially decreasing rate. (see [51]). Usingthe ltered temperatures we
can easilywrite the discretetime modelof the jth cellin the heat exchanger.
ÆT
One can see from(2.19)and (2.20) thatthe only unknown parameter inthis model
is the HTC, i.e. U
j
that carries the diagnostic informationabout the deterioration
of the heat transfer surface in the jth cell.
The prediction errors for (2.19) and (2.20)in the jth cellare given by
j
U
j
(k) has tobecalculated
jc
Since we don't know the real value of U
j
we can approximate
jc
and
jh
in the
following way (see [51]).
^
The standardrecursivescheme for estimatingU
j
fromthe cold and hot side
respec-tively is asfollows
^
Thechoiceof in(2.28)and (2.30) isatradeo between goodtracking capabilities
of the HTCs and sensitivity tonoise and unmodeled dynamics. A small value of
meansthat the estimatewilltrack variations inthe HTC quicker, but onthe other
hand,the itwillalsobecomemoresensitivetonoiseand unmodeleddynamics. The
actual estimates of the HTCs in the jth cell can be computed by using a convex
combination:
where shouldreect the relative condence wehave inthe two estimates.
Estimating the cold side volume
Theonlydierencebetweentheassumptionsinthisandthepreviouscaseisthatnow
theHTC ineachcellisassumed tobeknownand constantandthecold sidevolume
becomes a time-varying parameter. In most cases this is a reasonable assumption,
sincethechangeoftheHTCisusuallymuchslowerthanthatofthecoldsidevolume
during leakage. According to these assumptions the discrete time form of the cold
side energy balanceequation of the jth cellis given by
ÆT
of the same size
where V
c
(k) denotes the overall cold side uid volume inthe heat exchanger and n
isthe number of cells the heat exchanger is divided into.
For the purpose of tracking the cold side uid volume let us introduce the
fol-lowing variable
V
Thuswe can rewrite (2.32)as
ÆT
Starting from (2.35) let us now present a possible identication scheme for the
trackingof the cold side uid volume.
Tracking with the gradient algorithm. Using (2.35) the prediction error is
writtenas
Its negative gradient with respect toV
jcr
The approximation of
j
(k) is writtenas
^
Using(2.38)and (2.37)therecursiveschemeforthe estimationofthe coldsideuid
volumecan bewritten as
^
is again a tradeo between quick tracking capabilities of the parameter
and sensitivity tonoise.
In this case both the HTC and the cold side volume are unknown parameters and
wewould liketoestimatethem simultaneously. An obviousstrategyto achievethis
goal is to combine the two algorithms described in Eqns.(2.23)-(2.31) and
(2.37)-(2.41) respectively. Within the jth heat exchanger cell we have two parameters
to be estimated in this case, namely V
jc
and U
j
. The idea of the solution is to
let the two gradient algorithms use each other's estimations. This approach works
very well if both V
jc
and U
j
vary slowly in time, but when abrupt jumps occur in
the HTC, problems arise with the estimation of the cold side liquidvolume. These
problems can be handled successfullyif weutilize allthe informationwehave from
the complete availabilityof measurement data.
The suitable form of the discrete time energy balance equation of the cold side
inthis case iswritten as
ÆT
while Eqn.(2.20) describing the hot side can be used here without change. After
combining the algorithms described in Eqns.(2.23)-(2.31) and (2.37)-(2.41) we get
thefollowingprocedureforthesimultaneousestimationofthecoldsideuidvolume
and the HTC inthe jthcell
V
The estimate for the cold side liquid volume in the whole heat exchanger is given
by the sum of the estimates of the cells, i.e.
^
wheren denotesthe numberofcells intheheatexchanger. Thetuningknobs ofthe
algorithmare
jUcc ,
jUch and
jVc
respectively.
2.6.2 Linear cascade model - only cold side measurements
available
Estimating the HTC
The algorithm used in this case is based on exactly the same idea as the one for
estimating the cold side liquid volume described in the next subsection. However,
only one of them can be presented in a detailed way due to the space limitations.
Usingthisapproachwecanapproximatelycalculatethehotsidetemperaturesalong
theheatexchangerandestimatetheHTCbyprocessingthemeasuredandcalculated
temperature data.
Estimating the cold side volume
In this section the identicationalgorithmwillbe extended to the case where only
measurementsofthecoldside temperaturesand thehotinlettemperatureare
avail-able. It is assumed that the HTC is constant. Furthermore, we assume that we
know the constant parameters of the heat exchanger (U, A, V
h
). Thus the only time-varying parameter which is to be estimated in order to
detectleaking isthe cold side liquidvolume,V
c .
Wecanconsideracascademodelconsistingofthreecellsforthesakeofsimplicity,
without the loss of generality. Let T
nc
denote the cold side output temperature in
cellno. n. Let usintroducethe following notations
Thusthe energy balanceequations take the form
dT
2h
The system can approximately be described by a linear time invariant model by
assumingthattheowratesandtheHTCareapproximatelyconstant. AfterLaplace
transforming (2.59) we get (for the sake of simplicity, variable s in the argument
of the Laplace transforms of v
c
will be suppressed in the following
equations).
Since, according to our assumptions, we do not have access to the hot side
tem-perature measurements along the heat exchanger (we can only measure the hot
inputandoutput), weneedtoeliminatethehot sidetemperaturesT
1h andT
2h from
Eqns.(2.59)-(2.64). In order todoit we express T
1h
The Laplacetransform of (2.60) isgiven by
sT
Writing (2.68) into (2.69)gives
sT
fromwhichit follows that
Similarly to (2.68) we can express T
2h
from the Laplace transform of (2.61) in the
following way
T
(s+v
Substituting(2.72) into (2.73)givesthe transfer functionfor T
2c
which reads
AftertakingtheLaplacetransformof(2.63)and(2.64),andexpressingT
3h
fromthe
Laplacetransformof (2.63) andsubstituting itinto theLaplace transformof (2.64)
we obtainthe following transfer functionfor T
3c
Notice the pattern inthe transfer functions (2.71),(2.74) and (2.75) as the number
ofcells grows. Next,itwillbe shown that (2.74)and (2.75)canbebroughtintothe
form(2.71) by introducingltered variables.
Let usintroduce the signal
T
Then (2.74)can be writtenas
which has exactly the same form as (2.71). The following ltered variable is
intro-duced
Now Eqn.(2.75)reads
which againhas the sameformas(2.71), andthe patterncontinues if weconsider a
model with more than three cells.
Our identication strategy is now to use a discrete time version of (2.71) to
obtainestimatesforV
1c
to(2.76) usingthe estimates of v
c and
1c
inthe implementationof the lter. Then
we use a discrete time version of (2.77) with the newly obtained ltered signal to
getanestimateof
2c
. Werepeatthe lteringandidenticationprocedureon(2.78)
and(2.79)toobtainestimateof
3c
. Becauseofthepatterninthetransferfunctions,
anarbitrary number of cells.
Inorder toobtainadiscretetimeversionof(2.71)wesubstitute swiththe delta
operator. Wethenobtain (forthe sakeof simplicitywe suppress thetime argument
of v
Toavoidnumericaldierentiationwelterallthetemperaturemeasurementsthrough
a second order linear lter as it was described in Section2.5. Recall that the lter
a second order linear lter as it was described in Section2.5. Recall that the lter