International Journal of Heat and Mass Transfer 116 (2018) 931–939

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

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

Experimental transfer functions identification: Thermal impedance and

transmittance in a channel heated by an upstream unsteady volumetric
heat source
Waseem Al Hadad ⇑, Denis Maillet, Yves Jannot
Université de Lorraine, LEMTA (UMR 7563), ENSEM, 2 Avenue de la Forêt de Haye, BP 90161, 54505 Vandœuvre-lès-Nancy cedex, France
CNRS, LEMTA (UMR 7563), BP 90161, 54505 Vandœuvre-lès-Nancy cedex, France

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

Article history: Linear Time Invariant (LTI) systems can be modeled by transfer functions linking inputs to outputs
Received 12 May 2017 through convolution products. This concept is experimentally validated in the case of an upstream heated
Received in revised form 15 September channel with thick walls and steady liquid flowrate using two calibration experiments with different
2017
unsteady upstream heating powers. Proper identification of transfer functions (impedances and transmit-
Accepted 19 September 2017
tances) requires the inversion of experimental temperature measurement through regularization tech-
niques. Truncated Singular Value Decomposition (TSVD) and Zero-order Tikhonov regularization are
tested and compared for this application.
Keywords:
Transfer function
Ó 2017 Elsevier Ltd. All rights reserved.
Experimental identification
Convolution
Unsteady heat transfer
Function estimation
Conduction and advection
Conjugate heat transfer

1. Introduction (conductivities, heat capacities, emissivities, etc.) as well as the

velocity field in cases where forced convection is present, espe-
Considerable efforts have been made during the last few dec- cially in the turbulent case.
ades for improving numerical simulations in every branch of engi- We focus here on the second point, which relates to the
neering, and in thermal engineering in particular. When the unsteady simulation of heat transfer in processes occurring in a
physical laws of conduction, convection and radiation are the only furnace or in a heat exchanger for example. In these types of
ones at stake, which is often the case for material systems at the devices, the thermophysical parameters of the different con-
mesoscopic scale for non-coupled problems, direct simulation of stituents of the system are often not accurately known and their
heat transfer can be made, for increasingly large systems, if prior precise geometry may be difficult to represent in a numerical
pertinent assumptions have been set. It stems from the numerous model. If a flowing fluid is present in the system, the real velocity
numerical methods that are now available for solving the equa- fields may be difficult to retrieve numerically, because of this lack
tions of the model as well as from high-performance computa- of geometrical precision: this is the true for example in the case of
tional means, with ever increasing processing speeds and digital a mal-distribution of the flow in the outlet sections of a header in a
storage capacities. heat exchanger [1].
However, at least two bottle-necks remain: first, the writing of For all these reasons detailed numerical modeling can be sup-
the boundary conditions, that allow to compensate for our igno- plemented by another experiment-fed method, that is model iden-
rance of what happens outside the studied geometrical domain, tification. In this type of approach, originally developed in the
and, second, the lack of knowledge about the precise values of control science community [2], measurements of both inputs and
the thermophysical parameters of the media inside this domain outputs of a real dynamical system allow to construct a reduced
model. In this kind of model, the whole field, here temperature,
is not modeled: it is observed at locations of interest only, where
⇑ Corresponding author at: Université de Lorraine, LEMTA (UMR 7563), ENSEM, 2 sensors are present in a calibration experiment corresponding to
Avenue de la Forêt de Haye, BP 90161, 54505 Vandœuvre-lès-Nancy cedex, France. a given source history, (e.g. thermal power input).
E-mail address: waseem.al-hadad@univ-lorraine.fr (W. Al Hadad).

https://doi.org/10.1016/j.ijheatmasstransfer.2017.09.079
/Ó 2017 Elsevier Ltd. All rights reserved.
932 W. Al Hadad et al. / International Journal of Heat and Mass Transfer 116 (2018) 931–939

In this configuration, it is possible to look for a reduced model Our paper is organized as follows: in Section 2 we present the
using either the output point temperatures of a detailed model experimental bench (half heat exchanger) and the corresponding
(model reduction) or the corresponding measured temperatures measurements. In Section 3 we show how the transfer function
to construct a reduced model of the same structure but of consid- can be identified whatever the geometry. At the end of this section,
erably smaller size (model identification). Such an approach is pos- we show that experimental identification is not always stable. In
sible through the Modal Identification Method [3,4], with Section 4 we show how we move from an unstable identification
applications, once the reduced model is constructed, to heat source to a stable one through a regularization technique. In Section 5
estimation in a so-called Inverse Heat Conduction (or convection) we validate this methodology on an experimental basis: the iden-
Problem for example [5]. tified transfer functions obtained by regularized inversion are com-
We will consider here the specific case of advection and diffu- pared for two different time scenarios of the heating power.
sion in a conjugated way in a real system composed of the flowing
fluid and of its solid walls, with upstream heating and only two
points of measurement, in the inlet and outlet sections of the flow. 2. Experimental setup
Here, if the velocity field is steady, if the initial temperature field,
before the heating starts, is at steady state and if the variation with A sketch of the experimental setup is presented in Fig. 1. This
temperature of the thermophysical parameters of both fluid and setup is composed of a channel in which different transfer func-
walls is neglected, the relationship between the inlet and outlet tions are to be identified, of a syringe pump (to impose the flow
temperatures of the fluid and the upstream unsteady heat source rate of the used fluid), of a heating and cooling bath thermostat
is a convolutive one, because the Partial Differential Heat Equation (to impose the fluid inlet temperature at the upper plenum cham-
and its associated boundary and interface conditions in this non ber inlet) and of an electrical resistance (heat source). The channel
homogeneous material system is linear with invariant coefficients is constituted by two polycarbonate plane walls of thermal con-
in time (Linear Time Invariant = LTI), but not in space. ductivity ks and volumetric heat capacity qcs which are separated
So transfer functions, a tool widely used in control can be used by a water flow of thermal conductivity kf and volumetric heat
for reduction or for identification of models. These are either impe- capacity qcf . Two plenum chambers of larger thicknesses are
dances relating a local temperature to the cause of its variation, the located upstream and downstream of the constant rectangular sec-
unsteady heat source, or a ‘‘transmittance” linking a local temper- tion channel which is equal to wef , where ef is its thickness and w
ature to another one whose variation is prior. In that last case, the its width in the dimension normal to this figure. The total length of
prior variating local temperature will be called a ‘‘pseudo-source” the channel, which corresponds to the distance between these two
here, since it is also the consequence of the common heat source. chambers is, 2‘c , while the length of its part where the thickness of
In a previous work [6] this approach was validated numerically: its walls stay constant is, 2‘, which is considered here as the dis-
the corresponding analytical transfer functions could be found by tance between the inlet and the outlet of the channel. Between
numerical inversion of its analytical expression in Laplace domain, the inlet and the outlet of the channel, the thicknesses of the walls
or by the solution of a linear inverse problem where the data are e1 and e2 , see Fig. 1. An electrical foil heater has been set at the
(inputs and outputs) were generated through a commercial Finite exit of the upper plenum chamber. It creates a heat source of inten-
Elements Code (Comsol). sity: Q ¼ RI2 , where Q is the dissipated heat power, R the electrical

eh
Upper plenum chamber
h

Electrical resistance

Tin (t )
2
2 c

Tout (t )

Lower plenum chamber

Thermostat
Syringe
Polycarbonate channel

Fig. 1. Sketch of experimental setup.

 W. Al Hadad et al. / International Journal of Heat and Mass Transfer 116 (2018) 931–939 933

Table 1
Geometric parameters.

e1 ¼ e2 ðmmÞ ef ðmmÞ 2‘ ðmmÞ 2‘c ðmmÞ w ðmmÞ ‘h ðmmÞ eh ðmmÞ

2 1 65 120 50 10 0:3

Table 2 3. Identification of the transfer function

Thermophysical properties of each medium.
As we have shown in our previous work [6], as long as the stud-
ks (W m1 K1 ) qcs (kJ m3 K1 ) kf (W m1 K1 ) qcf (kJ m3 K1 )
ied system is LTI (linear and time invariant), the output (response
0:215 1440 0:63 4186 or consequence) y at a given point, or its average in any subdomain
of the system, is a convolution product between the input (excita-
tion or cause) u and a corresponding thermal transfer function H:
resistance of the heater and I the current. This thermal power can yðt Þ ¼ Hðt Þ  uðt Þ ð1Þ
be considered as a distributed heat source of volumetric density
where  denotes the convolution product. The integral form of the
P v ¼ Q =V, with V ¼ ‘h weh , where ‘h ; w and eh are respectively
convolution product (1) can be defined in two different ways, using
the length, the width and the thickness of the electrical resistance.
its commutative property [7]:
The evolution of the inlet and outlet bulk temperatures are mea-
Z t
sured by two thermocouples of very small diameters
yðtÞ  Hðt  t 0 Þ uðt0 Þ dt 0 ð2aÞ
(d ¼ 50:8 lm) to avoid disturbing the fluid flow. They are located 0
at the inlet and outlet of the channel as shown in Fig. 1. The geo- Z t
metric parameters and the corresponding thermophysical proper-  uðt  t0 Þ Hðt 0 Þ dt 0 ð2bÞ
0
ties are given respectively in Table 1 and in Table 2.
Two experiments were made with two different temporal forms Parameterization of the excitation uðtÞ, over a base of piecewise
of the volumetric heat source. In the first experiment, a constant constant functions defined on a constant time step 4t ¼ tf =N t ,
thermal excitation (step function) has been applied between times where tf is the duration of the thermal excitation and N t the num-
t ¼ 0 to t ¼ tf . In the second experiment, a piecewise constant peri- ber of time steps, allows a sampling yk ¼ yðtk Þ of the response, for
odic thermal excitation (square periodic function) with a 129 s per- times tk ¼ k 4 t (for k ¼ 1 to N t ). This can be put under a column
iod has been applied between t ¼ 0 to t ¼ tf . Initially, in both vector/matrix form:
experiments (before starting the volumetric heat source), the con- y ¼ M ðH Þ u ð3aÞ
dition of uniformity of the initial temperature was verified in the
¼ M ðuÞ H ð3bÞ
solid and in the fluid, such as Tðx; y; t 6 0Þ ¼ T 1 so
hðx; y; t 6 0Þ ¼ 0 with hðx; y; tÞ ¼ Tðx; y; tÞ  T 1 , where T 1 is both where
the external ambient temperature and the temperature of the ther- 2 3
w1 0 0  0 2 3
mostat. This means that the only source of heat is the volumetric
6 7 w1
heat source. 6 .. .. 7 6 7
6 w2 w1 0 . . 7 6 w2 7
The evolution of the imposed electric power, Q from t ¼ 0 to 6 7 6 7
6 .. .. 7 6 w3 7
t ¼ tf is plotted in Fig. 2 for the step excitation (first experiment) MðwÞ ¼ 4t6 w3 w2 w1 . . 7 with w ¼ 6 7
6 7 6 .. 7
and in Fig. 3 for the periodic excitation (second experiment). In 6 .. .. .. .. .. 7 6 7
6 7 4 . 5
the same figures, the corresponding evolutions of the bulk temper- 4 . . . . . 5
wNt
atures at the inlet hin out wNt wNt 1 wNt 2    w1
b and at the outlet hb of the channel are
shown. The parameters of these experiments are given in Table 3. ð4Þ

25 4

3.5
20
3

15 2.5

2
10
1.5

1
5
0.5

0 0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

Fig. 2. Step excitation and the corresponding measurement of inlet and outlet bulk temperatures.
934 W. Al Hadad et al. / International Journal of Heat and Mass Transfer 116 (2018) 931–939

20 6.5
6
5.5
15 5
4.5
4
3.5
10
3
2.5
2
5 1.5
1
0.5
0 0
0 1000 2000 3000 4000 5000 6000

Fig. 3. Periodic excitation and the corresponding measurement of inlet and outlet bulk temperatures.

Table 3
Parameters of the experiments.

Q (W) V m (m s1 ) T 1 (°C) 4t (s) tf (s)

Step 3:6161 4:4  104 20:3 1 5000

Periodic 6:4286 4:4  104 21:5 1 5950

and inverse problem whose formal solution, an estimate Z ^ of vector Z

Z derives from application of Eq. (3b) applied to this experiment:

tk
1
wk ¼ wðtÞ dt for w  u or H ð5Þ ^ c ¼ ½M ðQ Þ1 hc
Z with c  in or out ð6Þ
4t t k1 b

Let us note that this calibration problem constitutes here a linear

Here MðwÞ is a square matrix of size N t  N t (a lower triangular Toe- inverse problem, in the case where no noise is present in the heat-
plitz matrix) with w  H or u. Let us note that Eqs. (3a) and (3b), the ing power Q , that is if the system matrix MðQ Þ is not a stochastic
numerical quadrature of the time-continuous convolution product one.
(1), is valid only if the time step is small enough with respect to
the characteristic times of both input uðtÞ and transfer function 3.2. Transmittance identification
HðtÞ.
In this work, since the velocity field is stationary and with the In the case where the two quantities, excitation and its
assumption that the thermophysical properties remain indepen- response, have the same physical unit, the corresponding transfer
dent of temperature and, as a consequence, of time, the considered function H is called a transmittance here, noted W. Experimentally,
detailed model (conduction and advection) can be considered as in some applications it is easier to measure a temperature than a
LTI and the convolution model (3a and 3b), where y is a tempera- thermal heating power. Here the inlet bulk temperature hin b is a
ture h and u either the heat power or another temperature, remains response to the heating power Q but conversely it can be consid-
valid. In the next two subsections we will show how to adapt this
ered as a pseudo-source for the outlet bulk temperature hout
b . There-
model to identify impedances and transmittance.
fore by analogy to the identification of the impedances shown
above, the transmittance W connecting the pseudo-source hin
b to
out
3.1. Impedance identification Z its response hout
b , noted W in here can be identified as:

The transfer function H that connects the excitation or cause

(the thermal power input) to its response or consequence in tem- Z out (t )
perature, is generally called an impedance, noted Z. This term is
commonly used in electricity, where u is an alternating current
and y the corresponding potential difference between two points Z in (t )
of the circuit, but it can be extended to unsteady and transient s
regimes for LTI systems in other fields in physics, using a Laplace, θ bin (t ) e1 θ bout (t )
and not a Fourier, representation. This temperature may be the Q Vm ef f
temperature in the solid or in the fluid. We are interested here in e2
the inlet and outlet bulk temperature responses, hin out s
b and hb as
shown in Fig. 4. The impedances Z in and Z out can be identified using
the convolutive model (3b) and by measuring both the source Winout (t )
(heating power Q) and its response (inlet and outlet bulk temper-
atures hin out
b and hb respectively). This calibration problem is an Fig. 4. Sketch of polycarbonate channel.
 W. Al Hadad et al. / International Journal of Heat and Mass Transfer 116 (2018) 931–939 935

h  i1
^ out ¼ M hin
W hout ð7Þ 4.1. Truncated Singular Value Decomposition (TSVD) regularization
in b b

The identification of the impedance, Eq. (6) and transmittance, Instead of keeping all the N t singular values Ri of M ðuÞ in the
Eq. (7) is not always stable because of the presence of noise in OLS identification (9), only the a first largest singular values Ri
the measurements, especially when M ðÞ is very ill-conditioned are kept unchanged in R, while the remaining (smaller) ones are
[8,9]. In the next section, we will present two most usual regular- given infinite values, so the truncated version of this estimator is:
ization techniques to make the linear estimation more stable. b a ¼ VR1 U T y with R1
H a a
 
4. Regularized identification of a transfer function ¼ diag R1 ; R2 ; . . . ; R1
1 1
a ; 0; . . . ; 0 ð11Þ

Let us consider the general form of the direct model that con- where a is the regularization parameter which varies from 1 to N t . If
nects the source to its response by a transfer function, Eq. (3b). a ¼ Nt , this corresponds to the case without regularization (OLS
The Singular Value Decomposition (SVD) of square matrix M ðuÞ estimation, Eq. (9)).
decomposes it into a product of three matrices:
4.2. Zero-order Tikhonov regularization
M ðuÞ ¼ URV T with U T U ¼ UU T ¼ VV T ¼ V T V ¼ I Nt
R ¼ diag ðR1 ; R2 ; . . . ; RNt Þ; where R1 P R2 P . . . P RNt P 0 b l is
The estimation using zero order Tikhonov regularization, H
U ¼ ½U 1 ; U 2 ; . . . ; U Nt  and V ¼ ½V 1 ; V 2 ; . . . ; V Nt  defined by:
  
ð8Þ b l ¼ arg min J ðHÞ
H ð12Þ
l
where R is the diagonal matrix composed of the singular values Ri
of M ðuÞ (for i ¼ 1 to Nt ), U the matrix of its N t left singular (column) This estimation is the solution of the minimization of the fol-
vectors U i and V the matrix of its N t right singular (column) vectors lowing criterion:
V i . By replacing M ðuÞ by its decomposition, the general form of the J l ðHÞ ¼ kM ðuÞH  yk2 þ l2 kHk2 ð13Þ
Ordinary Least Estimator (OLS) of the impedance or transmittance
(model (6) and model (7)) can be rewritten as: where l is the Tikhonov regularization parameter which varies
from 0 to 1. If l ¼ 0, this corresponds to the case without regular-
b OLS ¼ ½M ðuÞ y ¼ VR1 U y
H 1 T
ð9Þ ization (OLS estimation, Eq. (9)). This Tikhonov estimate can be
The measurement response y is always corrupted by noise, such given an inverse SVD-like form:
as y ¼ yexact þ e. If the noise vector e can be considered as indepen- b l ¼ VR1 U y with R1
H l l
dent and identically distributed (i.i.d.) it is characterized by a stan- !
dard deviation r, and we can write cov ðeÞ ¼ r2 I Nt , where I Nt is R1 R2 RN t
¼ diag ; ;...; ð14Þ
identity matrix of size N t and cov ðÞ designates the variance- R21 þ l2 R22 þ l2 R2Nt þ l2
covariance matrix of a stochastic vector. In this case the diagonal
coefficients of the variance-covariance matrix are the variances So, instead of replacing the inverse of the smallest singular values
R2
of the coefficients of the H estimates, that is the squares of the by zeroes, all of them are premultiplied by a filter factor f i ¼ R2 þi l2
i
standard deviations of their values (a measure of the dispersion
that is hence smaller than unity. Matrix R1
l in Eq. (14) can hence
of the estimated values around their expectation). It is defined as:
   1 be given the following forme:
cov Hb OLS ¼ r2 M ðuÞT M ðuÞ ¼ r2 VR2 V T ð10Þ  
R1
l ¼ Fl R
1
with F l ¼ diag f 1 ; f 2 ; . . . ; f Nt
If the matrix to be inverted is ill-conditioned (the ratio of the The choice of the hyperparameter or the regularization coeffi-
largest to the smallest singular value R1 =RNt is large), the smallest cient (a for TSVD regularization and l for Tikhonov regulariza-
singular values which are present in matrix R2 will bring a dom- tion) is delicate. A low regularization (high a or small l,
 
inant contribution to the coefficients of cov H b OLS including its depending on the regularization technique) will lead to a solu-
tion that will remain as faithful as possible to the measurements
diagonal values. Consequently the identification error will become
(the residuals go to zero) with a high estimation error. Con-
very sensitive to the measurement noise [8].
versely, a strong regularization will aggravate the bias. This
So the ordinary way (9) of implementing the identification
means that there is an optimum between the two. If the stan-
starting from the experimental input/output distributions is ill-
dard deviation of the measurement noise r is known, the opti-
posed if the matrix M ðuÞ is ill-conditioned. In the literature there
mal hyperparameter value can be found by Morozov’s
are two techniques to treat this undesirable character: either by fil-
discrepancy principle. This optimal value gives an estimate with
tering the noise r ! 0 or by regularizing the matrix to be inverted
a level of residuals slightly higher than the measurement noise
to make it well-conditioned [9]. It is preferable to use the second
[9]. This can be expressed as:
technique of treatment because, in some applications, the matrix sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
to be inverted M ðuÞ is very ill-conditioned and, consequently, the PNt rec 2
i¼1 ðyi  yi Þ
numerical noise of the machine which is unavoidable, can play RMSR P r with RMSR ¼ ð15Þ
Nt  1
the same role as the measurement noise [10].
In this work therefore, we will regularize the matrix to be where RMSR is the Root Mean Square Residual, r the standard devi-
inverted (also called sensitivity matrix), M ðuÞ to get a stable iden- ation of the noise and yrec
c the recalculated signal using estimation
tification. In other words, one modifies its singular values, to make b c with c  a or l, which is defined as:
H
it well-conditioned. We have tested and compared here two of the
yrec b
most popular methods that are the truncated singular value c ¼ MðuÞ H c ð16Þ
decomposition TSVD [11–14] and Tikhonov regularization [15,16].
936 W. Al Hadad et al. / International Journal of Heat and Mass Transfer 116 (2018) 931–939

−1 5
10 −1 10
10

0
10
−2
10

0 50 100 −5
10
−2
10
−10
10

−15
10
0 1000 2000 3000 4000 5000 6000
−3
10
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Fig. 7. Singular values Ri of MðQ Þ for step and periodic excitation.

Fig. 5. RMSR for TSVD impedance estimation.

−3
x 10
6
0.12
5
0.1
4
0.08
3

0.06 2

0.04 1

0.02 0

−1
0 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
0 50 100 150 200 250 300

b in and Z
Fig. 8. Estimated impedances Z b out for step excitation.
Fig. 6. RMSR for Tikhonov impedance estimation.

The estimated impedances Z b in and Z

b out by TSVD and Tikhonov
5. Results and discussion
regularization at the optimal values of hyper-parameters given in
Table 4 are plotted in Fig. 8 for step excitation and in Fig. 9 for peri-
5.1. Experimental identification of impedances Z in and Z out
odic excitation. These curves show that two inversion techniques
give nearly the same results.
To find the optimal hyper-parameters of each regularization
In Fig. 10, the estimated impedances for the first and second
method (aopt and lopt ) for each experiment, RMSR can be plotted
experiments are compared with Tikhonov regularization for each
for each estimation as a function of the corresponding hyper-
calibration. One notes that the impedances estimated from the
parameter. In the same type of figure, the level of the standard
two different experiments are the same with a small offset (nor-
deviation of the noise (r^ ¼ 0:008 K, estimated from prior steady
mally the two functions should overlap, because of the intrinsic
state measurement) is shown. The optimum value is selected fol- character of this transfer function). This can be explained by the
lowing Morozov’s discrepancy principle (intersection of the level of water in the plenum at the inlet. In the two experiments,
standard deviation line with RMSR curve). We present in Figs. 5 the channel was in vertical position but the level of water was
and 6 RMSR as a function of a and l respectively only for the first higher in the first than in the second experiment. A higher level
experiment. The corresponding optimal values for each experiment means more inertia, so the characteristic time is naturally larger.
are summarized in Table 4. This table shows that to estimate Z at
the inlet, we regularize less than at the outlet (aopt
in > aout and
opt
5.2. Experimental identification of transmittance W out
lopt
in < lout ). This can be explained by a higher signal-to-noise ratio
opt in

at the inlet. To give an idea of the singular values Ri of the matrix to The estimated transmittance starting form the temperatures
be inverted MðQ Þ, we plot in Fig. 7 Ri as a function of their index
profiles (hin out
b and hb ) using TSVD and Tikhonov regularization are
i for the first and the second experiment. The largest and smallest
plotted in Fig. 11 for the first experiment and in Fig. 12 for the sec-
singular values R1 and RNt respectively are given in Table 4.
ond experiment. This figure shows that the two regularization

Table 4
Step, periodic experiments and its corresponding optimal hyper-parameters.

N t ¼ tf (s) b (W/K)
Z R1 RNt

aopt
in
aopt
out lopt
in
lopt
out

Step 5000 77 44 49:25 84:25 1:15  104 1:8

Periodic 5958 150 74 72:5 180:5 1:23  104 3:10  1013
 W. Al Hadad et al. / International Journal of Heat and Mass Transfer 116 (2018) 931–939 937

techniques give practically the same estimate. The general shape of The upper bar denotes the Laplace transform here and p (in
such a transfer function (impedance and transmittance) is always Hertz) is the Laplace parameter. For a step excitation, we get:
characterized by a zero asymptotic limit for t ! þ1 with no neg-
uss
ative values in between (it is a response to a Dirac pulse). So the ðpÞ ¼ HðpÞ
y ð18bÞ
p
oscillations that appear in each estimation, are not physical. These
oscillations can be explained by the fact that the input and the out-
put of the identification model are both noisy while, on the con-
−3
trary, in the impedance identifications, only the response was 6 x 10
noisy.
As we have shown in our previous work [6], the transmittance 5
can be estimated also from the impedances corresponding to the 4
same temperature profiles:
3
^ out ¼ ½MðZ in Þ1 Z out
W ð17Þ
in 2
To check this equation and to show that the transmittances are 1
as intrinsic as the impedances (independent of the excitation
0
form), the transmittances estimated from temperatures (hinb and
b in
hout , see Eq. (7)) and from previously estimated impedances ( Z −1
b 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
and Z b out , see Eq. (17)), for the first and the second experience are
plotted in Fig. 13 and by zooming (short times, t 6 1000 s) in
Fig. 14 for Tikhonov regularization only. This figure shows that b in and Z
Fig. 10. Estimated impedances Z b out for step and periodic excitation
from two different excitations (heat sources) and their responses, (Tikhonov regularization).
we obtain globally the same transmittances, this means that the
transmittances are intrinsic quantities. One also notices that the
short times zero limit is more difficult to retrieve for transmittance
identification starting from previously identified impedances than −4
x 10
from measured temperatures. It may stem from the bias in both 15
previously estimated impedances. So we can say that, for transmit-
tance estimation, better results are obtained starting from the
measured temperatures. In the same way, both estimation tech- 10
niques yield lower physical transmittance estimates (absence of
zero limit or large oscillations) for the periodic source than from
the step one: the cause may be the sampling period 1 s that is 5
not small enough with respect to the 129 s period of the excitation.
The corresponding optimal regularization parameters are given in
Table 5. 0

Steady state transfer function Hss : There is also a permanent

transfer function which can be calculated as the integral of its tem-
−5
poral distribution. This can be easily shown when the excitation 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
uðtÞ and therefore its response yðt Þ tend toward an asymptote
when t tends to infinity, with their limits respectively denoted
c out for step excitation starting from both
uss and yss . Let us consider Eq. (1) in Laplace domain which can Fig. 11. Estimated transmittance W in
temperature responses.
be written as:

y  ðpÞ
ðpÞ ¼ HðpÞu ð18aÞ

−4
−3 x 10
x 10 20
6

5 15

4
10
3

2 5

1
0
0
−5
−1 0 1000 2000 3000 4000 5000 6000
0 1000 2000 3000 4000 5000 6000

c out for periodic excitation starting from both

Fig. 12. Estimated transmittance W in
b in and Z
Fig. 9. Estimated impedances Z b out for periodic excitation. temperature responses.
938 W. Al Hadad et al. / International Journal of Heat and Mass Transfer 116 (2018) 931–939

 
−4 ðpÞ ¼ uss lim HðpÞ
yss ¼ limyðtÞ ¼ lim½py
20 x 10 t!1 p!0 p!0

¼ u Hð0Þ
ss
ð18; d; e; f; gÞ
15
Comparison of (18g) and (18c) leads to the expression of the
10 corresponding steady state transfer function Hss :
Z 1 Z 1
5
Hss ¼ Hð0Þ ¼ e0t HðtÞdt ¼ HðtÞdt
0 0
0
Note here that the unity of the steady state transfer function is
that of its temporal version, multiplied by time. If the output y is a
−5
0 1000 2000 3000 4000 5000 6000 temperature variation ðKÞ caused by an input that is a heat power
in Watts, Hss ¼ Z ss has the same units ðK=WÞ as a thermal resis-
c out from the temperatures (hin and hout ) or from
tance but it does not involve the notion of area linked to a given
Fig. 13. Estimated transmittance W in b b
the estimated impedances ( Z b in and Z b out ) for step and periodic excitation using surface.
Tikhonov regularization. By integrating the impedances and transmittance profiles (see
Figs. 8–12) for step and periodic excitation (TSVD and Tikhonov reg-
ss
ularization), one obtains Z ss ss out
in ; Z out and W in . The corresponding val-
ues are given in Table 6. This result can also be found by taking the
−4 ratio of asymptotic power and temperature values Table 7.
20 x 10

15
6. Conclusion

10 We have shown that transfer functions linking a forced temper-

ature response, usually a local temperature, and a unique thermal
5 source, that is either a time varying heat power or a variation of
temperature, can be identified experimentally in a heterogeneous
0 system where conduction and advection occur. It should be noted
that the accuracy of the identified transfer function depends on the
−5 quality of the measurements and on the corresponding inversion
0 100 200 300 400 500 600 700 800 900 1000 technique. This function, once identified, takes into account all
the effects that are difficult to model (ex: wall roughness, turbu-
c out for t 6 1000 s from the temperatures (hin lence, contact resistance, defect in the materials that can have ther-
Fig. 14. Estimated transmittance W in b
and hout ) or from the estimated impedances b in and Z
(Z b out ) for step and periodic mophysical properties different from their a priori known nominal
b
excitation using Tikhonov regularization. values,. . .). If the physical system is linear (no thermodependance
of its thermophysical coefficients) with time invariance of its coef-
ficients (stationary fluid flow, no ‘‘ageing” of the system) its iden-
For long times, steady state values are reached which means tified transfer function allows the use of a convolutive model
that the steady state output yss is proportional to the steady state that has the right structure.
input uss . We call naturally Hss the proportionality constant defined Application to a half heat exchanger, where only a change of
as: either an upstream volumetric thermal power dissipation or of
the inlet temperature of one fluid occurs at initial time, has been
yss ¼ Hss uss ð18cÞ
made. Two regularization techniques (TSVD and Tikhonov) were
By applying the theorem of the final value, the asymptotic value compared here to make the identification well posed. These two
of the output, or the steady state response, yss can be calculated as: regularization techniques give practically the same results. We

Table 5
c out from temperature profiles ðhin ; hout Þ or from impedances profiles
Step, periodic experiments and its corresponding optimal hyper-parameters for transmittance estimation W in b b
b in ; Z
ðZ b out Þ.

Q (W) 4t (s) N t ¼ tf (s) c ¼ f ðhin ; hout Þ

W c ¼ f ðZ
W b in ; Z
b out Þ
b b

a
opt
l opt
lopt
Step 3:6161 1 5000 19 150 0:3
Periodic 6:4286 1 5958 15 150 0:4

Table 6
Identified steady state impedances and transmittance from the temporal distribution of the input and output (step and periodic excitation).
R1 R1
b ss ¼
Z b ðt Þdt
Z c outss ¼
W c out ðtÞdt
W
0 in 0 in

Inlet Outlet TSVD Tikh.

TSVD Tikh. TSVD Tikh.
Step 5:97 5:97 2:68 2:68 0:46 0:46
Periodic 5:87 5:87 2:75 2:75 0:47 0:48
 W. Al Hadad et al. / International Journal of Heat and Mass Transfer 116 (2018) 931–939 939

Table 7
Identified steady state impedances and transmittance from asymptotic values of the input and output (step excitation case).
1 1
Q ss (W) Z ss Z ss
ss ss ss
hin
b (K) hout
b (K) in (K W ) out (K W ) W out
in (s1)

Step 3:6161 21:59 9:70 5:97 2:68 0:45

have shown from two different experiments (different excitations) [6] W. Al Hadad, D. Maillet, Y. Jannot, Modeling unsteady diffusive and advective
heat transfer for linear dynamical systems: a transfer function approach, Int. J.
that we obtain the same transfer function, which confirms our pre-
Heat Mass Transfer 115 (2017) 304–313.
vious analytical work [6]. [7] D. Maillet, Y. Jannot, A. Degiovanni, Analysis of the estimation error in a
Once adhoc transfer functions are identified for a given system, parsimonious temperature-temperature characterization technique, Int. J.
they can be used experimentally in different types of applications: Heat Mass Transfer 62 (2013) 230–241.
[8] J.V. Beck, B. Blackwell, C.R.S. Clair Jr, Inverse Heat Conduction: Ill-Posed
inverse input estimation (case of impedances) or in direct simula- Problems, James Beck, 1985.
tion of the outlet bulk temperature variation (case of a transmit- [9] R.C. Aster, B. Borchers, C.H. Thurber, Parameter Estimation and Inverse
tance with known inlet temperature) or as a virtual temperature Problems, Elsevier Academic Press, 2012.
[10] W. Al Hadad, Heat Transfer in Mini-Channels: Unsteady Behaviour and
sensor of the inlet temperature variation (case of a transmittance Convolutive Approach, (in French: Thermique des mini-canaux:
with known outlet temperature). Let us note, to conclude, that comportement instationnaire et approche convolutive), PhD thesis,
transfer functions HðtÞ are a powerful tool that can be used in Université de Lorraine, September 22, 2016.
[11] P.C. Hansen, Truncated singular value decomposition solutions to discrete ill-
model reduction (from analytical or numerical detailed simula- posed problems with ill-determined numerical rank, SIAM J. Sci. Stat. Comput.
tions) or in model identification (from experimental 11 (3) (1990) 503–518.
measurements). [12] W. Al Hadad, D. Maillet, S. André, B. Rémy, Regularization using truncated
singular value decomposition for estimating the fourier spectrum of a noised
space distribution over an extended support, Comput. Assis. Meth. Eng. Sci. 21
References (3–4) (2014) 211–222.
[13] Y. Rouizi, D. Maillet, Y. Jannot, Fluid temperature distribution inside a flat
[1] S. Lalot, P. Florent, S. Lang, A. Bergles, Flow maldistribution in heat exchangers, mini-channel: semi-analytical wall transfer functions and estimation from
Appl. Therm. Eng. 19 (8) (1999) 847–863. temperatures of external faces, Int. J. Heat Mass Transfer 64 (2013) 331–342.
[2] L. Ljung (Ed.), System Identification Theory for the User, second ed., Prentice [14] Y. Rouizi, W. Al Hadad, D. Maillet, Y. Jannot, Experimental assessment of the
Hall PTR, Upper Saddle River, NJ, USA, 1999. fluid bulk temperature profile in a mini channel through inversion of external
[3] E. Videcoq, D. Petit, Model reduction for the resolution of multidimensional surface temperature measurements, Int. J. Heat Mass Transfer 83 (2015) 522–
inverse heat conduction problems, Int. J. Heat Mass Transfer 44 (10) (2001) 535.
1899–1911. [15] A.N. Tikhonov, V.Y. Arsenin, Solutions of Ill-Posed Problems, VH Winston &
[4] M. Girault, D. Petit, Resolution of linear inverse forced convection problems Sons, 1977.
using model reduction by the modal identification method: application to [16] A.N. Tikhonov, A. Goncharsky, V. Stepanov, A.G. Yagola, Numerical Methods for
turbulent flow in parallel-plate duct, Int. J. Heat Mass Transfer 47 (17) (2004) the Solution of Ill-Posed Problems, vol. 328, Springer Science & Business
3909–3925. Media, 2013.
[5] E. Videcoq, D. Petit, A. Piteau, Experimental modelling and estimation of time
varying thermal sources, Int. J. Therm. Sci. 42 (3) (2003) 255–265.