See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/251729481

Estimation of front surface temperature and

heat flux of a locally heated plate from
distributed sensor data on the back surface

Article in International Journal of Heat and Mass Transfer July 2011

DOI: 10.1016/j.ijheatmasstransfer.2011.03.043

CITATIONS READS

15 50

4 authors, including:

Zaichun Feng Yuwen Zhang

University of Missouri University of Missouri
116 PUBLICATIONS 1,497 CITATIONS 354 PUBLICATIONS 4,362 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Yuwen Zhang on 25 January 2016.

The user has requested enhancement of the downloaded file.

 International Journal of Heat and Mass Transfer 54 (2011) 34313439

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

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

Estimation of front surface temperature and heat ux of a locally heated plate

from distributed sensor data on the back surface
Z.C. Feng a,, J.K. Chen a, Yuwen Zhang a, James L. Griggs Jr. b
a
Department of Mechanical and Aerospace Engineering, University of Missouri, Columbia, MO 65211, USA
b
Science Applications International Corporation, 6200 Uptown Blvd. NE, Albuquerque, NM 87110, USA

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

Article history: We present a new method of solving the three-dimensional inverse heat conduction (3D IHC) problem
Received 29 September 2010 with the special geometry of a thin sheet. The 3D heat equation is rst simplied to a 1D equation
Received in revised form 22 February 2011 through modal expansions. Through a Laplace transform, algebraic relationships are obtained that
Accepted 22 February 2011
express the front surface temperature and heat ux in terms of those same thermal quantities on the back
Available online 13 April 2011
surface. We expand the transfer functions as innite products of simple polynomials using the Hadamard
Factorization Theorem. The straightforward inverse Laplace transforms of these simple polynomials lead
Keywords:
to relationships for each mode in the time domain. The time domain operations are implemented through
Inverse problem
Transfer function
iterative procedures to calculate the front surface quantities from the data on the back surface. The iter-
Sensor compensation ative procedures require numerical differentiation of noisy sensor data, which is accomplished by the
Temperature measurement SavitzkyGolay method. To handle the case when part of the back surface is not accessible to sensors,
we used the least squares t to obtain the modal temperature from the sensor data. The results from
the proposed method are compared with an analytical solution and with the numerical solution of a
3D heat conduction problem with a constant net heat ux distribution on the front surface.
2011 Elsevier Ltd. All rights reserved.

1. Introduction
To conduct thermal measurements under harsh environment, it
has been proposed that sensors be located away from direct con-
tact with the environment and mathematical models be used to
calculate the desired quantities from the sensor measurement
data. Specically, the front surface temperature can be determined
indirectly by solving an inverse heat conduction problem (IHCP)
[118] based on the transient temperature and/or heat ux mea-
sured on the back surface.
In our previous work [19], we applied the Laplace transform to
the one-dimensional heat conduction problem. Algebraic equa-
tions relating temperature and heat ux on the front and back sur-
faces were obtained. Although these equations were no different
from those given in [2023], we proposed approximations to the
transfer functions through which the front surface temperature
and heat ux can be obtained from the data on the back surface
through mathematical operations in the time domain. No compli-
cated inverse Laplace transform is required. Following that work,
we have replaced the ad hoc approximations to the transfer func-
tions by a procedure which achieves better approximations
Corresponding author. Tel.: +1 573 884 4624; fax: +1 573 884 5090.
E-mail address: fengf@missouri.edu (Z.C. Feng).
through iterations [24]. The method was shown to be both accu-
rate and easy to implement for one-dimensional IHCPs. However,
it is formulated to handle one-dimensional IHCPs only.
The above method for one-dimensional IHCPs is too restrictive
in measurement practices. We thus wish to generalize the method
to make it applicable to realistic measurement problems. Since our
motivation lies in measurement instrumentation, we have the ex-
ibility of selecting the geometry and even boundary conditions that
help to simplify the solution of the IHCP. We may restrict the
geometry to that of a rectangular sheet. Furthermore, we assume
that the sheet thickness is much smaller than the transverse
dimensions. Therefore, heat conduction is mostly across the thick-
ness. This special geometry and boundary condition allows us to
introduce simplication from 3-dimension to 1-dimension. Inte-
gral transformation methods have been proposed in [22]. In this
paper, we use Fourier series to convert the 3-dimensional heat con-
duction equation into a system of 1-dimensional equations. Each
equation is solved by the method we proposed in [24].
In the following section, we present the reduction to the one-
dimensional problem. Since the resulting one-dimensional problem
contains an additional source term, we present in Section 3 the
polynomial representation of the transfer function based on the
Hadamard Factorization Theorem; the polynomial representation
is a generalization of our prior result. From the polynomial
representation of the transfer function, an iterative procedure is

0017-9310/$ - see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijheatmasstransfer.2011.03.043
3432 Z.C. Feng et al. / International Journal of Heat and Mass Transfer 54 (2011) 34313439

Nomenclature

cps mass specic heat of the slab, J/(kg K) X dimensionless spatial coordinate variable along the
E total error length direction
f normalized heat ux, K Y dimensionless spatial coordinate variable along the
g notation for the integrand width direction
G transfer function Z dimensionless spatial coordinate variable along the
ks thermal conductivity, W/(m K) thickness direction
lx slab length, m
ly slab width, m Greek symbols
lz slab thickness, m a thermal diffusivity, m2/s
Lx ratio of slab length to thickness, lx/lz d Kronecker delta
Ly ratio of slab width to thickness, ly/lz qs density of the slab, kg/m3
M maximum mode number along the length direction s dimensionless time
N maximum mode number along the width direction h temperature, K
p pole of a transfer function H Laplace transform of the temperature
q heat ux, W/m2 U Laplace transform of the normalized heat ux
s Laplace transform variable x dimensionless frequency
t time, s n dimensionless length variable
tc characteristic time, s
Ts temperature of the slab, K Subscripts
u unit step function b back surface quantity
U Laplace transform of temperature f front surface quantity
x spatial coordinate variable along the length, m m mode number along the length direction
y spatial coordinate variable along the width, m n mode number along the width direction
z spatial coordinate variable along the thickness, m

obtained. From the data on the back surface, we need to extract the @T s
ks qx; y; t for z 0 5
data for each one-dimensional problem. In Section 4, we present @z
the equations we need when the sensor distribution on the back To simplify, introduce the following changes to dimensionless
surface is arbitrary. In Section 5, we validate our method for a sin- variables:
gle 3D mode by comparing our results with those from analytical
solutions. In Section 6, we provide a comparison of our IHC results t t c s; x lz X; y lz Y; z lz Z 6
with a direct three-dimensional heat conduction calculation.
and to variables with dimensions of temperature (K)
2. Mathematical model and reduction to one-dimensional ks
qx; y; t f X; Y; s 7
problem lz
T s x; y; z; t hX; Y; Z; s 8
Although the formulation for a three-dimensional anisotropic
heat conduction problem is given in [21], we give the following for- where
mulation for isotropic problem since it is the application we have
in mind. Since our interest lies in the heat input and temperature qs cps l2z
tc 9
on the front surface, we are at liberty in selecting the geometry ks
and boundary conditions that are most convenient in instrumenta- and f(X, Y, s) is the normalized front surface heat ux (with unit K).
tion and subsequent data processing. Thus we consider the following With these substitutions, Eq. (1) becomes
heat conduction equation:
! @h @ 2 h @ 2 h @ 2 h
@T @2T s @2T s @2T s 10
qs cps s ks 2 2 1 @ s @X 2 @Y 2 @Z 2
@t @x2 @y @z
with boundary conditions:
where qs, ks and cps are the density, thermal conductivity and spe-
cic heat of the solid, all assumed to be independent of tempera- @h
0 for X 0 and X Lx 11
ture. We assume adiabatic boundary conditions on the back @X
surface and the four edge surfaces:
@h
@T s 0 for Y 0 and Y Ly 12
0 for x 0 and x lx 2 @Y
@x
@T s @h
0 for y 0 and y ly 3 0 for Z 1 13
@y @Z
and where Lx = lx/lz and Ly = ly/lz. On the front surface, the heat ux is
specied:
@T s
0 for z lz 4
@z @h
 f X; Y; s for Z 0 14
On the front surface, the heat ux is specied: @Z
 Z.C. Feng et al. / International Journal of Heat and Mass Transfer 54 (2011) 34313439 3433

The rectangular geometry and the boundary conditions in the temperature on the corresponding surfaces. Since we assume the back
above are most amenable to solutions using integral transform surface heat ux to be zero, we have the simplied relationships:
method [22]. The problem can be converted into one-dimensional q
heat transfer problem as follows. We express the temperature and Hmn s; 0 cosh s c2mn Hmn s; 1 22
heat ux as superposition of two-dimensional harmonic functions
and
X mpX npY q q
hX; Y; Z; s hmn s; Z cos cos 15
Lx Ly Umn s; 0 s c2mn sinh s c2mn Hmn s; 1 23
m;n0;1;...

and the normalized front surface heat ux In the Laplace domain, Eqs. (22) and (23) give the front surface
X temperature and heat ux based on the back surface temperature.
mpX npY
f X; Y; s fmn s cos cos 16 These algebraic relationships are called transfer functions and they
m;n0;1;...
Lx Ly are exact. The physically meaningful quantities in the time domain
are obtained if inverse Laplace transform is found. When restricting
Note that Eqs. (15) and (16) are Fourier series expansions of the
s to be purely imaginary, i.e. s = jx, the relationships in the fre-
temperature and heat ux in X and Y directions. These substitutions
quency domain are obtained. In [19], such relationships are
yield a one-dimensional heat conduction equation,
examined to obtain the attenuation of the sinusoidal front surface
@ 2 hmn @hmn heat input as it is conducted to the back surface. The attenuation is
 c2mn hmn  0 for m; n 0; 1; 2; . . . 17 found to be especially drastic for high frequency input when time
@Z 2 @s
is non-dimensionalized by the characteristic time dened in (9).
where Because of the unavoidable noises on the back surface measure-
 2  2 ment, a limit exists on the input frequency above which it is not
mp np
c2mn 18 feasible to calculate the front surface temperature and heat ux
Lx Ly
based on the noisy back surface measurement.
The boundary conditions on the side walls are all met by the cosine Instead of resorting to inverse Laplace transform, we expand the
form of the expansion. The prescribed heat ux boundary condition transfer functions as innite product of simple polynomials based
on the front surface Z = 0 now becomes: on the Hadamard Factorization Theorem [26]. We then took advan-
@hmn tage of the straightforward inverse Laplace transform of the simple
 fmn s 19 polynomials. Since the transfer functions in (22) and (23) contain
@Z
an additional term c2mn , the generalized innite product expansions
and the adiabatic boundary condition on the back surface Z = 1 are given in the following. In [24], we have the following
becomes representation:
@hmn ( )
 0 20 p Y1
s
@Z cosh s 1 24
k1 2k  1p=22
The function hmn(s, Z) represents the modal temperature. The gov-
erning equation (17) for the modal temperature now resembles that Replacing s by s c2mn in the above equation, we have
of a one-dimensional heat conduction problem. However, the sec- ( )
ond term in (17) represents a source term that typically occurs in q Y
1
2
s c2mn
a special one-dimensional problem representing ns with convec- cosh s cmn 1 25
k1 2k  1p=22
tion [25].
The term in the braces can be rewritten as
3. Solutions of the one-dimensional equations using Laplace ( )
transform s c2mn c2mn
1 1
2k  1p=22 2k  1p=22
The analytical solution for Eq. (17) can be obtained for a few ( )
s
special cases. (A solution that is used later in this paper for compar-  1 26
ison is given in the Appendix A.) In general, numerical methods are c2mn 2k  1p=22
required. Eq. (17) is conveniently solved in the Laplace domain.
We thus have
Take the Laplace transform of (17). The resulting second order or-
( )
dinary differential equation is then solved with two undetermined q Y
1
c2mn
coefcients. These two coefcients are determined using any two cosh s c2mn 1
k1 2k  1p=22
of the four possible quantities on the two surfaces. Once these ( )
two coefcients are found, the remaining two quantities on the Y
1
s
 1 27
surfaces are thus known. The results can be concisely expressed c2mn 2k  1p=22
k1
as the linear relationship between the Laplace transforms of the
temperature and heat ux on the front surface and on the back sur- Using (24), we note that the rst innite product in (27) equals
face as in the following [22]: cosh cmn. Therefore, we have
  " p p # ( )
Hmn s; 0 cosh s c2mn p
1
sinh s c2mn q
p
scmn Y
1
s
p p cosh s c2mn cosh cmn 1 28
Umn s; 0 s c2mn sinh s c2mn cosh s c2mn
  k1 c2mn 2k  1p=22
Hmn s; 1
 21
Umn s; 1 In [24], we also obtained the following equation:
p " #
where Umn(s, 0) and Umn(s, 1) are the Laplace transform of the mod- sinh s Y 1
s
al heat ux of the front surface and back surface respectively; p 1 29
s k1 kp2
Hmn(s, 0) and Hmn(s, 1) are the Laplace transform of the modal
3434 Z.C. Feng et al. / International Journal of Heat and Mass Transfer 54 (2011) 34313439

Replacing s by s c2mn in the above and following the steps analo- cmn = 0; for that case we have shown that doubling the number of
gous to those leading to Eqs. (25)(28), we have iterations from 8 to 16 has negligible effect.
p " #
sinh s c2mn sinh cmn Y 1
s
p 1 30 4. Extraction of modal temperature from back surface sensor
s c2mn cmn k1 c2mn kp2 data
Substituting Eqs. (28) and (30) into Eqs. (22) and (23), we have
The iteration formulas (35)(38) require the modal temperature
the following equations in the Laplace domain:
on the back surface. Given the back surface temperature measure-
" # ment, the modal back surface temperature can be calculated as
Y
1
s
Hmn s; 0 cosh cmn 1 Hmn s; 1 31 follows:
k1 c2mn 2k  1p=22 Z Z
22d0m d0n Lx Ly
mpX
hmn s; 1 hX; Y; 1; s cos
" # Lx Ly 0 0 Lx
sinh cmn Y1
s
Umn s; 0 s c2mn 1 Hmn s; 1 32 npY
cmn k1 cmn kp2
2  cos
Ly
dX dY 39

where d0m and d0n are Kronecker delta, m = 0, 1, 2, . . . , M and

These equations correspond to the following operations in the time
n = 0, 1, 2, . . . , N. When back surface temperature is known through
domain:
sensors placed at the center of uniformly distributed rectangular
hmn s; 0 cosh cmn grids, we can evaluate the double integral approximately. The
( ) straight forward approach is to divide the integration area also into
Y1
1 d
 1 hmn s; 1 33 rectangular grids. On each grid (X e (X0, X1), Y e (Y0, Y1)), approxi-
k1 c2mn 2k  1p=22 ds mate the integrand by the function

 Y ( ) gX; Y g 00 1  n1  g g 01 1  ng g 10 n1  g g 11 ng
1
sinh cmn d 1 d
fmn s c2mn 1 hmn s;1 34 40
cmn ds k1 c2mn kp2 ds
where
Recall that hmn(s, 0) and hmn(s, 1) are the modal temperature on the X  X0 Y  Y0
front surface and back surface respectively and fmn(s) is the modal n ; g 41
X1  X0 Y1  Y0
heat ux on the front surface. They represent exact relationships
in the time domain similar to the result in [27]. These exact rela- and g00, g01, g10, g11 are the values of the integrand at the corners of
tionships involve time derivatives to innite order. Practical utility the grid. Then
of these exact relationships is based on their approximations as Z Z
X1 Y1
shown in [28]. 1
gX; YdX dY g g 01 g 10 g 11 X 1  X 0
Eqs. (33) and (34) contain derivatives of innite order. As k be- X0 Y0 4 00
comes large, the coefcient in front of the derivative term becomes  Y 1  Y 0 42
small. Therefore approximation can be obtained by truncating the
innite product. This can be implemented in an iterative scheme. The average of the values of the integrand at the four corners can be
Starting with the back surface temperature for a given mode, i.e. approximated by the value of the integrand at the center point of
hmn(s, 1), the initial approximations to the front surface modal the grid, i.e. the location of the sensor. Therefore,
temperature and heat ux are
" # 22d0m d0n X X mpX i
1 dhmn s; 1 hmn s; 1 hX i ; Y j ; 1; s cos
1
hmn s; 0 cosh cmn hmn s; 1 35 Lx Ly i j
Lx
c2mn p=22 ds
npY j
 cos dX i dY j 43
  Ly
1 sinh cmn 2 dhmn s; 1
fmn s cmn hmn s; 1 36
cmn ds where (Xi, Yj) is the center of the (i, j)th grid and dXi and dYj are the
sides of the grid.
From these initial approximations, we obtain the successive If the back surface is divided into I  J rectangular grids, we
approximations using iterations: have
K
    
K1 K 1 dhmn s; 0 1 Lx 1 Ly
hmn s; 0 hmn s; 0 37 X i ; Y j i ; j 44
cmn 2K 1p=2
2 2 ds 2 I 2 J
Since we have a total of I  J sensor measurements, we cannot
K
K1 K 1df mn s include more total modes than the number of sensors. Therefore,
fmm s fmn s 38
c2mn K p2 ds
M 1N 1 6 I  J 45
for K = 1, 2, 3, . . .
to avoid aliasing.
Therefore, based on the back surface modal temperature, the
front surface temperature and heat ux can be obtained using Eq. (43) can be used to obtain the back surface modal tempera-
the iterative formulas given in (35)(38). Iteration stops when fur- ture when sensors are uniformly distributed on the back surface.
ther iteration results in changes smaller than a set value which is Under some circumstances, data from certain parts of the back sur-
called truncation tolerance. An inspection of (37) and (38) leads face are not available because of limitations of the xture. We refer
us to conclude that the iteration converges faster for cmn > 0 than to this situation as arbitrarily distributed sensors. As a result, we
for cmn = 0. In our previous work [24] for 1D IHC problem, we have have data at a number of known points. Let Ns denote the total
 Z.C. Feng et al. / International Journal of Heat and Mass Transfer 54 (2011) 34313439 3435

number of sensors and si be the reading of the ith sensor. With the X
Ns

temperature sensor data, we can determine the modal tempera- dj si Sj X i ; Y i : 53

i1
ture by minimizing the total errors dened below:
" #2 The mode decomposition calculation has been veried by
X
Ns X
M X
N
mpX i npY i starting from the sensor data of a known back surface temperature
E si  hmn s; 1 cos cos 46
i1 m0 n0
Lx Ly distribution consisting of a few modes. The calculation is found to
recover the assumed modal temperature very accurately. We must
where (Xi, Yi) is the coordinate of the ith sensor. For notational clar-
ity, we organize the modal temperatures into a vector [hl] with
l = m + 1 + n(M + 1). Then [hl] contains Nm = (M + 1)(N + 1) total
modal temperature. Therefore, Front Surface Heat Flux
1
" #2
X
Ns X
Nm
E si  hl Sl X i ; Y i 47
i1 l1
0.8

where
0.6
mpX i npY i
Sl X i ; Y i cos cos 48
Lx Ly Front Surface Temperature
0.4

and the mode numbers are calculated by the following:

  0.2
l1
m modl  1; M 1; n 49 Back Surface Temperature
M1
0
In order to minimize E, we differentiate E with respect to hl and
set it zero. We obtain the following algebraic equations: 0 0.5 1 1.5

X
Nm X
Ns X
Ns
hl Sl X i ; Y i Sj X i ; Y i si Sj X i ; Y i 50
l1 i1 i1 Front Surface Heat Flux
1
for j = 1, 2, . . . , Nm, which can be written in matrix form:

ch d 51 0.8

where
0.6
X
Ns
clj Sl X i ; Y i Sj X i ; Y i ; 52 Front Surface Temperature
i1 0.4

and
0.2
Back Surface Temperature
0
Front Surface Heat Flux
0 0.5 1 1.5
1

0.8
Front Surface Heat Flux
1
0.6
0.8
Front Surface Temperature
0.4
0.6

0.2 Front Surface Temperature

0.4
Back Surface Temperature

0
0.2
Back Surface Temperature
0 0.5 1 1.5
0

Fig. 1. Convergence of the iteration to obtain modal heat ux and temperature for 0 0.5 1 1.5
m = n = 12. The front surface truncation tolerance is set at 0.001. All lines represent
the theoretical values. Sampled back surface modal temperatures represented by +
are the only sensor data used by the program. Symbols and o represent the Fig. 2. The effect of the order of polynomials on convergence, with noised data
recovered front surface modal heat ux and modal temperature respectively, for (a1 = 0.01, a2 = 0.005). The seven point SavitzkyGolay data smoothing and
m n 12, Lx Ly 20. Data smoothing and differentiation use the Savitzky differentiation was used. The orders of polynomials, from top to bottom, are 4, 3,
Golay method, with seven points and tted by fourth order polynomials. and 2, respectively.
3436 Z.C. Feng et al. / International Journal of Heat and Mass Transfer 54 (2011) 34313439

keep in mind that the total number of modes must not exceed the
total number of the sensors. Otherwise, singularity in matrix c oc-
curs. Singularity in matrix c can also result from a bad distribution
of sensors such as clustering sensors over a small area on the back
surface.
5. Code validation against one-mode analytical solution
The proposed method of solving IHCPs is to rst decompose the
back surface temperature data into modes, using Eq. (43). Next
using the time domain relationships (33) and (34), the modal front
surface temperature and heat ux for each mode is calculated.
Using superposition, the temperature and heat ux at a given point
on the front surface are then obtained by summing up the contri-
bution from each mode, using Eqs. (15) and (16).
Since the available data are from the nite number of sensors on
the back surface, the number of modes cannot exceed the number
of sensors. Truncation error cannot be controlled by increasing the
number of modes at will. Problems whose spatial variations are
characterized by short characteristic length within the plane may
require unrealistically large number of sensors. Therefore, our
method is more suited for problems whose spatial variations are
not too drastic.
Aside from the truncation issue, our method treats each mode
one at a time. In the following, we provide validation of our pro-
gram for the front surface heat ux
2pmX 2pnY
f X; Y; s cos cos us
Lx Ly
where u(s) is a unit step function. The analytical solution for this
problem is the following:
2pmX 2pnY
hX; Y; Z; s hmn s; Z cos cos
Lx Ly
where hmn is given in Appendix A based on the solutions available in
the literature [29,30]. The plate geometry we considered has
Lx = Ly = 20. For a chosen mode, m = n = 12, we rst calculate the
modal back surface temperature hmn for 0 6 s 6 1.5 sampled with
an interval of 0.1. The sampled modal temperature is then used to
obtain the modal front surface temperature and heat ux using
the iterations given in (37) and (38). Iteration is stopped when
the maximum change from the latest iteration results in change
which is smaller than a preset value. Fig. 1 shows the comparison
of the computed results with the theoretical values.
The iteration formulas (35)(38) require the function and its
rst order derivative for each iteration. Since data in actual mea-
surement are discrete sampled data, derivative calculation can
introduce noises which get further amplied at succeeding itera-
tions. There is a vast literature in digital signal processing in deal-
ing with discrete data. We have selected the SavitzkyGolay (SG)
method [3134] for its proven satisfactory performance in han-
dling similar problems. We specically select seven points for
smoothing and derivative calculations. The SG method uses
least-squares t of a point and three points on either side of it with
a chosen order of polynomial. For the rst and last three points,
this scheme would have required data that are not in the data se-
quence. The method compromises by using the closest data avail-
able. Therefore, the smoothed values for the rst four points are
calculated using the same seven data points in the beginning of
sequence, although the weights used to calculate the slopes at each
point are different. The last four points are handled similarly.
The SavitzkyGolay method amounts to a low-pass lter. The
bandwidth of the lter is obviously affected by the order of the
polynomials used to t the seven points. To investigate how well
the SG algorithm works in the presence of unavoidable sensor
noise, we include additive and multiplicative noise terms to cor-
54
rupt the data,
hi 1 a1 n1i hmn si ; 1 a2 n2i 56
where hi is the noised data, hmn(si, 1) is sampled back surface modal
55 temperature, a1 and a2 are relative and absolute scaling factors
respectively, and n1i and n2i are pseudorandom numbers following
standard normal distribution. We show the results with noise in
Fig. 2, where a1 = 0.01 and a2 = 0.005. The order of polynomials in
the t is 4, 3 and 2 from top to bottom panels respectively. From
the calculated heat ux, we conclude that ltering by low order

Front
Surface
Net Heat
Flux (W/cm2)
at t = 1.00 s

500
450
400
60 350
54 300
250
48
450-500 200
42 150
400-450
36 100
350-400
300-350 30 50
24 0
250-300
60

18
200-250
54
48

12
42

150-200
36
30

6
100-150
24
18

0
50-100
12
6
0

0-50

Fig. 3. Distribution of the net front surface heat ux. The distribution is an elliptical Gaussian with major and minor axes of 18 mm and 9 mm (1/e radii), respectively. The
center of the ellipse is at (36, 33) mm which is displaced from the sheet geometric center. The axes of the ellipse are rotated 24 relative to the grid. The dimensions along the
x and y axes are millimeters. The peak heat ux is 500 W/cm2.
 Z.C. Feng et al. / International Journal of Heat and Mass Transfer 54 (2011) 34313439 3437

polynomials is more effective in ltering out the high frequency
noise and less responsive to the dynamic input. If the process is very
dynamic, we should minimize the error by ltering the signal before
sampling and using higher order polynomials (4, 5, or 6). Noise in
the sampled data is not easy to handle since there is no easy method
to separate the noise from the actual data. Filtering by Savitzky
Golay with different orders of polynomials amounts to altering the
data in different ways. It is difcult to know which order is better
other than the fact that low order polynomials are less accurate in
capturing fast dynamics.
Fig. 4. Distribution of an 8 by 8 sensor array on the back surface. The array occupies
a 42 mm by 42 mm square with 6 mm spacing. The edges of the plate are assumed
to be inaccessible to sensors. The IHCP assumes that the data from the 64 sensors
are available.
6. Code validation against three-dimensional numerical
solutions for a constant but non-uniform front surface heat ux
We have validated our program against a three-dimensional
numerical solution [35]. We consider a 3 mm thick by 60 mm by
60 mm square stainless steel plate with assumed constant thermal
conductivity of 17 W/m K and thermal diffusivity of 4.4 mm2/s. The
front surface of the square plate is subjected to the constant net
heat ux prole shown in Fig. 3. All four edges and the back surface
are isolated. Corresponding to these boundary conditions, a
numerical solution for the heat conduction problem is obtained
using a commercial nite element package. The solution includes
the temperature on the front and the back surfaces.
Temperature on the back surface is used as the original data for
input to the IHCP. We assume an 8 by 8 temperature sensor array
provides the measurement data. The sensors are distributed on the
back surface as shown in Fig. 4. Note that the sensor coverage does
not fully extend to the plate edges. The sensor data are sampled
with 0.1 s sampling interval.
Starting from the sampled data from the 8 by 8 sensor array, we
rst computed the 64 modal back surface temperatures by solving
Eq. (51). Next, the modal front surface temperature and heat ux
are computed for each of the 64 modes. Finally, the front surface
temperature and heat ux are calculated at locations opposite to
the sensors on the back surface.
The IHCP was solved using the iterative procedure. Sixteen iter-
ations were used. The data smoothing and differentiation were
conducted with SavitzkyGolay method using seven points tted
by fourth order polynomials. We especially point out that only
the sampled back surface temperature and the plate geometric
and material properties are used for solving the IHCP. The back

S8
Front IHC
Temp S7 Front
S6 Temp
800-850
750-800 Error
S5
700-750 @ 0.7 s
650-700 S4
600-650 2-4
550-600 S3 0-2
500-550 -2-0
450-500 S2 -4--2
400-450 -6--4
350-400 S1 -8--6
300-350 1 2 3 4 5 6 7 8 -10--8

(a) (b)
Fig. 5. Contour plots related to front surface temperature (K) at t = 0.7 s: (a) Truth temperature (peak = 841 K); (b) IHC temperature error rms is 4 K (over all points), and 9 K
at peak.

Front Net IHC Front

Heat Flux Net Heat
(constant) Flux Error
@0.7 s
450-500
400-450
350-400
300-350
250-300 4-6
200-250 2-4
150-200 0-2
100-150 -2-0
50-100 -4--2
0-50 -6--4
(a) (b)
Fig. 6. Contour plots related to front surface heat ux (W/cm2): (a) Constant net heat ux; (b) IHC heat ux error at t = 0.7 s. The rms error is 2.2 W/cm2, and is 5 W/cm2 at
peak.
3438 Z.C. Feng et al. / International Journal of Heat and Mass Transfer 54 (2011) 34313439
surface temperature data required for the IHCP were generated by
a direct heat conduction solution. The same solution furnished the
front surface temperature, which served as truth to evaluate the
accuracy of the IHC solution. The IHC results for front surface tem-
perature and heat ux were compared with the truth values. It was
found that the rms temperature error was 4 K (scoring all 64 sen-
sors at times from 0.2 to 0.8 s at 0.1 s intervals, inclusive). This is
0.5% of the peak temperature of 841 K. Contour plots of the front
surface temperature (truth and IHC error) at t = 0.7 s are shown
in Fig. 5. At this time, the peak temperatures on the front and back
surfaces are 841 K and 439 K, respectively. (In Figs. 5 and 6, values
are plotted at the intersections of the 8  8 grids. Then, contours
are drawn to help interpret the proles.)
Although the front surface heat ux was constant, the solution
of the IHCP does not use this information and treats the problem as
it does for dynamic front surface heat ux. The rms heat ux error
is 2.2 W/cm2 (0.4% of the peak 500 W/cm2) when scoring all 64
sensors at time of 0.7 s.
The IHC solutions for the temperature and heat ux are found to
have accuracies better than 1%. The IHC solution procedures pre-
sented above are also computationally efcient. The calculation
above was completed within 12 s when implemented in Matlab
on a Dell personal computer.
7. Conclusion
When the in-plane temperature gradient is smaller than the
temperature gradient across the thickness of the slab, the modal
representation provides a convenient method to reduce the
three-dimensional problem to one-dimensional problem. The
one-dimensional problem is easily solved through generalizing
the previously developed method [24]. The two-dimensional dis-
tributions of temperature and heat ux on the front surface can
be recovered by using Eqs. (15) and (16) with quantities hmn(s, 0)
and fmn(s) determined from Eqs. (33) and (34). It is found that
the method provided results which agree with analytical and
numerical results very accurately.
While we have demonstrated the accuracy of the proposed
method in recovering the front surface heat ux and temperature
when certain noise is assumed in the measurements, the accuracy
will deteriorate for thick and thermally insulating slabs and for
heat input with high rate of change and high gradients. Analysis
leading to a quantitative criterion on the validity of the method
is an important task yet to be carried out.
Acknowledgments
The authors would like to thank the Test Resource Management
Center (TRMC) Test and Evaluation/Science & Technology (T&E/
S&T) Program for their support. This work is funded by the T&E/
S&T Program through the US Army Program Executive Ofce for
Simulation, Training and Instrumentations contract number
W900KK-08-C-0002.
Appendix A. Analytical solutions for non-uniform constant heat
ux on the front surface
The one-dimensional equation in (17) is the same form as an
important problem in heat conduction of ns [29]. Analytical solu-
tions have been obtained for a few special cases. For example, if the
modal heat ux on the front surface is constant, i.e.
0 [23] A. Sutradhar, G.H. Paulino, L.J. Gray, Transient heat conduction in
fmn s fmn A1
The solution is given in [29] as the following:
coshcmn 1  Z 1 2
0
hmn s; Z fmn  2 ecmn s
cmn sinh cmn cmn
)
X 1
1k1 c2mn k2 p2 s
2 coskp1  Ze A2
2 2
k1 cmn k p
2
The above solution is valid for general mode numbers. For
(m, n) = (0, 0), taking the limit cmn ? 0, and recognizing that
X
1
1 p2
2
A3
k 6
k1
we obtain the solution for the temperature at depth within a slab
with a constant uniform front surface heat ux [30]:
" #
0 Z2 1 2 X1
coskpZ kp2 s
h00 s; Z f00 s Z  2 e A4
2 3 p k1 k
2
References
[1] E.M. Sparrow, A. Haji-Sheikh, T.S. Lundgren, The inverse problem in transient
heat conduction, ASME J. Appl. Mech. 86 (1964) 369375.
[2] J.V. Beck, B. Blackwell, C.R. St-Clair, Inverse Heat Conduction: Ill Posed
Problems, Wiley, New York, 1985.
[3] R. Pasquetti, C.L. Niliot, Boundary element approach for inverse heat
conduction problems: application to a bidimensional transient numerical
experiment, Numer. Heat Transfer, Part B 20 (1991) 29112919.
[4] O.M. Alifanov, Inverse Heat Transfer Problems, Springer-Verlag, Berlin/
Heidelberg, 1994.
[5] C.-Y. Yang, C.-K. Chen, The boundary estimation in two-dimensional inverse
heat conduction problems, J. Phys. D: Appl. Phys. 29 (1996) 333339.
[6] C.-H. Huang, S.-P. Wang, A three-dimensional inverse heat conduction problem
in estimating surface heat ux by conjugate gradient method, Int. J. Heat Mass
Transfer 42 (1999) 33873403.
[7] M.N. zisik, H.R.B. Orlande, Inverse Heat Transfer: Fundamentals and
Applications, Taylor & Francis, New York, 2000.
[8] A.F. Emery, A.V. Nenarokomov, T.D. Fadale, Uncertainties in parameter
estimation: the optimal experimental design, Int. J. Heat Mass Transfer 43
(2000) 33313339.
[9] M. Monde, H. Arima, Y. Mitsutake, Estimation of surface temperature and heat
ux using inverse solution for one-dimensional heat conduction, ASME J. Heat
Transfer 125 (2003) 213223.
[10] X. Xue, R. Luck, J.T. Berry, Comparisons and improvements concerning the
accuracy and robustness of inverse heat conduction algorithms, Inverse Prob.
Sci. Eng. 13 (2) (2005) 177199.
[11] A. Azimi, S.K. Hannani, B. Farhanieh, Structured multiblock body-tted grids
solution of transient inverse heat conduction problems in an arbitrary
geometry, Numer. Heat Transfer, Part B: Fundam. 54 (3) (2008) 260290.
[12] M. Samai, T. Loulou, A comparative study of heat ux and temperature based
objective functional to solve inverse heat conduction problems, Numer. Heat
Transfer, Part B: Fundam. 56 (1) (2009) 75104.
[13] S. Vakili, M.S. Gadala, Effectiveness and efciency of particle swarm
optimization technique in inverse heat conduction analysis, Numer. Heat
Transfer, Part B: Fundam. 56 (2) (2009) 119141.
[14] J. Zhou, Y. Zhang, J.K. Chen, Z.C. Feng, Inverse heat conduction in a nite slab
with measured back heat ux, in: Proceedings of the 47th AIAA Aerospace
Sciences Meeting, Orlando, FL, January 59, 2009.
[15] S.K. Kim, Resolving the nal time singularity in gradient methods for inverse
heat conduction problems, Numer. Heat Transfer, Part B: Fundam. 57 (1)
(2010) 7488.
[16] K.-C. Liu, C.-T. Lin, Solution of an inverse heat conduction problem in a bi-
layered spherical tissue, Numer. Heat Transfer, Part B: Fundam. 58 (10) (2010)
802818.
[17] J. Zhou, Y. Zhang, J.K. Chen, Z.C. Feng, Inverse heat conduction using measured
back surface temperature and heat ux, J. Thermophys. Heat Transfer 24
(2010) 95103.
[18] J. Zhou, Y. Zhang, J.K. Chen, Z.C. Feng, Inverse heat conduction in a composite
slab with pyrolysis effect and temperature-dependent thermophysical
properties, ASME J. Heat Transfer 132 (2010). 034502-1.
[19] Z.C. Feng, J.K. Chen, Y. Zhang, Real-time solution of heat conduction in a nite
slab for inverse analysis, Int. J. Therm. Sci. 49 (2010) 762768.
[20] M.N. zisik, Heat Conduction, second ed., Wiley-Interscience, New York, 1993.
[21] G.E. Myers, Analytical Methods in Conduction Heat Transfer, second ed.,
AMCHT Publications, Madison, WI, 1998.
[22] D. Maillet, S. Andre, J.C. Batsale, A. Degiovanni, C. Moyne, Thermal
Quadrupoles: Solving the Heat Equation through Integral Transforms, John
Wiley & Sons, New York, 2000.
homogeneous and non-homogeneous materials by the Laplace transform
 Z.C. Feng et al. / International Journal of Heat and Mass Transfer 54 (2011) 34313439
Galerkin boundary element method, Eng. Anal. Boundary Elem. 26 (2002)
119132.
[24] Z.C. Feng, J.K. Chen, Y. Zhang, Temperature and heat ux estimation from
sampled transient sensor measurements, Int. J. Therm. Sci. 49 (2010) 2385
2390.
[25] D.Q. Kern, A.D. Kraus, Extended Surface Heat Transfer, McGraw-Hill, New York,
1972.
[26] J.B. Conway, Functions of One Complex Variable, Graduate Texts in
Mathematics, second ed., vol. 11, Springer-Verlag, New York/Berlin, 1978.
[27] O.R. Burggraf, An exact solution of the inverse problem in heat conduction
theory and application, J. Heat Transfer 86C (1964) 373382.
[28] A.E. Segall, An inverse solution for determining arbitrary boundary-conditions
using a least-squares approach, J. Heat Transfer 127 (2005) 14031405.
[29] A.D. Kraus, A. Aziz, J. Welty, Extended Surface Heat Transfer, John Wiley &
Sons, Inc., New York, 2001 (Chapter 16).
View publication stats
3439
[30] A. Faghri, Y. Zhang, J.R. Howell, Advanced Heat and Mass Transfer, Global
Digital Press, Columbia, MO, 2010.
[31] A. Savitzky, M.J.E. Golay, Smoothing and differentiation of data by simplied
least squares procedures, Anal. Chem. 36 (1964) 16271639.
[32] P.A. Goray, General least-squares smoothing and differentiation by the
convolution (SavitzkyGolay) method, Anal. Chem. 62 (1990) 570573.
[33] M. Jakubowska, W.W. Kubiak, Adaptive-degree polynomial lter for
voltammetric signal, Anal. Chim. Acta 512 (2004) 241250.
[34] J. Luo, K. Ying, P. He, J. Bai, Properties of SavitzkyGolay digital differentiators,
Digital Signal Process. 15 (2005) 122126.
[35] J.L. Griggs, High energy laser target surface temperature estimation by a 1D
inverse heat conduction algorithm, in: Directed Energy Test and Evaluation
Conference, Albuquerque, NM, August 35, 2010.