Upload
salty
View
10
Download
0
Embed Size (px)
DESCRIPTION
project
Citation preview
Applied Mathematical Modelling 39 (2015) 5480–5494
Contents lists available at ScienceDirect
Applied Mathematical Modelling
journal homepage: www.elsevier .com/locate /apm
A one-dimensional inverse problem in composite materials:Regularization and error estimates
http://dx.doi.org/10.1016/j.apm.2015.01.0040307-904X/� 2015 Elsevier Inc. All rights reserved.
⇑ Corresponding author.E-mail address: [email protected] (Y.C. Hon).
X.T. Xiong a, W.X. Shi a, Y.C. Hon b,⇑a Department of Mathematics, Northwest Normal University, Lanzhou, Gansu, Chinab Department of Mathematics, City University of Hong Kong, Hong Kong, China
a r t i c l e i n f o a b s t r a c t
Article history:Received 29 November 2012Received in revised form 1 March 2014Accepted 5 January 2015Available online 14 January 2015
Keywords:Inverse problemIll-posednessError estimateRegularizationComposite materials
In this paper we investigate an inverse one-dimensional heat conduction problem in multi-layer medium. The inverse problem is first formulated in the frequency domain via Fouriertransform technique. An effective regularization method for the stable reconstruction ofsolution is given with proven error estimates. Several numerical examples are constructedto demonstrate the effectiveness of the proposed method.
� 2015 Elsevier Inc. All rights reserved.
1. Introduction
The transient temperature distribution in a composite medium consisting of several layers in contact has numerous appli-cations in engineering. For instance, it is important in the aerospace engineering to reconstruct the surface temperature andheat flux in high temperature composite materials. One of these works evolved as a boundary identification problem [1] inseveral-layer domain arises from steel-making industry in which the iron was heated and melted in one container composedof different materials. The result of this work is important to detect any corrosion inside the inner-surface of the containerand prevent any disaster damage due to the leakage of the steel fluid. Mathematically, these kinds of problems can be treatedas boundary identification problems. Another example is the design problem of shielded thermocouple which is a measure-ment device used for monitoring the temperature in the hostile environments [2]. The shielded thermocouple deviceconsists of composite materials each with different thermal property. In order to simulate the concrete measurement situ-ations, a mathematical model of the shielded thermocouple is usually formulated as an Inverse Heat Conduction Problem(IHCP) [3] in several-layers domain.
The IHCP arises in thermal manufacturing processes of solids and has attracted much attention [4–7]. In this inverse IHCPproblem, it is required to recover the surface temperature and heat flux on an inaccessible boundary from the measurementon an accessible boundary, which is also called as a non-characteristic Cauchy problem of heat equation. It is well knownthat non-characteristic Cauchy problem of heat equation is ill-posed [8] in the Hadamard sense that any ‘‘small’’ measure-ment error in the data can induce extremely ‘‘large’’ error in the solution. Under an additional condition, a continuousdependence of the solution on the Cauchy data can be obtained. This is called conditional stability [9]. In other words, in
X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494 5481
theory we can stably reconstruct the solution of the ill-posed problem under a priori condition which is called the ‘‘sourcecondition’’. Due to the severe ill-posedness of the problem, most classical numerical methods failed to produce satisfactorystable approximation to the solution of the Cauchy problem of heat conduction equation. Some kinds of regularizationstrategies [10] needs to be employed. For the IHCP in a single-layer domain, theoretical investigation and computationalimplementation have been well studied, e.g., [11–27]. In alternative to the single-layer IHCP, the multi-layer IHCP is muchmore difficult in both numerical and theoretical studies. In general, the IHCP in multi-layer domain will be dissolved into anIHCP in each layer and the solution the multi-layer IHCP will be obtained by solving the IHCP in each single layer. Thisapproach is computational inefficient but the author can refer to the recent works [28–30] in which the authors used themethod of fundamental solutions combined with regularization to solve each IHCP layer by layer.
This paper aims at establishing a mathematical framework to recover the surface temperature and heat flux on the inac-cessible boundary in a two-layer domain [31] as follow.
Consider a two-layer body that consists of the first layer in 0 6 x 6 l1 and the second layer in l1 6 x 6 l2. The two layersare in perfect thermal contact at x ¼ l1 as displayed in Fig. 1.
Let k1; k2 > 0 be the thermal conductivities and a1;a2 > 0 be the thermal diffusivities for the first and second layer,respectively. The temperature distributions in the first and the second layers are denoted by u1ðx; tÞ and u2ðx; tÞ respectively.These temperature distributions satisfy the following partial differential equations in the two domains D1 :¼ fxj0 6 x 6 l1gand D2 :¼ fxjl1 6 x 6 l2g:
@u1
@t� a1
@2u1
@x2 ¼ 0; 0 < x < l1; t > 0; ð1:1Þ
@u2
@t� a2
@2u2
@x2 ¼ 0; l1 < x < l2; t > 0; ð1:2Þ
subject to the initial and boundary conditions
u1ðx;0Þ ¼ u2ðx;0Þ ¼ 0; 0 < x < l2; ð1:3Þu2ðl2; tÞ ¼ gðtÞ; t > 0; ð1:4Þ@u2
@xðl2; tÞ ¼ 0; t > 0; ð1:5Þ
u1ðl1; tÞ ¼ u2ðl1; tÞ; t > 0; ð1:6Þ
k1@u1
@xðl1; tÞ ¼ k2
@u2
@xðl1; tÞ; t > 0: ð1:7Þ
For the similar IHCP, Shcheglov [32,33] analyzed the convergence of the problem by using a hyperbolic equation perturba-tion method. For 2D IHCP in two-layer Cartesian bodies, some numerical approaches have been proposed [34,35]. To theknowledge of the authors, the convergence rate on the IHCP in multi-layer domain has not yet been given.
In this paper, we aim at obtaining an analytical solution to the above IHCP in multi-layer domain via Laplace and Fouriertransform techniques. Due to the severe ill-posedness of the problem, a regularization strategy is derived for the stablereconstruction of the solution. For illustration, we present the regularization method based on the analytical solution forthe IHCP in a two-layer domain in which we adapt a spectral regularization method and derive the error estimate withproven optimal order. The proposed method is simple and effective [14]. Furthermore, the proposed method can easily begeneralized to solve IHCP in the multi-layer domain.
The paper is organized as follows. We first give a theoretical analysis on the IHCP problem in Section 2. In Section 3, thespectral regularization technique is adopted for the reconstruction of stable solution with proven error estimates. Finally,several numerical examples are constructed in the last section to demonstrate the validity of the proposed regularizationmethod although the so-called ‘‘inverse crime’’ [36] has been used.
0
D2D1
l1xl2
Lateral Interface Lateral
Fig. 1. Schematic illustration of a one-dimensional inverse heat conduction problem in a two-layer domain.
5482 X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494
2. Mathematical analysis of the problem
Throughout this paper, we suppose that the exact data g 2 L2ð0;1Þ. As a solution of the problem (1.1)–(1.7), the functionsu1ðx; tÞ;u2ðx; tÞ satisfy (1.1)–(1.7) in the classical sense. Assume that, for a fixed x 2 ½0; l2�, the functions u1ðx; �Þ;u2ðx; �Þ andtheir derivatives @u1ðx;�Þ
@x ; @u2ðx;�Þ@x belong to L2ð0;1Þ.
The inverse problem is to seek the solutions @u1@x ðx; tÞ; u1ðx; tÞ for 0 6 x 6 l1 in the space L2ð0;1Þ from the given data
gðtÞ 2 L2ð0;1Þ and the insulated condition at the accessible boundary x ¼ l2. In practice, gðtÞ can only be measured withmeasurement errors to give some data functions gdðtÞ 2 L2ð0;1Þ, for which
kgdð�Þ � gð�Þk 6 d; ð2:1Þ
where the constant d > 0 represents a bound on the measurement error, k � k denotes the L2-norm, and there exists a con-stant E > 0, such that the following a priori bound exist for the problem
ku1ð0; �Þkp 6 E; ð2:2Þ
with p P 0, where k � kp denotes the norm of Sobolev space HpðRÞ (when p ¼ 0;H0ð�Þ ¼ L2ð�Þ).In fact, in the domain D2, we have
@u2
@t� a2
@2u2
@x2 ¼ 0; l1 < x < l2; t > 0 ð2:3Þ
u2ðx;0Þ ¼ 0; l1 < x < l2; ð2:4Þu2ðl2; tÞ ¼ gðtÞ; t > 0; ð2:5Þ@u2
@xðl2; tÞ ¼ 0; t > 0: ð2:6Þ
Applying the Laplace transform to both sides of (2.3) with respect to t and according to the property of Laplace transform,we get
sU2ðx; sÞ � u2ðx;0Þ � a2@2U2
@x2 ðx; sÞ ¼ 0; ð2:7Þ
where s denotes the variable of Laplace transform on t. Applying the homogenous initial condition, we have
a2@2U2
@x2 ðx; sÞ ¼ sU2ðx; sÞ; ð2:8Þ
which is a second-order ordinary differential equation. From the boundary conditions, we obtain
U2ðx; sÞ ¼ coshðffiffiffiffiffiffiffiffiffiffis=a2
pðl2 � xÞÞGðsÞ; ð2:9Þ
whereffiffiffiffiffiffiffiffiffiffis=a2
pis taken to be the principal square root and GðsÞ is the Laplace transform of gðtÞ.
Throughout this paper, we extend all the functions to the whole line �1 < t <1. Let
f ðnÞ ¼ 1ffiffiffiffiffiffiffi2pp
Z 1
�1f ðtÞe�intdt
be the Fourier transform of the function f ðtÞ 2 L2ðRÞ.For the function f ðtÞ which vanishes on the negative t axis, the Fourier and Laplace transforms are related via
FðinÞ ¼ffiffiffiffiffiffiffi2pp
f ðnÞ:
Therefore, from (2.9), on setting s ¼ in, the solution of problem (2.3)–(2.6) can be formulated in the frequency domain forl1 6 x 6 l2:
u2ðx; nÞ ¼ coshððl2 � xÞ
ffiffiffiffiffiina2
sÞgðnÞ; ð2:10Þ
where
ffiffiffiffiffiina2s¼ jn=ð2a2Þj
12ð1þ isgnðnÞÞ; the symbol sgn is the usual sign function: ð2:11Þ
Inverse Fourier transform on u2ðx; nÞ gives u2ðx; tÞ. From (2.10), we have for l1 6 x 6 l2:
@u2
@xðx; nÞ ¼ �
ffiffiffiffiffiina2
ssinhððl2 � xÞ
ffiffiffiffiffiina2
sÞgðnÞ: ð2:12Þ
X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494 5483
In the domain D1;u1 satisfies
@u1
@t� a1
@2u1
@x2 ¼ 0; 0 < x < l1; t > 0 ð2:13Þ
u1ðx;0Þ ¼ 0; 0 < x < l1; ð2:14Þu1ðl1; tÞ ¼ u2ðl1; tÞ; t > 0; ð2:15Þ
k1@u1
@xðl1; tÞ ¼ k2
@u2
@xðl1; tÞ; t > 0: ð2:16Þ
Similarly, from Laplace transform we have
sU1ðx; sÞ � u1ðx;0Þ � a1@2U1
@x2 ðx; sÞ ¼ 0; 0 < x < l1; ð2:17Þ
U1ðl1; sÞ ¼ U2ðl1; sÞ; ð2:18Þ
k1@U1
@xðl1; sÞ ¼ k2
@U2
@xðl1; sÞ: ð2:19Þ
From the relation between the Fourier transform and the Laplace transform, we have for 0 6 x 6 l1
u1ðx; nÞ ¼ cosh
ffiffiffiffiffiina1
sðl1 � xÞ
0@ 1A cosh
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1AgðnÞ þ k2ffiffiffiffiffia1p
k1ffiffiffiffiffia2p sinh
ffiffiffiffiffiina1
sðl1 � xÞ
0@ 1A sinh
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1AgðnÞ;
ð2:20Þ
which yields
@u1ðx; nÞ@x
¼ �
ffiffiffiffiffiina1
ssinh
ffiffiffiffiffiina1
sðl1 � xÞ
0@ 1A cosh
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1AgðnÞ
� k2ffiffiffiffiffia1p
k1ffiffiffiffiffia2p
ffiffiffiffiffiina1
scosh
ffiffiffiffiffiina1
sðl1 � xÞ
0@ 1A sinh
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1AgðnÞ: ð2:21Þ
Taking inverse Fourier transform on (2.20) and (2.21), we have u1ðx; tÞ and @u1ðx;tÞ@x . In fact, we need to prove that the inverse
Fourier transforms u1ðx; tÞ and u2ðx; tÞ satisfy (1.3). In fact, if bgðnÞ decays rapidly in the frequency domain, we can prove thatu1ðx; tÞ is a unique solution to the problem. Refer [37] for the detailed proof. However. since the given data usually containmeasurement error, we cannot expect the error to have the same decay rate in frequency domain as the exact data bgðnÞ. Inother words, the corresponding perturbed solution will in general not lie in L2ðRÞ. Therefore, the problem considered in thispaper is an ill-posed problem.
Denote
k :¼ k2ffiffiffiffiffia1p
k1ffiffiffiffiffia2p ; ð2:22Þ
bAðx; nÞ :¼ cosh
ffiffiffiffiffiina1
sðl1 � xÞ
0@ 1A cosh
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1Aþ k sinh
ffiffiffiffiffiina1
sðl1 � xÞ
0@ 1A sinh
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1A; ð2:23Þ
bBðx; nÞ :¼ �
ffiffiffiffiffiina1
ssinh
ffiffiffiffiffiina1
sðl1 � xÞ
0@ 1A cosh
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1A� k
ffiffiffiffiffiina1
scosh
ffiffiffiffiffiina1
sðl1 � xÞÞ sinhð
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1A: ð2:24Þ
Formulating (2.20) and (2.21) as operator equations, we have
u1ðx; nÞ ¼ bAðx; nÞgðnÞ; ð2:25Þ@u1ðx; nÞ
@x¼ bBðx; nÞgðnÞ: ð2:26Þ
From above two equations, we can see that bAðx; nÞ; bBðx; nÞ : L2ðRÞ ! L2ðRÞ are unbounded multiplication operators for0 6 x 6 l1.
In order to obtain the error estimates, first we need two inequalities on bAðx; nÞ and bBðx; nÞ.
5484 X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494
Proposition 2.1 [31]. Let 0 < x < l1, then
ð1Þ jbAðx; nÞj 6 94ð1þ kÞe
ðl1�xÞffiffiffiffiffijnj
2a1
qeðl2�l1Þ
ffiffiffiffiffijnj
2a2
q; ð2:27Þ
ð2Þ jbBðx; nÞj 6 94ð1þ kÞ
ffiffiffiffiffiffiffiffijnj
2a1
seðl1�xÞ
ffiffiffiffiffijnj
2a1
qeðl2�l1Þ
ffiffiffiffiffijnj
2a2
q: ð2:28Þ
3. Regularization method and error estimates
In this section, we prove the error estimates for the regularized solutions between the solution u1ðx; tÞ and the heat fluxsolution @u1
@x ðx; tÞ, respectively.
3.1. On temperature u1ðx; tÞ
In the frequency domain, the spectral method for solving the problem can be given as:
unmax1 ðx; nÞ ¼ u1ðx; nÞvmax;
@u1
@x
nmax
ðx; nÞ ¼ @u1
@xðx; nÞvmax;
where nmax is the regularization parameter, vmax denotes the characteristic function on the interval ½�nmax; nmax�, i.e.,
vmaxðnÞ ¼1; �nmax 6 n 6 nmax;
0; jnj > nmax:
�
Let us recall dfðtÞ ¼ Fðf ðtÞÞ ¼ bf ðnÞ ¼ 1ffiffiffiffiffiffiffi2pp
Z 1
�1f ðtÞe�intdt;
represents the Fourier transform of f ðtÞ 2 L2ðRÞ, and
f ðtÞ ¼ F�1ðbf ðnÞÞ ¼ 1ffiffiffiffiffiffiffi2pp
Z 1
�1
bf ðnÞeintdn;
represents the corresponding inverse Fourier transform.Denote
Rðu1ðx; tÞÞ ¼ F�1ðunmax1 ðx; nÞÞ ¼ 1ffiffiffiffiffiffiffi
2pp
Z 1
�1u1ðx; nÞvmaxeinxdn; ð3:1Þ
R@u1
@xðx; tÞ
� �¼ F�1ð@u1
@x
nmax
ðx; nÞÞ ¼ 1ffiffiffiffiffiffiffi2pp
Z 1
�1
@u1
@xðx; nÞvmaxeinxdn: ð3:2Þ
In order to derive the error estimates, we need a lower bound on bAð0; nÞ.Proposition 3.1. If jnj > ð
ffiffiffiffiffiffi2a2
pl2�l1
ln 2Þ2
, it holds
jbAð0; nÞjP 1þ k8
el1
ffiffiffiffiffijnj
2a1
qeðl2�l1Þ
ffiffiffiffiffijnj
2a2
q: ð3:3Þ
Proof. Let a :¼ l1
ffiffiffiffiffiffijnj
2a1
q; b :¼ ðl2 � l1Þ
ffiffiffiffiffiffijnj
2a2
q; a1 :¼ asgnðnÞ; b1 :¼ bsgnðnÞ, we have
jbAð0;nÞj ¼ cosh
ffiffiffiffiffiina1
sl1
0@ 1A cosh
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1Aþ k sinh
ffiffiffiffiffiina1
sl1Þ sinhð
ffiffiffiffiffiina2
sðl2 � l1Þ
0@ 1A������������
¼ cosh l1
ffiffiffiffiffiffiffiffijnj
2a1
s !ð1þ isgnðnÞÞ
!cosh ðl2 � l1Þ
ffiffiffiffiffiffiffiffijnj
2a2
s !ð1þ isgnðnÞÞ
!þ k sinh l1
ffiffiffiffiffiffiffiffijnj
2a1
s !ð1þ isgnðnÞÞ
!������ sinh ðl2 � l1Þ
ffiffiffiffiffiffiffiffijnj
2a2
s !ð1þ isgnðnÞÞ
!�����¼ j coshðaþ ia1Þ coshðbþ ib1Þ þ k sinhðaþ ia1Þ sinhðbþ ib1Þj¼ jA1 þ kA2j;
X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494 5485
where
A1 ¼ coshðaþ ia1Þ coshðbþ ib1Þ ¼eaþia1 þ e�a�ia1
2ebþib1 þ e�b�ib1
2
¼ 12½cosh ðaþ bÞ þ iða1 þ b1Þð Þ þ cosh ða� bÞ þ iða1 � b1Þð Þ�;
A2 ¼ sinhðaþ ia1Þ sinhðbþ ib1Þ ¼eaþia1 � e�a�ia1
2ebþib1 � e�b�ib1
2
¼ 12½cosh ðaþ bÞ þ iða1 þ b1Þð Þ � cosh ða� bÞ þ iða1 � b1Þð Þ�:
Indeed, first we have
j coshððaþ bÞ þ iða1 þ b1ÞÞj � j coshðða� bÞ þ iða1 � b1ÞÞj ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffifficosh2ðaþ bÞ cos2ða1 þ b1Þ þ sinh2ðaþ bÞ sin2ða1 þ b1Þ
q�
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffifficosh2ða� bÞ cos2ða1 � b1Þ þ sinh2ða� bÞ sin2ða1 � b1Þ
qP
12ðeaþb � e�ðaþbÞÞ � 1
2ea�b ¼ 1
2eaþb � 1
2e�ðaþbÞ þ ea�b� �
P12
eaþb � ea�b:
If jnj > ðffiffiffiffiffiffi2a2
pl2�l1
ln 2Þ2
, i.e., b > ln 2, then 14 eaþb � ea�b P 0, therefore
jbAð0; nÞj ¼ 12jð1þ kÞ coshððaþ bÞ þ iða1 þ b1ÞÞ þ ð1� kÞ coshðða� bÞ þ iða1 � b1ÞÞj
P12jð1þ kÞ coshððaþ bÞ þ iða1 þ b1ÞÞj � jð1� kÞ coshðða� bÞ þ iða1 � b1ÞÞjj j
P1þ k
2j coshððaþ bÞ þ iða1 þ b1ÞÞj � j coshðða� bÞ þ iða1 � b1ÞÞjj jP 1þ k
212
eaþb � ea�b
� �P
1þ k8
eaþb:
By the definitions of a; b, we have the inequality holds. h
For readability, we stress that the following assumptions: the noisy data gdðtÞ satisfies kgdðtÞ � gðtÞk 6 d and the ‘‘sourcecondition’’ ku1ð0; �Þkp 6 E; ðp P 0Þ holds. Meanwhile, we define a symbol:
‘ :¼ l1ffiffiffiffiffiffiffiffi2a1p þ ðl2 � l1Þffiffiffiffiffiffiffiffi
2a2p : ð3:4Þ
The following lemma shows the stability of the regularization solution.
Lemma 3.1. If Rðu1ðx; tÞÞ;Rðud1ðx; tÞÞ are the regularized solutions with data given by gðtÞ and gdðtÞ, respectively, and
kgdðtÞ � gðtÞk 6 d, if we choose the regularization parameter
ffiffiffiffiffiffiffiffi2a2pl2 � l1ln 2
� �2
< nmax ¼1‘
lnEd
lnEd
� ��2p !" #2
; ð3:5Þ
then
kRðud1ðx; tÞÞ � Rðu1ðx; tÞÞk 6 c1 ln
Ed
� ��2p !1�
xffiffiffiffiffi2a1
p‘
E1�
xffiffiffiffiffi2a1
p‘ d
xffiffiffiffiffi2a1
p‘ ; ð3:6Þ
where c1 is a positive constant which is independent on d and E.
Proof. Using Parseval’s equality, we have
kRðud1ðx; �ÞÞ � Rðu1ðx; �ÞÞk2 ¼ k dRðud
1ðx; �ÞÞ � dRðu1ðx; �ÞÞk2 ¼Z nmax
�nmax
jbAðx; nÞj2jbgdð�Þ � bgð�Þj2dn
6 supjnj<nmaxjbAðx; nÞj2 Z nmax
�nmax
jbgdð�Þ � bgð�Þj2dn:
5486 X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494
From kgdð�Þ � gð�Þk 6 d, we have
kRðud1ðx; �ÞÞ � Rðu1ðx; �ÞÞk2
68116ð1þ kÞ2e
2ðl1�xÞffiffiffiffiffiffiffiffijnmax j
2a1
qe
2ðl2�l1Þffiffiffiffiffiffiffiffijnmax j
2a2
q Z nmax
�nmax
jbgdð�Þ � bgð�Þj2dn
68116ð1þ kÞ2e
2l1�xffiffiffiffiffi
2a1
p þ l2�l1ffiffiffiffiffi2a2
p� � ffiffiffiffiffiffiffi
nmax
pd2
By nmax ¼ ½1‘ lnðEd ðln EdÞ�2pÞ�
2, we have
xxffiffiffiffiffip
kRðud1ðx; �ÞÞ � Rðu1ðx; �ÞÞk2
68116ð1þ kÞ2d2eð2�2
xffiffiffiffiffi2a1
p‘Þ ln E
d lnEdð Þ�2p
� �6
8116ð1þ kÞ2d2 E
d
� �2�2
ffiffiffiffiffi2a1
p‘
lnEd
� ��2p !2�2
2a1‘
6 c21 ln
Ed
� ��2p !2�2
xffiffiffiffiffi2a1
p‘
E2�2
xffiffiffiffiffi2a1
p‘ d2
xffiffiffiffiffi2a1
p‘ ;
i.e., The proof of (3.6) is completed. h
The following lemma shows the convergence of the regularized solution.
Lemma 3.2. Suppose Rðu1ðx; tÞÞ; u1ðx; tÞ are the regularized solution and the exact solution with the exact data gðtÞ respectively,and nmax is chosen as (3.5), it holds error estimate for d! 0
kRðu1ðx; tÞÞ � u1ðx; tÞk 6 c2 lnEd
� ��2p" #1�
xffiffiffiffiffi2a1
p‘
E1�
xffiffiffiffiffi2a1
p‘ d
xffiffiffiffiffi2a1
p‘ ð1þ oð1ÞÞ; ð3:7Þ
where c2 is a positive constant which is independent on d and E.
Proof. Using the Parseval’s equality, we have the follow relation
kRðu1ðx; �ÞÞ � u1ðx; �Þk2 ¼ k dRðu1ðx; �ÞÞ � du1ðx; �Þk2 ¼Zjnj>nmax
jbAðx; nÞj2jbgðnÞj2dn:
Since bu1ð0; nÞ ¼ bAð0; nÞbgðnÞ, then � �
kRðu1ðx; �ÞÞ � u1ðx; �Þk2 ¼ k dRðu1ðx; �ÞÞ � du1ðx; �Þk2 ¼Zjnj>nmax
bAðx; nÞbAð0; nÞ���� ����
2
jbu1ð0; nÞj2dn:
According to ku1ð0; �Þkp 6 E, it yields
kRðu1ðx; �ÞÞ � u1ðx; �Þk2 ¼ k dRðu1ðx; �ÞÞ � du1ðx; �Þk2 ¼Zjnj>nmax
bAðx; nÞbAð0; nÞ�����
�����ð1þ n2Þ�p2
" #2
ð1þ n2Þpjbu1ð0; nÞj2dn
6 supjnj>nmax
bAðx; nÞbAð0; nÞ�����
�����2
n�2p
24 35Zjnj>nmax
ð1þ n2Þpjbu1ð0; nÞj2dn 6 E2supjnj>nmax
bAðx; nÞbAð0; nÞ�����
�����2
n�2p
8<:9=;:
Using Proposition 2.1 and 3.1, then
supjnj>nmax
bAðx; nÞbAð0; nÞ�����
����� 6 ce�x
ffiffiffiffiffijnj
2a1
q; c is a constant:
By (3.5), we have 0 1 0 1
kRðu1ðx; �ÞÞ � u1ðx; �Þk26 c1‘
lnEdðln E
d�2p
!" #�2p@ A2
E2 exp �2
xffiffiffiffiffiffi2a1
p
‘ln
Edðln E
d�2p
!@ A
¼ c22 ln
Edðln E
d�2p
! !�2p24 352
lnEd
� ��2p" #�2
xffiffiffiffiffi2a1
p‘
E2�2
xffiffiffiffiffi2a1
p‘ d2
xffiffiffiffiffi2a1
p‘ :
The proof has been completed. h
X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494 5487
Now we can claim the main conclusion of this subsection. For the temperature u1ðx; tÞ, we have the following errorestimate:
Theorem 3.1. Suppose that Rðud1ðx; tÞÞ; u1ðx; tÞ are the regularized solution with the noisy data and the exact solution with the
exact data respectively, and the a priori condition ku1ð0; �Þkp 6 E with p > 0 holds, ifffiffiffiffiffiffi2a2
pl2�l1
ln 2� �2
< nmax ¼ 1‘ ln E
d ln Ed
� ��2p h i2
,
then for d! 0 it holds the error estimate
kRðud1ðx; tÞÞ � u1ðx; tÞk 6 c3 ln
Ed
� ��2p !1�
xffiffiffiffiffi2a1
p‘
E1�
xffiffiffiffiffi2a1
p‘ d
xffiffiffiffiffi2a1
p‘ ð1þ oð1ÞÞ: ð3:8Þ
where c3 is a positive constant which is independent on d and E.
Proof. Due to Parseval’s equation and triangle inequality, it yields
kRðud1ðx; tÞÞ � u1ðx; tÞk
6 kRðud1ðx; tÞÞ � Rðu1ðx; tÞÞk þ kRðu1ðx; tÞÞ � u1ðx; tÞk:
According to Lemma 3.1 and 3.2 (see (3.6)), we have
kRðud1ðx; tÞÞ � u1ðx; tÞk 6 c1 ln
Ed
� ��2p !1�
xffiffiffiffiffi2a1
p‘
E1�
xffiffiffiffiffi2a1
p‘ d
xffiffiffiffiffi2a1
p‘ þ c2 ln
Edðln E
d�2p
!�2p24 35 ln
Ed
� ��2p" #� xffiffiffiffiffi
2a1
p‘
E1�
xffiffiffiffiffi2a1
p‘ d
xffiffiffiffiffi2a1
p‘
¼ c3 lnEd
� ��2p !1�
xffiffiffiffiffi2a1
p‘
E1�
xffiffiffiffiffi2a1
p‘ d
xffiffiffiffiffi2a1
p‘ ð1þ oð1ÞÞ; d! 0:
Thus, we have the conclusion (3.8). h
Remark 3.1. When p ¼ 0, i.e., ku1ð0; �Þk 6 E holds, for 0 < x < l1 we have the following error estimate
kRðud1ðx; tÞÞ � u1ðx; tÞk 6 c3E1�
xffiffiffiffiffi2a1
p‘ d
xffiffiffiffiffi2a1
p‘ :
It is a Hölder-type error estimate.
Remark 3.2. If x ¼ 0 and kuð0; �Þkp 6 E with p > 0 holds, we have the following error estimate
kRðud1ð0; tÞÞ � u1ð0; tÞk 6 c3E ln
Ed
� ��2p
ð1þ oð1ÞÞ; d! 0:
It is a logarithm-type error estimate.Obviously, above two results in the Remarks are consistent with the results obtained in single-layer case [13].In the forthcoming subsection, we will deal with the problem on estimating the regularized heat flux. Usually, this prob-
lem is more difficult than the problem of estimating the regularized temperature (see [27]). Using the proposed method forsingle-layer case [27], fortunately, we can obtain the corresponding results for the multi-layer case.
3.2. On heat flux @u1@x ðx; tÞ
In this subsection, a stronger a priori condition ku1ð0; �Þkp 6 E for p > 1=2 is assumed. Here, we need to estimate @u1@x ðx; tÞ.
Similarly, we first give the stability result.
Lemma 3.3. Suppose that R @ud1ðx;tÞ@x
; R @u1ðx;tÞ
@x
are the regularized solutions with the noisy data and the exact data respectively,
and kgdðtÞ � gðtÞk 6 d, then if the nmax satisfiesffiffiffiffiffiffi2a2
pl2�l1
ln 2� �2
< nmax ¼ 1‘ ln E
d ln Ed
� ��2p h i2
, which is the same as (3.5), it holds for
d! 0x
!
R@ud1
@xðx; tÞ
� �� R
@u1
@xðx; tÞ
� ����� ���� 6 c4 lnEd
� �1�2p 1�
ffiffiffiffiffi2a1
p‘
E1�
xffiffiffiffiffi2a1
p‘ d
xffiffiffiffiffi2a1
p‘ ð1þ oð1ÞÞ; ð3:9Þ
where c4 is a positive constant which is independent on d and E.
5488 X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494
Proof. From Parseval’s equality, it yields
R@ud
1
@xðx; �Þ
� �� R
@u1
@xðx; �Þ
� ����� ����2
¼Z nmax
�nmax
jbBðx; nÞj2jbgdðnÞ � bgðnÞj2dn 6 supjnj<nmaxjbBðx; nÞj2 Z nmax
�nmax
jbgdðnÞ � bgðnÞj2dn
6
8116 ð1þ kÞ2
2a1jnmaxje
2ðl1�xÞffiffiffiffiffiffiffiffijnmax j
2a1
qe
2ðl2�l1Þffiffiffiffiffiffiffiffijnmax j
2a2
q Z nmax
�nmax
jbgdðnÞ � bgðnÞj2dn:
Noting that nmax ¼ 1‘
ln Ed ln E
d
� ��2p h i2
, we have
R@ud
1
@xðx; �Þ
� �� R
@u1
@xðx; �Þ
� ����� ����2
ð3:10Þ
6
8116 ð1þ kÞ2
2a1d2jnmaxj exp 2ðl1 � xÞ
ffiffiffiffiffiffiffiffiffiffiffiffijnmaxj2a1
s !exp 2ðl2 � l1Þ
ffiffiffiffiffiffiffiffiffiffiffiffijnmaxj2a2
s !
¼ 9ð1þ kÞ4ffiffiffiffiffiffiffiffi2a1p
� �2 1‘
lnEd
lnEd
� ��2p !" #2
d2 exp 2ðl1 � xÞ
ffiffiffiffiffiffiffiffiffiffiffiffijnmaxj2a1
s !exp 2ðl2 � l1Þ
ffiffiffiffiffiffiffiffiffiffiffiffijnmaxj2a2
s !
¼ c24 ln
Edðln E
d�2p
!" #2
d2 exp 2ðl1 � xÞ
ffiffiffiffiffiffiffiffiffiffiffiffijnmaxj2a1
s !exp 2ðl2 � l1Þ
ffiffiffiffiffiffiffiffiffiffiffiffijnmaxj2a2
s !¼ c2
4 lnEdðln E
d�2p
!" #2
E2�2
xffiffiffiffiffi2a1
p‘ d2
xffiffiffiffiffi2a1
p‘
¼ c24 ln
Ed
� �� 2�4p 1�
xffiffiffiffiffi2a1
p‘
!E2�2
xffiffiffiffiffi2a1
p‘ d2
xffiffiffiffiffi2a1
p‘ ð1þ oð1ÞÞ; d! 0:
Thus (3.9) holds. h
Secondly, we have the following convergence result.
Lemma 3.4. Suppose that R @u1@x ðx; tÞ� �
; @u1@x ðx; tÞ are the regularized solution and the exact solution with the exact data respectively,
and the a priori condition kuð0; tÞkp 6 E with p > 12 holds, if the regularization parameter nmax is chosen as (3.5), then for d! 0 it
holds xffiffiffiffiffi2ap
xffiffiffiffiffip xffiffiffiffiffip
kRð@u1@xðx; tÞÞ � @u1
@xðx; tÞk 6 c5ðln
EdÞ
1�2pð1� 1‘ Þ
E1�2a1‘ d
2a1‘ ð1þ oð1ÞÞ; ð3:11Þ
where c5 is a positive constant which is independent on d and E.
Proof. As before, we have
R@u1
@xðx; tÞ
� �� @u1
@xðx; tÞ
���� ����2
¼Zjnj>nmax
jbBðx; nÞbgðnÞj2dn; ð3:12Þ
According to bu1ðx; nÞ ¼ bAðx; nÞbgðnÞ, bgðnÞ ¼ bu1ð0;nÞbAð0;nÞ , we have
R@u1
@xðx; tÞ
� �� @u1
@xðx; tÞ
���� ����2
¼Zjnj>nmax
bBðx; nÞbAð0; nÞ�����
�����2
jcu1ð0; nÞj2dn;
Due toffiffi2p
a2l2�l1
ln 2 2
< nmax ¼ 1‘
ln Ed ðln E
d�2p
h i2, and ku1ð0; �Þkp 6 E, it yields
R@u1
@xðx; tÞ
� �� @u1
@xðx; tÞ
���� ����2
6 c2Zjnj>nmax
ffiffiffiffiffiffiffiffin
2a1
se�x
ffiffiffiffiffin
2a1
q������������
2
ð1þ n2Þ�pð1þ n2Þpjbu1ð0; nÞj2dn
61
2a1c2Zjnj>nmax
jnj1�2pe�2x
ffiffiffiffiffijnj
2a1
qð1þ n2Þpjbu1ð0; nÞj2dn 6
12a1
c2supjnj>nmaxjnj1�2pe
�2x
ffiffiffiffiffijnj
2a1
q Zjnj>nmax
ð1þ n2Þpjcu1ð0; nÞj2dn
61
2a1c2E2n1�2p
max e�2x
ffiffiffiffiffiffiffinmax2a1
q6 c2
5½lnðEdÞ�
2�4pð1�
xffiffiffiffiffi2a1
p‘ÞE2�2
xffiffiffiffiffi2a1
p‘ d2
xffiffiffiffiffi2a1
p‘ ð1þ oð1ÞÞ; d! 0:
Thus, we have (3.11). h
Now we have the main conclusion of this subsection.
X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494 5489
Theorem 3.2. Suppose R @ud1ðx;tÞ@x
; @u1ðx;tÞ
@x are the regularized solution with the noisy data and the exact solution with the exact data
respectively, and kgdðtÞ � gðtÞk 6 d; ku1ð0; �Þkp 6 E with p > 12 hold, if
ffiffiffiffiffiffi2a2
pl2�l1
ln 2� �2
< nmax ¼ 1‘ ln E
d ðln EdÞ
� ��2ph i2
, we have the
following error estimate for d! 0 !
R@ud1
@xðx; tÞ
� �� @u1
@xðx; tÞ
���� ���� 6 c6 lnEd
� �1�2p 1�
xffiffiffiffiffi2a1
p‘
E1�
xffiffiffiffiffi2a1
p‘ d
xffiffiffiffiffi2a1
p‘ ð1þ oð1ÞÞ; ð3:13Þ
where c6 is a positive constant which is independent on d and E.
Proof. Similar to Theorem 3.1, according to the triangle inequality and Lemma 3.3 and 3.4, we have (3.13). h
In practice, we are more interesting in the convergence rate at x ¼ 0. By Theorem 3.2, we have the following remark.
Remark 3.3. Suppose R @ud1ð0;tÞ@x
; @u1ð0;tÞ
@x are the regularized solution with the noisy data and the exact solution with the exact
data respectively, and kgdðtÞ � gðtÞk 6 d; ku1ð0; �Þkp 6 E with p > 12 hold, if nmax is chosen as (3.5), we have the following error
estimate for d! 0
R@ud
1
@xð0; tÞ
� �� @u1
@xð0; tÞ
���� ���� 6 c6 lnEd
� �1�2p
ð1þ oð1ÞÞ; d! 0: ð3:14Þ
4. Numerical examples
In general, for an ill-posed problem, we can only obtain the worst-case error for regularized methods. However, in prac-tical computation, the errors in numerical computation of regularization methods are far less than the worst-case errors. Thisphenomenon has been observed in many literatures, e.g. [6]. In this paper, the formula (3.5) is difficult for computationexcept that one can compute the exact E. In the following examples, we found that the regularization parameter nmax withina large range is available for good numerical reconstruction. Moreover, in practice the test of an inversion process avoidingthe ‘‘inverse crime’’ can be done using a model for the numerically simulated data and a different one to invert the data. Inthis paper, we can use a simple traditional numerical method (e.g. finite difference method), which is different from Fouriermethod, to solve the direct problem. However, this is a time-consuming task [38]. Based on the theoretical analysis derivedin the above section, we construct in the following several numerical examples to verify the convergence of the proposedmethod.
The following numerical implementation is performed by using Matlab. The regularized solutions are computed by usingthe discrete Fast Fourier Transform (FFT) and inverse discrete Fast Fourier Transform according to the formulas (3.1) and(3.2) in Section 3.
Firstly we solve the corresponding direct problem
@u1
@t� a1
@2u1
@x2 ¼ 0; 0 < x < l1; t > 0; ð4:1Þ
@u2
@t� a2
@2u2
@x2 ¼ 0; l1 < x < l2; t > 0; ð4:2Þ
subject to the initial and boundary conditions
u1ðx;0Þ ¼ u2ðx;0Þ ¼ 0; 0 < x < l2; ð4:3Þu1ð0; tÞ ¼ f ðtÞ; t > 0; ð4:4Þ@u2
@xðl2; tÞ ¼ 0; t > 0; ð4:5Þ
u1ðl1; tÞ ¼ u2ðl1; tÞ; t > 0; ð4:6Þ
k1@u1
@xðl1; tÞ ¼ k2
@u2
@xðl1; tÞ; t > 0: ð4:7Þ
The solution of (4.1)–(4.7) in the frequency domain is given by
u2ðl2; nÞ :¼ gðnÞ ¼ 1bAð0; nÞ f ðnÞ: ð4:8Þ
Inverse Fourier transform gives the data gðtÞ. Then we generate the noisy data gd:
gd ¼ g þ gmax � r randðsizeðgÞÞ; ð4:9Þ
5490 X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494
r indicates the error level, gmax is the maximum value of sampled data g, RMS denotes the root mean square for a sampledfunction W which is defined by
RMSðWÞ ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
N þ 1
XNþ1
j¼1
ðWðtjÞÞ2vuut ; ð4:10Þ
where N þ 1 is the total number of test points. According to (4.10) we can define the root mean square error (RMSE) for eachpair of the computed data and exact data. The symbol randðsizeð�ÞÞ is a random number uniformly distributed on ½0;1�.
In the numerical test, we take a1 ¼ 9; k1 ¼ 6; a2 ¼ 4; k2 ¼ 3; l1 ¼ 0:5; l2 ¼ 1; x ¼ 0; N ¼ 100. The numerical simulationis produced by the following steps:
� Step 1. Use discrete FFT to obtain the data in the frequency domain according (4.8), then perform inverse FFT to obtain thedata gðtÞ.� Step 2. According to (4.9), generate the noisy data gdðtÞ.� Step 3. Use discrete FFT to obtain the regularized solution with the noisy data in the frequency domain, according to (3.1)
and (3.2), then perform inverse FFT to obtain the data ud1;að0; tÞ and
@ud1;a@x ð0; tÞ.
Example 1. Let the exact data for the corresponding direct problem be given by
f ðtÞ ¼1; if 0:1 6 t 6 0:4 and 0:5 6 t 6 0:9;0; else:
�
This function belongs to HpðRÞ with p < 1=2.First, we can computed the exact input data gðtÞ according to (4.8). The result is shown in Fig. 2.To show the effect of the proposed method, we compute the solution f ðtÞ :¼ u1ð0; tÞ according to the formula (2.20)
directly without any regularization. The result is displayed in Fig. 3 where r ¼ 1%.Now in order to show the convergence of the proposed regularization method, we computed the regularization solutions
by tuning the regularization parameter nmax gradually. The results are shown in Fig. 4 where r ¼ 1%.From Fig. 4, we can see that the important role of regularization parameter nmax is obvious for accurate
approximation. If the parameter nmax is too large or too small, the numerical approximation to the solution willcompletely fail.
We give two computation results to close this example.Fig. 5 shows the result for the reconstruction of temperature with r ¼ 1% and nmax ¼ 80.Fig. 6 shows the result for the reconstruction of flux with r ¼ 1% and nmax ¼ 200.Despite the discontinuities of the temperature in the form of square-wave step function and the heat flux in the form of
sharp corner, the numerical approximations are in reasonably match with the exact solutions of the temperature and theheat flux.
From the above figures, we can see that the proposed method works well too.
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
the
inpu
t dat
a g(
t)
Fig. 2. The exact input data gðtÞ.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
the
solu
tion
u 1(0,t)
and
its
appr
oxim
atio
n
exact solutionno regularization
Fig. 3. The result without any regularization.
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
the s
olu
tion u
1(0
,t)
and it
s appro
xim
atio
n
exact solutionFourier regularization
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
the s
olu
tion u
1(0
,t)
and it
s appro
xim
atio
n
exact solutionFourier regularization
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
the s
olu
tion u
1(0
,t)
and it
s appro
xim
atio
n
exact solutionFourier regularization
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
t
the
solu
tion
u 1(0,t)
and
its
appr
oxim
atio
n
exact solutionFourier regularization
Fig. 4. (a): nmax ¼ 10 , (b): nmax ¼ 40, (c): nmax ¼ 70, (d) nmax ¼ 300.
X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494 5491
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
the
solu
tion
u 1(0,t)
and
its
appr
oxim
atio
n
exact solutionFourier regularization
Fig. 5. Reconstruction of temperature with RMSE = 0.13.
0 0.2 0.4 0.6 0.8 1−4
−3
−2
−1
0
1
2
3
4
t
the
solu
tion
(u1) x(0
,t) a
nd it
s ap
prox
imat
ion
exact solutionFourier regularization
Fig. 6. Reconstruction of gradient of temperature with RMSE = 0.40.
0 0.2 0.4 0.6 0.8 1−0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
t
the
inpu
t dat
a g(
t)
Fig. 7. the exact input data gðtÞ.
5492 X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
the
solu
tion
u 1(0,t)
and
its
appr
oxim
atio
n
exact solutionFourier regularization
Fig. 8. Reconstruction of temperature with RMSE = 0.02.
0 0.2 0.4 0.6 0.8 1−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
t
the
solu
tion
(u1) x(0
,t) a
nd it
s ap
prox
imat
ion
exact solutionFourier regularization
Fig. 9. Reconstruction of flux (gradient) of temperature of with RMSE = 0.04.
X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494 5493
Example 2. Let the exact data be given by
f ðtÞ ¼
23 ð5t � 1Þ; if 0:2 6 t 6 0:5;� 2
3 ð5t � 4Þ; if 0:5 6 t 6 0:8;0; else:
8><>:
This function belongs to HpðRÞ with p < 3=2.Without more explanation, we give the numerical results as follows:Fig. 7 shows the exact input data gðtÞ for the inverse problem.Fig. 8 shows the result for the reconstruction of temperature with r ¼ 3% and nmax ¼ 80.Fig. 9 shows the result for the flux(gradient) of temperature with r ¼ 3% and nmax ¼ 80.From these numerical results we conclude that the proposed method works well. In these examples we managed to solve
the inverse problem with acceptable accuracy. In the numerical tests, we found that the quality of the numerical solution isnot very sensitive to variations of the cutoff level within a suitable range. It is relatively easy to find an appropriate regular-ization parameter nmax, which is very important for practical computation.
5. Concluding Remarks
Although there are several regularization methods for stabilizing the inverse heat conduction problem in a single-layerbody by using an a priori information on the exact solution, the regularization error estimates for the inverse heat conduc-tion problem in a multi-layer body are still very rare. This is due to the complexity of the forward operators as shown in(2.23) and (2.24). Therefore, the direct extension of the existing methods for solving the IHCP in single-layer domain is
5494 X.T. Xiong et al. / Applied Mathematical Modelling 39 (2015) 5480–5494
unavailable. However, the idea for stabilizing the IHCP in single-layer domain can be used. In this paper, we found that thespectral regularization method is efficient in solving the IHCP in two layer domain.
Spectral method is a simple and effective method and has been investigated by many authors for IHCP, e.g.[15][35]. Themethod can be implemented by the mollification method with some special mollifier in the physical domain (see [35]).Relative to the IHCP in single-layer domain, the IHCP in multi-layer domain is much more difficult. As we see that thecorresponding estimates on the forward operators are more difficult to be established. Fortunately, we can successfullyestimate the forward operators in the frequency domain(see Proposition 2.1 of this paper). Furthermore, we obtain the errorestimates for the spectral regularization method for solving the IHCP in two-layer domain. In theoretical aspect, the order ofthe error estimates is ’optimal’. The constructed numerical examples also verify that the proposed regularization method iseffective for solving the IHCP in the two-layer domain.
Acknowledgement
The work described in this paper was partially supported by a grant from the Research Council of the Hong Kong SpecialAdministrative Region, China (Project No. CityU 101112). The research of X.T. Xiong was partially supported by a grant fromthe National Natural Science Foundation of China (No. 11001223), the Research Fund for the Doctoral Program of HigherEducation of China (No. 20106203120001). The authors would like to thank the reviewers for their valuable commentsand suggestions, in particularly the reminder of the concept of ‘‘inverse crime’’.
References
[1] T. Wei, M. Yamamoto, Reconstruction of a moving boundary from Cauchy data in one-dimensional heat equation, Inverse Prob. Sci. Eng. 17 (2009)551–567.
[2] F. Berntsson, An inverse heat conduction problem and improving shielded thermocouple accuracy, Numer. Heat Transfer Part A 61 (2012) 754–763.[3] J.V. Beck, Nonlinear estimation applied to the nonlinear inverse heat conduction problem, Int. J. Heat Mass Transfer 13 (1970) 703–716.[4] B.R. Bass, L.J. Ott, A finite element formulation of the two-dimensional nonlinear inverse heat conduction problem, Adv. Comput. Technol. 2 (1980)
238–248.[5] J. Baumeister, H.J. Reinhardt, On the approximate solution of a two-dimensional inverse heat conduction problem, in: H.G. Engl, C.W. Groetsch (Eds.),
Inverse and Ill-posed Problems, Academic Press, Orlando, FL, 1987. 325–44.[6] F. Berntsson, A spectral method for solving the sideways heat equation, Inverse Prob. 15 (1999) 891–906.[7] A.S. Carasso, Determining surface temperature from interior observations, SIAM J. Appl. Math. 42 (1982) 558–574.[8] J. Hadamard, Lectures on Cauchy Problem in Linear Partial Differential Equations, Yale University Press, New Haven, CT, 1923.[9] A.N. Tikhonov, V.Y. Arsenin, Solution of Ill-Posed Problems, Winston/Wiley, New York, 1977.
[10] H.W. Engl, M. Hanke, A. Neubauer, Regularization of Inverse Problems, Kluwer Academic Publisher, Dordrecht Boston London, 1996.[11] L. Eldén, Approximations for a Cauchy problem for the heat equation, Inverse Prob. 3 (1987) 263–273.[12] L. Eldén, Hyperbolic approximations for a Cauchy problem for the heat equation, Inverse Prob. 4 (1988) 59–71.[13] L. Eldén, Numerical solution of the sideways heat equation by difference approximation in time, Inverse Prob. 11 (1995) 913–923.[14] L. Eldén, F. Berntsson, T. Reginska, Wavelet and Fourier methods for solving the sideways heat equation, SIAM J. Sci. Comput. 21 (6) (2000) 2187–2205.[15] C.L. Fu, Simplified Tikhonov and Fourier regularization methods on a general sideways parabolic equation, J. Comput. Appl. Math. 167 (2004) 449–463.[16] L. Guo, D.A. Murio, C. Roth, A mollified space marching finite algorithm for the inverse heat conduction problem with slab symmetry, Comput. Math.
Appl. 19 (1990) 75–89.[17] L. Guo, D.A. Murio, A mollified space-marching finite-difference algorithm for the two-dimensional inverse heat conduction problem with slab
symmetry, Inverse Prob. 7 (1991) 247–259.[18] D.N. Hào, A mollification method for ill-posed problems, Numer. Math. 68 (1994) 469–506.[19] M. Imber, Temperature extrapolation mechanism for two-dimensional heat flow, AIAA J. 12 (1974) 1089–1093.[20] L. Liu, D.A. Murio, Numerical experiments in 2-D IHCP on bounded domains part I: the interior cube problem, Comput. Math. Appl. 31 (1996) 15–32.[21] C.F. Weber, Analysis and solution of the ill-posed inverse conduction problem, Int. J. Heat Mass Transfer 24 (1981) 1783–1792.[22] T. Yoshimura, K. Ikuta, Inverse heat conduction problem by finite element formulation, Int. J. Syst. Sci. 16 (1985) 1365–1376.[23] N. Zabaras, J.C. Liu, An analysis of two-dimensional linear inverse heat transfer problems using an integral method, Numer. Heat Transfer 13 (1988)
527–533.[24] D.A. Murio, Stable numerical solution of a fractional-diffusion inverse heat conduction problem, Comput. Math. Appl. 53 (2007) 1492–1501.[25] Y.C. Hon, T. Wei, The method of fundamental solutions for solving multidimensional inverse heat conduction problems, Comput. Model. Eng. Sci. 7 (2)
(2005) 119–132.[26] D.N. Hào, Methods for Inverse Heat Conduction Problems, Peter Lang, Frankfurt am Main, 1998.[27] X.T. Xiong, Regularization theory and algorithm for some inverse problems for parabolic differential equations (Ph.D. dissertation), Lanzhou University,
2007 (in Chinese).[28] T. Wei, Y.S. Li, An inverse boundary problem for one-dimensional heat equation with a multilayer domain, Eng. Anal. Bound. Elem. 33 (2009) 225–232.[29] T. Wei, Uniqueness of moving boundary for a heat conduction problem with nonlinear interface conditions, Appl. Math. Lett. 23 (2010) 600–604.[30] Y.S. Li, T. Wei, identification of a moving boundary for a heat conduction problem in a multilayer medium, Heat Mass transfer 46 (2010) 779–789.[31] X.T. Xiong, Y.C. Hon, Regularization error analysis on a one-dimensional inverse heat conduction problem in multi-layer domain, Inverse Prob. Sci. Eng.
21 (2013) 865–887.[32] A. Yu. Shcheglov, A method for the approximate solution of an inverse problem for the heat equation, Zh. Vychisl. Mat. iMat. Fiz. 32 (1992) 904–916
(Russian) Translation in Comput. Math. Math. Phys., 32 (1992) 781-791..[33] A. Yu. Shcheglov, Uniform approximation of a solution of an inverse problem by the quasir eversibility method, Mat. Zametki 53 (1993) 168–174
(Russian) Translation in Math. Notes 53 (1993) 235-239.[34] C. Aviles-Ramos, A. Haji-Sheikh, J.V. Beck, Exact solution of heat conduction in composites and application to inverse problems, ASME J. Heat Transfer
120 (1998) 592–599.[35] C. Aviles-Ramos, A. Haji-Sheikh, Estimation of thermophysical properties of composites using multi-parameter estimation and zeroth-order
regularization, Inv. Prob. Eng. 9 (2001) 507–536.[36] J.P. Kaipio, E. Somersalo, Statistical and Computational Inverse Problems, Applied Mathematical Sciences, Springer Publisher, 2005.[37] D.N. Hào, H.J. Reinhardt, On a sideways parabolic equation, Inverse Prob. 13 (1997) 297–309.[38] R.I. Hicksona, S.I. Barrya, G.N. Mercera, H.S. Sidhua, Finite difference schemes for multilayer diffusion, Math. Comput. Modell. 54 (2011) 210–220.