数学杂志  2019, Vol. 39 Issue (3): 464-474   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
闵涛
仝云莉
逆散射问题求解的正则化GAUSS-NEWTON法
闵涛, 仝云莉    
西安理工大学理学院应用数学系, 陕西 西安 710054
摘要:本文研究了一类重要的反问题–逆散射问题,它是从远场散射数据识别散射体中的障碍物的形状问题.利用迭代正则化Gauss-Newton法,通过数值模拟比较分析,验证了本文所提出的方法在求解逆散射问题时是可行有效的.
关键词逆散射    噪声    正则化    Gauss-Newton法    数值解    
A REGULARIZED GAUSS-NEWTON METHOD FOR SOLVING INVERSE SCATTERING PROBLEM
MIN Tao, TONG Yun-li    
Department of Applied Mathematics, Xi'an University of Technology, Xi'an 710054, China
Abstract: In this paper, an important inverse problem-inverse scattering problem, is studied. It is a problem of identifying the shape of an obstacle in a scatterer from far-field scattering data. Here, we consider the problem using the Iteratively Regularized Gauss-Newton method. Through the numerical simulation, it is proved that the proposed method is feasible and effective in solving the inverse scattering problem.
Keywords: inverse scattering     noise     regularization     Gauss-Newton method     numerical solution    
1 引言

逆散射问题在遥感、无损探测、地球物理、医用成像和雷达目标识别等方面[1-4]有着广泛的应用.但由于此类问题具有非线性、不适定的特点, 给数值求解带来了很大困难. 20世纪60年代, 前苏联数学家吉洪诺夫、伊万诺夫和拉弗伦捷耶夫等提出了不适定问题正则化方法, 随后, 很多学者对逆散射问题进行了卓有成效的研究. Jost等人[5]提出了逆散射微扰论, Moses [6]和Prosser [7-10]等人对此作了进一步的研究, 研究了一维和三维逆散射问题, 对解误差进行了分析和估计, 随后Newton等人[7-8, 11-14]对一维和三维逆散射问题进行了进一步的研究, 并对解的唯一性进行了讨论. Colton等人[15-18]利用积分方程方法对声波和电磁波逆问题作了深入研究.其中, 对逆散射理论发展贡献最大的是Bleistein [19-26]和Cohen等人[27-29], 他们以小扰动理论和Born近似为理论基础, 建立了利用Fourier变换等方法进行反演的理论, 最终形成了逆散射反演方法.本文在对障碍物的边界进行参数化处理的基础上, 将逆散射问题转化为一非线性积分方程, 通过数值积分离散, 提出利用迭代正则化高斯-牛顿法来求解, 并进行了数值模拟.

2 逆散射模型的建立

散射是指由光波、音波、电磁波或粒子在通过均匀介质时, 遇到区域大小为$ \Omega $、边界为$ \Gamma $的障碍物而改变其直线运动轨迹的物理现象.为了计算方便, 我们主要考虑二维区域并假定散射区域$ \Omega $是呈星形的, 即在散射区域$ \Omega $中, 存在连续的正函数$ R(\theta)\; (0\leq\theta<2\pi) $, 使得边界$ \Gamma $上的点用极坐标表示为$ (R(\theta), \theta) $.假定入射波为波矢为$ k_{0} $的简谐波, 即$ \psi^{in}(r) = \exp(ik_{0}\cdot r) $, 障碍物中的波数$ k_{1} $是一个常量.令$ \psi^{sc}(r) $表示散射区域, 总区域$ \psi = \psi^{in}+\psi^{sc} $满足微分方程

$ (\bigtriangledown^{2}+k_{1}^{2})\psi(r) = 0, r\in\Omega, $
$ \begin{align} (\bigtriangledown^{2}+k_{0}^{2})\psi(r) = 0, r\in R^2-\bar{\Omega}, \end{align} $ (1)

其中$ \psi $和导数$ \frac{\partial\psi}{\partial n} $在区域$ \Omega $的边界$ \Gamma $上连续.

如果平面波正常照射介质柱的轴线, 并且电场$ E $的两极沿轴线分布在两端, 该散射问题转化为求解方程(1), 其中$ \psi = E $, 波数$ k_{0}^{2} = \varepsilon_{0}\mu_{0}\omega^{2} $, $ k_{1}^{2} = \varepsilon_{0}\varepsilon_{r}\mu_{0}\omega^{2}+i\omega\mu_{0}\sigma $, $ \varepsilon_{0} $, $ \mu_{0} $是真空中的介电常数和磁导率, $ \varepsilon_{r} $, $ \sigma $是障碍物的相对介电常数和电导率.

从入射角$ \varphi $处的远场散射数据, 得到散射振幅为

$ \begin{align} S(\varphi) = \lim\limits_{x\to\infty}\psi^{sc}(r)(\frac{1}{2}\pi k_{0}r)^{\frac{1}{2}}\exp(-ik_{0}r+\frac{1}{4}i\pi), \end{align} $ (2)

其中向量$ r $用极坐标表示为$ (r, \varphi) $.对于$ \frac{k_{1}}{k_{0}}\approx1 $, 障碍为弱散射, 则

$ \begin{align} S(\varphi) = S_{B}(\varphi)+\delta_{B}(\varphi). \end{align} $ (3)

这种情况下

$ \begin{align} S_{B}(\varphi) = \frac{1}{2}ik_{0}^{2}(\frac{k_{1}}{k_{0}}-1)\int_{\Omega}\exp(-i2k_{0}\cdot r)dr \end{align} $ (4)

表示散射振幅和$ \delta_{B} $的Born近似[30-32].与$ S_{B} $相比, Born近似导致的误差是很小的.

为了求方程(4)的导数近似, 首先利用方程(1)和辐射边界条件$ (\nabla\psi-k_{0}\psi = O(r^{-\frac{1}{2}})(r\rightarrow\infty)) $下区域$ \psi $中边界$ \Gamma $的光滑性, 得到格林函数表示为

$ \begin{eqnarray*} &&\psi^{sc}(r) = \frac{1}{4}i(k_{1}^{2}-k_{0}^{2})\int_{\Omega}H_{0}^{(1)}(k_{0}R)\psi(r')dr'\\ &&R = [r^{2}-2r\cdot r+(r')^{2}]^{\frac{1}{2}}, \end{eqnarray*} $

其中$ H_{0}^{(1)} $是第一类Hankel函数.用$ \psi(r) $代替$ \exp(ik_{0}\cdot r) $, 当$ r\rightarrow\infty $时, $ R\rightarrow r-\widehat{r}\cdot r $ ($ \widehat{r} $是反向散射方向$ (-k_{0}) $的单位矢量).由于$ k_{1}^{2}-k_{0}^{2}\approx2k_{0}^{2}(\frac{k_{1}}{k_{0}}-1) $, 因此

$ H(k_{0}R)\rightarrow(\frac{2}{\pi k_{0}R})^{\frac{1}{2}}\exp(ik_{0}R-\frac{1}{4}i\pi)\approx(\frac{2}{\pi k_{0}r}) ^{\frac{1}{2}}\exp(ik_{0}r-\frac{1}{4}i\pi)\exp(ik\cdot r'). $

对于模型的离散测量数据, 考虑

$ \begin{align} d_{i} = S_{B}(\varphi_{i})+\delta_{B}(\varphi_{i})+\varepsilon_{i}, 1\leq i\leq N, \end{align} $ (5)

其中$ \varepsilon_{i} $表示噪声.假设$ \varepsilon_{i} $是随机变量, 满足

$ \begin{align} E(\varepsilon_{i}) = 0, \; \; E(\varepsilon_{i}\varepsilon_{j}) = \sigma^{2}\delta_{ij}, 1\leq i, j\leq N, \end{align} $ (6)

通过$ \frac{1}{2}ik_{0}^{2}(\frac{k_{1}}{k_{0}}-1) $, 方程(5)可写为

$ \begin{align} d_{j} = \int\int I_{\Omega}(r)\exp(-iq_{j}\cdot r)dr+\eta_{j}, 1\leq j\leq N, \end{align} $ (7)

其中如果$ r\in\Omega $ (0除外), 则$ I_{\Omega}(r) = 1 $. $ q $用极坐标表示为$ (2k_{0}, \varphi) $, Born近似所导致的误差包含在$ \eta_{j} $中.

为了解决反问题的数据缺失和不适定性, 使问题更容易处理, 对问题进行降维处理.假设散射区域$ \Omega $呈星形, 方程(7)转化为极坐标表示

$ \begin{align} d_{i} = \int_{0}^{2\pi}f(\theta, \varphi_{i}, R(\theta))d\theta+\eta_{i}, 1\leq i\leq N, \end{align} $ (8)

其中

$ f(\theta, \varphi, R) = \frac{[\exp(\beta R)(\beta R-1)+1]}{\beta^{2}}, \; \; \beta = -i2k_{0}\cos(\varphi-\theta). $

方程(8)可以转化为一维非线性第一类Fredholm积分方程

$ \begin{align} d(\varphi) = \int_{0}^{2\pi}f(\theta, \varphi, R(\theta))d\theta. \end{align} $ (9)

一般来说, 该方程的解$ R(\theta) $不连续依赖于数据$ d(\varphi) $.

3 迭代正则化Gauss-Newton法

方程(9)可以归结为求解非线性方程$ F(\gamma) = y $, 这里

$ \begin{align} [F(\gamma)](\phi) = \int_{0}^{2\pi}f(\psi, \phi, \gamma(\psi))d\psi = y(\phi), \end{align} $ (10)

其中

$ f(\psi, \phi, \gamma) = \frac{[\exp(\beta\gamma)(\beta\gamma-1)+1]}{\beta^{2}}, \; \; \beta = -i2k_{0}\cos(\phi-\psi), $

$ k_{0} $是固定波数, $ \gamma(\psi) $未知[1, 2, 33].

在实际问题中, 右端数据$ y $通常是由测量得到的, 因而得到的数据是一个满足$ \|y^{\delta}-y\|\leq\delta $的近似数据$ y^{\delta} $, 这里$ \delta>0 $是给定的较小的扰动水平[34-35].

求解不适定问题的方法主要有两种: Tikhonov正则化和迭代法, 本文主要采用迭代正则化Gauss-Newton法求解逆散射问题.

定义泛函

$ \begin{align} \min\limits_{\gamma\in X}J_{\alpha}[\gamma, y] = \|F(\gamma)-y^{\delta}\|^{2}+\alpha\Omega(\gamma), \end{align} $ (11)

其中$ \alpha>0 $为正则化参数, $ \Omega(\gamma) $为稳定泛函.要求无约束最优化问题(11)的解, 首先将算子$ F $线性化, 利用$ F(\gamma) $在第$ k $次迭代点$ \gamma_{k} $处的泰勒展开式, 得到

$ \begin{align} \min\limits_{\gamma\in X}J_{\alpha}[\gamma, y^{\delta}] = \|y^{\delta}-F(\gamma_{k})-F'(\gamma_{k})(\gamma-\gamma_{k})\|^{2}+\alpha\Omega(\gamma), \end{align} $ (12)

其中$ \Omega(\gamma) = \|L(\gamma-\gamma_{0})\| $为稳定泛函, $ L $是单位矩阵$ (L = L_{0} = I\in R^{n\times n}) $或者是一阶或二阶导算子的离散近似, 即

$ \begin{eqnarray*} &&[L_{1}]_{ij} = \delta_{i, j}-\delta_{i, j-1}, \; \; i = 1, 2, \cdots, n-1, \\ &&[L_{2}]_{ij} = \delta_{i, j}-2\delta_{i, j-1}+\delta_{i, j-2}, \; \; i = 1, 2, \cdots, n-2, \end{eqnarray*} $

其中$ \delta_{ij} $为克罗内克符号, $ j = 1, 2, \cdots, n $.方程(12)通过一阶最优条件求解, 可得到

$ \begin{align} \gamma_{k+1} = \gamma_{k}-(\alpha_{k}L^{T}L+F'(\gamma_{k})^{T}F'(\gamma_{k}))^{-1}[F'(\gamma_{k})^{T}(F(\gamma_{k})-y^{\delta})+\alpha_{k}L^{T}L(\gamma_{k}-\gamma_{0})], \end{align} $ (13)

(13) 式称为迭代正则化Gauss-Newton法[36-37], 简记为IRGN法.

要利用此方法数值求解逆散射问题(10), 需要将其离散化, 本文采用梯形公式将其离散.将公式(13)改写为如下形式

$ [F'(\gamma_{k})^{T}F'(\gamma_{k})+\alpha_{k}L^{T}L]h_{k} = -[F'(\gamma_{k})^{T}(F(\gamma_{k})-y^{\delta})+\alpha_{k}L^{T}L(\gamma_{k}-\gamma_{0})], $

其中$ h_{k} = \gamma_{k+1}-\gamma_{k} $, 要得到$ F'(\gamma_{k})^{T}F'(\gamma_{k}) $, 需要求解非线性算子$ F $的Fréchet导数.

$ f(\psi, \phi, \gamma(\psi)) $为定义在$ 0\leq\psi, \phi\leq2\pi $, $ \|\gamma\|\leq r $上的函数, 则可知$ f(\psi, \phi, \gamma(\psi)) $, $ f'_{\gamma}(\psi, \phi, \gamma(\psi)) $处处连续.积分算子$ F $$ [0, 2\pi] $上的连续函数, 并且在开球$ \|\gamma\|\leq r $上Fréchet可导, 对任意的$ \gamma_{0}\in C[0, 2\pi] $$ \|\gamma_{0}\|\leq r $, $ h\in C[0, 2\pi] $, 可得到$ F $的Fréchet导数为

$ \begin{align} (F'(\gamma_{0})h)\phi = \int_{0}^{2\pi}f'_{\gamma}(\psi, \phi, \gamma_{0}(\psi))h(\psi)d\psi. \end{align} $ (14)

上述(14)式定义的算子是$ [0, 2\pi] $上的线性连续算子.

事实上, 对于确定的$ \theta = \theta(h)\in[0, 1] $, 有

$ \begin{eqnarray*} &&\|F(\gamma_{0}+h)-F(\gamma_{0})-F'(\gamma_{0})h\|\\ &\leq&\int_{0}^{2\pi}\sup\limits_{0\leq\phi\leq2\pi}|f(\psi, \phi, \gamma_{0}(\psi)+h(\psi))-f(\psi, \phi, \gamma_{0}(\psi))-f'_{\gamma}(\psi, \phi, \gamma_{0}(\psi))h(\psi)|d\psi\\ &\leq&\|h\|\int_{0}^{2\pi}\sup\limits_{0\leq\phi\leq2\pi}|f'_{\gamma}(\psi, \phi, \gamma_{0}(\psi)+\theta h(\psi))-f'_{\gamma}(\psi, \phi, \gamma_{0}(\psi))|d\psi. \end{eqnarray*} $

由于函数$ f'_{\gamma}(\psi, \phi, \gamma(\psi)) $$ \gamma $处连续, 有

$ \begin{eqnarray*} &&\lim\limits_{\|h\|\rightarrow0}\frac{1}{\|h\|}\|F(\gamma_{0}+\phi h)-F(\gamma_{0})-F'(\gamma_{0})h\|\\ &\leq&\lim\limits_{\|h\|\rightarrow0}\int_{0}^{2\pi}\sup\limits_{0\leq\phi\leq2\pi}|f'_{\gamma}(\psi, \phi, \gamma_{0}(\psi)+\theta h(\psi))-f'_{\gamma}(\psi, \phi, \gamma_{0}(\psi))|d\psi = 0. \end{eqnarray*} $

因此$ F $的Fréchet导数可以通过(14)式得到.

利用梯形公式将方程(14)进行离散:把区间$ [0, 2\pi] $等分成$ N $个小区间, 其步长为$ l = \triangle s = \frac{2\pi}{N} $, 令$ \psi_{0} = 0 $, $ \psi_{j} = j\triangle s $.同理$ \psi_{0} = \phi_{0} = 0 $, $ \psi_{N} = \phi_{N} = 2\pi $, $ \phi_{j} = j\triangle s(\psi_{j} = \phi_{j}). $$ \gamma(\psi_{j}) = \gamma_{j} $, 则有

$ \begin{eqnarray*} &&(F'(\gamma)h)\phi = \int_{0}^{2\pi}f'_{\gamma}(\psi, \phi, \gamma(\psi))h(\psi)d\psi\\ &\approx&\triangle s[f'_{\gamma}(\psi_{0}, \phi_{i}, \gamma_{0})h_{0}+f'_{\gamma}(\psi_{1}, \phi_{i}, \gamma_{1})h_{1}+\cdots+f'_{\gamma}(\psi_{N-1}, \phi_{i}, \gamma_{N-1})h_{N-1}+\frac{1}{2}f'_{\gamma}(\psi_{N}, \phi_{i}, \gamma_{N})h_{N}]\\ &\approx&\sum\limits_{j = 0}^{N}f'_{\gamma}(\psi_{j}, \phi_{i}, \gamma_{j})h_{j}\omega_{j}, \quad i = 0, 1, 2, \cdots, N, \end{eqnarray*} $

改写成矩阵形式为

$ F'_{\gamma}h = \left( \begin{matrix} b_{01}&b_{02}&\cdots&b_{0N}\\ b_{11}&b_{12}&\cdots&b_{1N}\\ \vdots&\vdots&\ddots&\vdots\\ b_{N1}&b_{N2}&\cdots&b_{NN} \end{matrix} \right)\left(\begin{matrix}h_{0}\\h_{1}\\\vdots\\h_{N}\end{matrix} \right), $

其中

$ \left( \begin{matrix} b_{01}&b_{02}&\cdots&b_{0N}\\ b_{11}&b_{12}&\cdots&b_{1N}\\ \vdots&\vdots&\ddots&\vdots\\ b_{N1}&b_{N2}&\cdots&b_{NN} \end{matrix} \right) = \triangle s\left( \begin{matrix} \frac{1}{2}f'_{\gamma}(\psi_{0}, \phi_{0}, \gamma_{0})&f'_{\gamma}(\psi_{1}, \phi_{0}, \gamma_{1})&\cdots&\frac{1}{2}f'_{\gamma}(\psi_{N}, \phi_{0}, \gamma_{N})\\ \frac{1}{2}f'_{\gamma}(\psi_{0}, \phi_{1}, \gamma_{0})&f'_{\gamma}(\psi_{1}, \phi_{1}, \gamma_{1})&\cdots&\frac{1}{2}f'_{\gamma}(\psi_{N}, \phi_{1}, \gamma_{N})\\ \vdots&\vdots&\ddots&\vdots\\ \frac{1}{2}f'_{\gamma}(\psi_{0}, \phi_{N}, \gamma_{0})&f'_{\gamma}(\psi_{1}, \phi_{N}, \gamma_{1})&\cdots&\frac{1}{2}f'_{\gamma}(\psi_{N}, \phi_{N}, \gamma_{N}) \end{matrix} \right). $

对于每一步迭代$ k $, 利用迭代正则化Gauss-Newton法可得到

$ \begin{align} \left[\left( \begin{matrix} b_{01}^{k}&b_{11}^{k}&\cdots&b_{N1}^{k}\\ b_{02}^{k}&b_{12}^{k}&\cdots&b_{N2}^{k}\\ \vdots&\vdots&\ddots&\vdots\\ b_{0N}^{k}&b_{1N}^{k}&\cdots&b_{NN}^{k} \end{matrix} \right)\left( \begin{matrix} b_{01}^{k}&b_{02}^{k}&\cdots&b_{0N}^{k}\\ b_{11}^{k}&b_{12}^{k}&\cdots&b_{1N}^{k}\\ \vdots&\vdots&\ddots&\vdots\\ b_{N1}^{k}&b_{N2}^{k}&\cdots&b_{NN}^{k} \end{matrix} \right)+\alpha_{k}L^{T}L\right]\left( \begin{matrix} h_{0}^{k}\\h_{1}^{k}\\\vdots\\h_{N}^{k} \end{matrix} \right) = g, \end{align} $ (15)

其中

$ \begin{eqnarray*} &&g = -\left( \begin{matrix} b_{01}^{k}&b_{11}^{k}&\cdots&b_{N1}^{k}\\ b_{02}^{k}&b_{12}^{k}&\cdots&b_{N2}^{k}\\ \vdots&\vdots&\ddots&\vdots\\ b_{0N}^{k}&b_{1N}^{k}&\cdots&b_{NN}^{k} \end{matrix} \right)\left( \begin{matrix} \sum\limits_{j = 0}^{N}f(\psi_{j}, \phi_{0}, \gamma_{j}^{k})\omega_{j}-y^{\delta}(\phi_{0})\\ \sum\limits_{j = 0}^{N}f(\psi_{j}, \phi_{1}, \gamma_{j}^{k})\omega_{j}-y^{\delta}(\phi_{1})\\ \vdots\\ \sum\limits_{j = 0}^{N}f(\psi_{j}, \phi_{N}, \gamma_{j}^{k})\omega_{j}-y^{\delta}(\phi_{N}) \end{matrix} \right)-\alpha_{k}L^{T}L\left( \begin{matrix} \gamma_{0}^{k}-\gamma_{0}^{0}\\ \gamma_{1}^{k}-\gamma_{1}^{0}\\ \vdots\\ \gamma_{N}^{k}-\gamma_{N}^{0} \end{matrix} \right), \\ &&F(\gamma_i^{(k)}) = \sum\limits_{j = 0}^{N}f(\psi_j, \phi_i, \gamma_j^{(k)})w_j, \quad i = 0, 1, 2, \cdots, N, \\ &&\gamma^{(0)} = (\gamma_0^{(0)}, \gamma_1^{(0)}, \cdots, \gamma_N^{(0)}). \end{eqnarray*} $

因此得到迭代格式$ \gamma^{(k+1)} = \gamma^{(k)}+h^{(k)} $, 通过此格式便可求出方程(10)的近似解.

4 数值模拟

考虑方程(10)所示的逆散射问题, 在数值求解之前, 先做如下规定

(1)  用$ L_\infty $表示误差的$ \infty $ -范数, 即$ L_\infty = \underset{0\leqslant j\leqslant N}\max\left|\gamma(\psi_j)-\overset {\thicksim}\gamma(\psi_j)\right| $.

(2)  用$ RE $表示相对误差, 即

$ RE = \sqrt{\sum\limits_{j = 0}^{N}\left|\gamma(\psi_j)-\overset {\thicksim}\gamma(\psi_j)\right|^2}/\sqrt{\sum\limits_{j = 0}^{N}\left|\gamma(\psi_j)\right|^2}, $

其中$ \psi_j $为节点, $ N $为区间$ [0, 2\pi] $上均匀分布的节点的个数, $ \gamma(\psi) $为精确解, $ \overset {\thicksim}\gamma(\psi) $为数值解.

(3)  在数值计算时, 假定扰动$ \{y^{\delta}(\phi_j)\}_{j = 0}^N $是随机的, 但是在实际应用时, 反复测试是相当困难甚至不可能的.因此, 考虑确定性误差.假设观测到的数据有以下扰动形式

$ y^{\delta}(\phi_j) = y(\phi_j)+\delta\sin(\sqrt{10}\pi\phi_j), j = 0, 1, \cdots, N. $

(4)  基于Sigmoid -型函数的性质, 选取正则化参数为

$ \alpha_{k} = \alpha(k) = \frac{1}{1+\exp(k)}, \quad k\text{为迭代次数}. $

不难验证其满足下列条件

(ⅰ)  $ 1<\frac{\alpha_k}{\alpha_{k+1}}<3, $

(ⅱ)  $ \underset{k\rightarrow\infty}\lim\alpha_k = 0. $

可以看到, 根据上述正则化参数的选取方法, 在迭代开始时能够充分对问题进行正则化, 然后随着迭代数的增加正则化参数逐渐减小, 达到解稳定的目的.

数值模拟时, 对所要识别的障碍物边界$ \gamma(\psi) $, 给定两种参数化模型的精确解, 分别对所得数值解和精确解进行比较, 以验证本文所提方法的有效性.

数值模拟一  考虑真解

$ \gamma(\psi) = (1.2+0.25\cos(3\psi)), \; \; 0\leqslant\psi\leqslant2\pi. $

数值模拟时, 取$ N = 100, L = L_1, k_0 = 2 $, 迭代次数$ k = 30 $, 初始猜测$ \gamma^{(0)} = (1, 1, \cdots, 1)^T $, 扰动分别为$ \delta = 0, 0.01, 0.05 $时, 所得$ L_{\infty} $$ RE $分别如表 1所示, 精确解与数值解的比较如图 1所示.

表 1 不同扰动、相同迭代数下, 迭代正则化Gauss-Newton法所得误差比较

图 1 取不同的扰动水平$ \delta $时, 所得精确解与数值解的比较

$ N = 100, k_0 = 2 $, 迭代次数$ k = 30 $, 初始猜测$ \gamma^{(0)} = (1, 1, \cdots, 1)^T $, $ L $分别取单位矩阵$ I $$ L_1 $$ L_2 $, 扰动分别为$ \delta = 0, 0.01, 0.05 $时, 所得$ L_{\infty} $$ RE $分别如表 2表 3所示, 精确解与数值解的比较如图 2所示.

表 2 不同扰动、相同迭代数下, $L$取不同时所得误差$(L_{\infty})$比较

表 3 不同扰动、相同迭代数下, $L$取不同时所得误差$(RE)$比较

图 2 无扰动时, $ L $取不同的值所得精确解与数值解的比较

数值模拟二  考虑真解

$ \gamma(\psi) = (1.2+0.25\cos(7\psi)), \; \; 0\leqslant\psi\leqslant2\pi. $

参数设置与模拟一相同, 所得$ L_{\infty} $$ RE $分别如下表 4所示, 精确解与数值解的比较如图 3所示.

表 4 不同扰动、相同迭代数下, 迭代正则化Gauss-Newton法所得误差比较

图 3 取不同的扰动水平$ \delta $时, 所得精确解与数值解的比较

$ N = 100, k_{0} = 2, $迭代次数$ k = 30 $, 初始猜测$ \gamma^{(0)} = (1, 1, \cdots, 1)^{T} $, $ L $分别取单位矩阵$ I $$ L_{1} $$ L_{2} $, 扰动分别为$ \delta = 0, 0.01, 0.05 $时, 所得$ L_{\infty} $$ RE $分别如下表 5表 6所示, 精确解与数值解的比较如下图 4所示.

表 5 不同扰动、相同迭代数下, $L$取不同时所得误差$(L_{\infty})$比较

表 6 不同扰动、相同迭代数下, $L$取不同时所得误差$RE$比较

图 4 无扰动时, $ L $取不同的值所得精确解与数值解的比较

比较上述两个数值模拟结果, 可以得到如下结论:从表 1表 4可以看出, 在迭代数相同的情况下, 当右端测量数据无扰动时, 所得解的误差很小, 数值解较准确, 而随着扰动的增加, 数值解与精确解误差的$ \infty $ -范数和相对误差也逐渐增加; 从表 2表 3表 5表 6可以看出, 在迭代数一定的情况下, 取不同的$ L $值时, 随着扰动的增加, 所得数值解与精确解误差的$ \infty $ -范数和相对误差都在非常小的范围内, 特别取$ L = L_{1} $时, 所得数值解与精确解误差的$ \infty $ -范数和相对误差都相对较小.

5 结论

本文采用迭代正则化Gauss-Newton法求解一类根据远场散射数据识别介质中障碍物形状的逆散射问题, 通过数值模拟可以看出:用此方法求解此类问题是可行的、有效的; 但是对于参数化后的边界函数, 在利用正则化方法求解时, 当稳定泛函中的$ L $取不同的算子时, 数值解的稳定性和准确性有微小差别, 而当$ L $取一阶导算子时, 数值求解所得结果更准确.

参考文献
[1] Lee K-M. Inverse scattering via nonlinear integral equations for a Neumann crack[J]. Inver. Prob., 2006, 22(6): 1989–2000. DOI:10.1088/0266-5611/22/6/005
[2] Qin H-H, Cakoni F. Nonlinear integral equations for shape reconstruction in the inverse interior scattering problem[J]. Inver. Prob., 2011, 27(3): 005–021.
[3] Cakoni F, Colton D. Qualitative methods in inverse scattering theory[M]. Springer, 2006.
[4] Cakoni F, Kress R. Integral equations for inverse problems in corrosion detection from partial Cauchy data[J]. Inver. Prob., 2007: 229–245.
[5] Jost R, Walte K. Construction of a potential from a pahase shift[J]. Phys. Rev., 1952, 87(6): 977–991. DOI:10.1103/PhysRev.87.977
[6] Moses H E. Calculation of the scattering potential from reflection coefficients[J]. Phys. Rev., 1955, 102(2): 559–567.
[7] Norton J. Itertive seismic inversion[J]. Geophys J., 1988, 94: 457–468. DOI:10.1111/j.1365-246X.1988.tb02268.x
[8] Prosser R T. Formal solution of inverse scattering problems:Ⅰ[J]. J. Math. Phys., 1969, 10(8): 1819–1822.
[9] Prosser R T. Formal solution of inverse scattering problems:Ⅱ[J]. J. Math. Phys., 1976, 17(11): 1775–1779.
[10] Prosser R T. Formal solution of inverse scattering problems:Ⅲ[J]. J. Math. Phys., 1980, 21(11): 2648–2653. DOI:10.1063/1.524379
[11] Newton R G. Inverse scattering:Ⅰ. One dimension[J]. J. Math. Phys., 1980, 21(3): 493–505.
[12] Newton R G. Inverse scattering:Ⅱ. Three dimensions[J]. J. Math. Phys., 1980, 21(7): 1698–1715. DOI:10.1063/1.524637
[13] Newton R G. Inverse scattering:Ⅱ. Three dimensions, continued[J]. J. Math. Phys., 1981, 22(10): 2191–2200. DOI:10.1063/1.524786
[14] Newton R G. Inverse scattering by a local impurity in a periodic potential in one dimension[J]. J. Math. Phys., 1983, 24(8): 2152–2162. DOI:10.1063/1.525968
[15] Colton D, Kress R. Inverse aconstic and electromagnetic scattering theory[M]. Berlin Heidelberg, New York: Springer Verlag, 1992.
[16] Colton D, Monk P. A novel method for solving the inverse scattering problem for time-harmonic aconstic waves in the resonance region[J]. SIAM J. Appl. Math., 1985, 45: 1039–1053. DOI:10.1137/0145064
[17] Colton D, Monk P. A novel method for solving the inverse scattering problem for time-harmonic aconstic waves in the resonance region Ⅱ[J]. SIAM J. Appl. Math., 1986, 46: 506–523. DOI:10.1137/0146034
[18] Colton D, Coyle J, Monk P. Recent developments in inverse acoustic scattering theory[J]. 2000, 42(3): 369-414.
[19] Bleistein N, Cohen J K, John W Stockwell Jr. Mathematics of Multi-dimensional Seismic Imaging, Migration, and Inversion[M]. New York: Springer Verlag, 1996.
[20] Bleistein N. On the imaging of reflectors in the earth[J]. Geophysics, 1987, 52(7): 931–942. DOI:10.1190/1.1442363
[21] Bleistein N, Cohen J K, Hagin F G. Two and one-half domension Born inversion with an arbitrary reference[J]. Geophysics, 1987, 52(1): 26–36.
[22] Bleistein N, Cohen J K, Hagin F G. Computional and asymptotic of velocity inversion[J]. Geophysics, 1985, 50(8): 1253–1265. DOI:10.1190/1.1441996
[23] Bleistein N, Cohen J K. The velocity inversion problem-Present status, new direction[J]. Geophysics, 1982, 47(11): 1497–1511. DOI:10.1190/1.1441300
[24] Bleistein N, Gray S H. An extension of the Born inversion method to a depth dependent reference profile[J]. Geophysical Prospecting, 1985, 33(8): 999–1022.
[25] Bleistein N. Two-and-one-half dimensional in-plane wave propagation[J]. Geophysical Prospecting, 1986, 34(6): 686–703.
[26] Bleistein N. Mathematical methods for wave phenomena[M]. Academic Press Inc., 1984.
[27] Coen S. Velocity, density profiles of a layered acoustic medium from common source-point data[J]. Geophysics, 1982, 47(6): 898–905. DOI:10.1190/1.1441356
[28] Cohen J K, Hagin F G. Velocity inversion using a stratified reference[J]. Geophysics, 1985, 50(11): 1689–1700. DOI:10.1190/1.1441860
[29] Cohen J K, Bleistein N. Three-dimensional Born inversion with an arbitrary referece[J]. Geophysics, 1986, 51(8): 1552–1558. DOI:10.1190/1.1442205
[30] Rose J H, Richardson T M. Time Domain Born Approximation[J]. J. Nondestructive Evaluation, 1982, 3: 45. DOI:10.1007/BF00566954
[31] Hsu D K, Rose J H, Thompson D O. Reconstruction of inclusions in solids using ultrasonic born inversion[J]. J. Appl. Phys., 1984, 55(1): 162–168. DOI:10.1063/1.332858
[32] Darmon M, Calmon P, Bele B. An integrated model to simulate the scattering of ultrasounds by inclusions in steels[J]. Ultrasonics, 2004, 42: 237–241. DOI:10.1016/j.ultras.2004.01.015
[33] D. Colton, R. Kress. Inverse acoustic and electromagnetic scattering theory[J]. Appl. Math. Sci., 2013, 93.
[34] 柳建军, 贺国强, 康传刚. 求解非线性不适定问题的隐式迭代法[J]. 应用数学和力学, 2009, 30(9): 1107–1116.
[35] 孙婷婷.求解非线性不适定问题的两类迭代法[D].哈尔滨: 哈尔滨理工大学, 2011.
[36] Adrian Doicu, Franz Schreier, Michael Hess. Iterative regularization methods for atmospheric remote sensing[J]. J. Quant. Spectr. Rad. Trans., 2002, 292: 1–15.
[37] Adrian Doicu, Franz Schreier, Michael Hess. An iterative regularization method with B-spline approximation for atmospheric temperature and concentration Retrievals[J]. Envir. Model. Soft., 2005: 1101–1109.