﻿ 病态加权总体最小二乘模型的正则化抗差解法
 文章快速检索 高级检索
 大地测量与地球动力学  2021, Vol. 41 Issue (11): 1106-1110  DOI: 10.14075/j.jgg.2021.11.002

### 引用本文

ZOU Shilin, WU Xing, WANG Fengwei. Regularized Robust Solution for Ill-Posed Weighted Total Least Squares Model[J]. Journal of Geodesy and Geodynamics, 2021, 41(11): 1106-1110.

### 第一作者简介

ZOU Shilin, associate professor, majors in survey data processing and geographic information application, E-mail: slzou@ecut.edu.cn.

### 文章历史

1. 东华理工大学测绘工程学院，南昌市广兰大道418号，330013;
2. 东华理工大学勘察设计研究院，江西省抚州市学府路56号，344000;
3. 同济大学测绘与地理信息学院，上海市四平路1239号，200092

1 等权病态总体最小二乘模型的正则化解

 ${\boldsymbol{y}} - {{\boldsymbol{e}}_y} = ({\boldsymbol{A}} - {{\boldsymbol{E}}_A}){\boldsymbol{x}}$ (1)

 $\left[ {\begin{array}{*{20}{c}} {{{\boldsymbol{e}}_y}}\\ {{{\boldsymbol{e}}_A}} \end{array}} \right]\sim \left( {\begin{array}{*{20}{c}} {\left[ {\begin{array}{*{20}{c}} 0\\ 0 \end{array}} \right]}&{\sigma _{\rm{0}}^{\rm{2}}\left[ {\begin{array}{*{20}{c}} {{{\boldsymbol{Q}}_y}}&{}\\ {}&{{{\boldsymbol{Q}}_A}} \end{array}} \right]} \end{array}} \right)$ (2)

 ${\boldsymbol{e}}_y^{\rm{T}}{{\boldsymbol{e}}_y} + {\boldsymbol{e}}_A^{\rm{T}}{{\boldsymbol{e}}_A} + \alpha {{\boldsymbol{x}}^{\rm{T}}}{\boldsymbol{x}} = \min$ (3)

 ${\boldsymbol{\hat x}} = {({{\boldsymbol{A}}^{\rm{T}}}{\boldsymbol{A}} + b{{\boldsymbol{I}}_n})^{ - 1}}{{\boldsymbol{A}}^{\rm{T}}}{\boldsymbol{y}}$ (4)

2 病态加权总体最小二乘模型的正则化抗差解 2.1 正则化解

 ${{\boldsymbol{Q}}_y}{\boldsymbol{ = P}}_y^{ - 1}{\boldsymbol{, }}{{\boldsymbol{Q}}_A} = {{\boldsymbol{Q}}_0} \otimes {{\boldsymbol{Q}}_{\boldsymbol{x}}}$ (5)

 ${\boldsymbol{e}}_y^{\rm{T}}{{\boldsymbol{P}}_y}{{\boldsymbol{e}}_y} + {\boldsymbol{e}}_A^{\rm{T}}({{\boldsymbol{P}}_0} \otimes {{\boldsymbol{P}}_{\boldsymbol{x}}}){{\boldsymbol{e}}_A} + \alpha {{\boldsymbol{x}}^{\rm{T}}}{\boldsymbol{x}} = \min$ (6)

 $\begin{array}{l} \varPhi ({{\boldsymbol{e}}_y}, {{\boldsymbol{e}}_A}, {\boldsymbol{\lambda }}, {\boldsymbol{x}}) = {\boldsymbol{e}}_y^{\rm{T}}{{\boldsymbol{P}}_y}{{\boldsymbol{e}}_y} + {\boldsymbol{e}}_A^{\rm{T}}({{\boldsymbol{P}}_0} \otimes {{\boldsymbol{P}}_{\boldsymbol{x}}}){{\boldsymbol{e}}_A} +\\ \alpha {{\boldsymbol{x}}^{\rm{T}}}{\boldsymbol{x}} + 2{{\boldsymbol{\lambda }}^{\rm{T}}}({\boldsymbol{y}} - {\boldsymbol{Ax}} - {{\boldsymbol{e}}_y} + ({{\boldsymbol{x}}^{\rm{T}}} \otimes {{\boldsymbol{I}}_n}) \cdot {{\boldsymbol{e}}_A}) \end{array}$ (7)

 $\frac{1}{2}\frac{{\partial \varPhi }}{{\partial {{\boldsymbol{e}}_y}}} = {\boldsymbol{Q}}_y^{ - 1}{{\boldsymbol{\tilde e}}_y} - {\boldsymbol{\hat \lambda }} = {\rm{0}}$ (8a)
 $\frac{1}{2}\frac{{\partial \varPhi }}{{\partial {{\boldsymbol{e}}_A}}} = ({\boldsymbol{Q}}_0^{ - 1} \otimes {\boldsymbol{Q}}_{\boldsymbol{x}}^{ - 1}){{\boldsymbol{\tilde e}}_A} + ({\boldsymbol{\hat x}} \otimes {{\boldsymbol{I}}_n}){\boldsymbol{\hat \lambda }} = {\rm{0}}$ (8b)
 $\frac{1}{2}\frac{{\partial \varPhi }}{{\partial {\boldsymbol{\hat \lambda }}}} = {\boldsymbol{y}} - {\boldsymbol{A\hat x}} - {{\boldsymbol{\tilde e}}_y} + ({{\boldsymbol{\hat x}}^{\rm{T}}} \otimes {{\boldsymbol{I}}_n}){{\boldsymbol{\tilde e}}_A} = {\rm{0}}$ (8c)
 $\frac{1}{2}\frac{{\partial \varPhi }}{{\partial {\boldsymbol{\hat x}}}} = \alpha {\boldsymbol{\hat x}} - {{\boldsymbol{A}}^{\rm{T}}}{\boldsymbol{\hat \lambda }} + {\boldsymbol{\tilde E}}_A^{\rm{T}}{\boldsymbol{\hat \lambda }} = {\rm{0}}$ (8d)

 ${{\boldsymbol{\tilde e}}_y} = {{\boldsymbol{Q}}_y}{\boldsymbol{\hat \lambda }}$ (9a)
 ${{\boldsymbol{\tilde e}}_A} = - ({{\boldsymbol{Q}}_0}{\boldsymbol{\hat x}} \otimes {{\boldsymbol{Q}}_{\boldsymbol{x}}}){\boldsymbol{\hat \lambda }}$ (9b)
 ${{\boldsymbol{\tilde E}}_A} = - {{\boldsymbol{Q}}_x}{\boldsymbol{\hat \lambda }}{{\boldsymbol{\hat x}}^{\rm{T}}}{{\boldsymbol{Q}}_0}$ (9c)

 ${\boldsymbol{\hat \lambda }} = {\left( {{{\boldsymbol{Q}}_y} + ({{{\boldsymbol{\hat x}}}^{\rm{T}}}{{\boldsymbol{Q}}_0}{\boldsymbol{\hat x}}){{\boldsymbol{Q}}_{\boldsymbol{x}}}} \right)^{ - 1}}({\boldsymbol{y}} - {\boldsymbol{A\hat x}})$ (10)

 $\begin{array}{l} - {{\boldsymbol{A}}^{\rm{T}}}{\boldsymbol{\hat \lambda }} = {{\boldsymbol{A}}^{\rm{T}}}{\left( {{{\boldsymbol{Q}}_y} + ({{{\boldsymbol{\hat x}}}^{\rm{T}}}{{\boldsymbol{Q}}_0}{\boldsymbol{\hat x}}){{\boldsymbol{Q}}_{\boldsymbol{x}}}} \right)^{ - 1}}({\boldsymbol{A\hat x}} - {\boldsymbol{y}})=\\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} - {\boldsymbol{\tilde E}}_A^{\rm{T}}{\boldsymbol{\hat \lambda }} - \alpha {\boldsymbol{\hat x}} \end{array}$ (11)

 ${{\boldsymbol{A}}^{\rm{T}}}{\left( {{{\boldsymbol{Q}}_y} + ({{{\boldsymbol{\hat x}}}^{\rm{T}}}{{\boldsymbol{Q}}_0}{\boldsymbol{\hat x}}){{\boldsymbol{Q}}_{\boldsymbol{x}}}} \right)^{ - 1}}({\boldsymbol{A\hat x}} - {\boldsymbol{y}}) = {{\boldsymbol{Q}}_0}{\boldsymbol{\hat x}}\hat v - \alpha {\boldsymbol{\hat x}}$ (12)

 ${\boldsymbol{\hat x}} = {({{\boldsymbol{A}}^{\rm{T}}}{{\boldsymbol{Q}}_{\boldsymbol{l}}}{\boldsymbol{A}} - \hat v{{\boldsymbol{Q}}_0} + \alpha {{\boldsymbol{I}}_n})^{ - 1}}{{\boldsymbol{A}}^{\rm{T}}}{{\boldsymbol{Q}}_{\boldsymbol{l}}}{\boldsymbol{y}}$ (13)

 $\hat \sigma = \frac{{{\boldsymbol{\tilde e}}_y^{\rm{T}}{{\boldsymbol{Q}}_y}{{{\boldsymbol{\tilde e}}}_y} + {\boldsymbol{\tilde e}}_A^{\rm{T}}{{\boldsymbol{Q}}_A}{{{\boldsymbol{\tilde e}}}_A}}}{{m - n}}$ (14)

 $\begin{array}{c} {{\boldsymbol{Q}}_{\boldsymbol{l}}} = {(1 + {{{\boldsymbol{\hat x}}}^{\rm{T}}}{\boldsymbol{x}})^{ - 1}}{{\boldsymbol{I}}_m}\\ {\boldsymbol{\hat \lambda }} = {(1 + {{{\boldsymbol{\hat x}}}^{\rm{T}}}{\boldsymbol{x}})^{ - 1}}({\boldsymbol{y}} - {\boldsymbol{A\hat x}})\\ \hat v = {(1 + {{{\boldsymbol{\hat x}}}^{\rm{T}}}{\boldsymbol{x}})^{ - 2}}{({\boldsymbol{y}} - {\boldsymbol{A\hat x}})^{\rm{T}}}({\boldsymbol{y}} - {\boldsymbol{A\hat x}}) \end{array}$ (15)

 ${\boldsymbol{\hat x}} = {\left( {{{\boldsymbol{A}}^{\rm{T}}}{\boldsymbol{A}} + (\alpha - \hat v)(1 + {{{\boldsymbol{\hat x}}}^{\rm{T}}}{\boldsymbol{\hat x}}){{\boldsymbol{I}}_n}} \right)^{ - 1}}{{\boldsymbol{A}}^{\rm{T}}}{\boldsymbol{y}}$ (16)

2.2 基于中位数法的正则化抗差解

 ${{\bar P}_i} = \left\{ \begin{array}{l} {P_i}{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} \left| {{{\hat e}_i} \le a} \right|\\ {P_i}a{\rm{/}}\left| {{{\hat v}_i}} \right|\;\;\;\;\;\;\;\;\;a < \left| {{{\hat e}_i}} \right| \le b\\ 0{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} b < \left| {{{\hat e}_i}} \right| \end{array} \right.$ (17)

 ${\boldsymbol{\hat X}} = \left[ {\begin{array}{*{20}{c}} {{{{\boldsymbol{\hat X}}}_1}}& \cdots &{{{{\boldsymbol{\hat X}}}_{C_m^n}}} \end{array}} \right] = \left[ {\begin{array}{*{20}{c}} {\hat x_1^1}& \cdots &{\hat x_{C_m^n}^1}\\ \vdots &{}& \vdots \\ {\hat x_1^n}& \cdots &{\hat x_{C_m^n}^n} \end{array}} \right]$ (18)

3 算例分析 3.1 数值算例

 $z(y) = \int_a^b {K(x, y)f(x){\rm{d}}x}$ (21)

 $\begin{array}{l} K(x, y) = \frac{{1.0}}{{1.0 + 100{{(y - x)}^2}}}\\ f(x) = \frac{{\exp ({\beta _1}) + \exp ({\beta _2})}}{{10.9550408}} - 0.052130913 \end{array}$ (22)

 $\begin{array}{l} z({y_i}) = \frac{1}{2}\Delta x\sum\limits_{j = 1}^{50} {[K({x_j},{y_i})f({x_j}) + } \\ \;\;\;\;\;\;\;\;\;K({x_{j + 1}},{y_i})f({x_{j + 1}})] \end{array}$ (23)

 $RMSE = \sqrt {\frac{{{{({\boldsymbol{\hat X}} - {\boldsymbol{\bar X}})}^{\rm{T}}}({\boldsymbol{\hat X}} - {\boldsymbol{\bar X}})}}{n}}$ (24)

 图 1 最小二乘解和总体最小二乘解 Fig. 1 Least squares solution and total least squares solution

 图 2 正则化解、正则化抗差解与真值对比 Fig. 2 Comparison between regularized solution, regularized robust solution and true values

 图 3 不同算法500次实验获得估值的RMSE Fig. 3 The root mean square error of parameter estimation of four algorithms for 500 experiments
3.2 病态测边网算例

 图 4 空间测边网平面点位分布 Fig. 4 The point position distribution of the space net in XY plane

4 结语

 [1] Li B F, Shen Y Z, Feng Y M. Fast GNSS Ambiguity Resolution as an Ill-Posed Problem[J]. Journal of Geodesy, 2010, 84(11): 683-698 DOI:10.1007/s00190-010-0403-5 (0) [2] Pan S G, Gao W, Wang S L, et al. Analysis of Ill Posedness in Double Differential Ambiguity Resolution of BDS[J]. Survey review, 2014, 46(339): 411-416 DOI:10.1179/1752270614Y.0000000121 (0) [3] 顾勇为, 归庆明, 张璇, 等. 大地测量与地球物理中病态性问题的正则化迭代解法[J]. 测绘学报, 2014, 43(4): 331-336 (Gu Yongwei, Gui Qingming, Zhang Xuan, et al. Iterative Solution of Regularization to Ill-Conditioned Problems in Geodesy and Geophysics[J]. Acta Geodaetica et Cartographica Sinica, 2014, 43(4): 331-336) (0) [4] Fu H Q, Wang C C, Zhu J J, et al. Estimation of Pine Forest Height and Underlying DEM Using Multi-Baseline P-Band PolInSAR Data[J]. Remote Sensing, 2016, 8(10): 820 DOI:10.3390/rs8100820 (0) [5] Save H, Bettadpur S, Tapley B D. Reducing Errors in the GRACE Gravity Solutions Using Regularization[J]. Journal of Geodesy, 2012, 86(9): 695-711 DOI:10.1007/s00190-012-0548-5 (0) [6] Freeden W, Schreiner M. Spaceborne Gravitational Field Determination by Means of Locally Supported Wavelets[J]. Journal of Geodesy, 2005, 79(8): 431-446 DOI:10.1007/s00190-005-0482-x (0) [7] Tikhonov A N, Goncharsky A V, Stepanov V V, et al. Numerical Methods for the Solution of Ill-Posed Problems[M]. Dordrecht: Springer Netherlands, 1995 (0) [8] Xu P L. Truncated SVD Methods for Discrete Linear Ill-Posed Problems[J]. Geophysical Journal International, 1998, 135(2): 505-514 DOI:10.1046/j.1365-246X.1998.00652.x (0) [9] Fierro R D, Golub G H, Hansen P C, et al. Regularization by Truncated Total Least Squares[J]. SIAM Journal on Scientific Computing, 1997, 18(4): 1223-1241 DOI:10.1137/S1064827594263837 (0) [10] 葛旭明, 伍吉仓. 病态总体最小二乘问题的广义正则化[J]. 测绘学报, 2012, 41(3): 372-377 (Ge Xuming, Wu Jicang. Generalized Regularization to Ill-Posed Total Least Squares Problem[J]. Acta Geodaetica et Cartographica Sinica, 2012, 41(3): 372-377) (0) [11] 孙同贺, 闫国庆. 病态总体最小二乘的混合正则化算法[J]. 大地测量与地球动力学, 2017, 37(4): 390-393 (Sun Tonghe, Yan Guoqing. A Hybrid Regularization Algorithm to Ill-Posed Total Least Squares[J]. Journal of Geodesy and Geodynamics, 2017, 37(4): 390-393) (0) [12] 王乐洋, 于冬冬. 病态总体最小二乘问题的虚拟观测解法[J]. 测绘学报, 2014, 43(6): 575-581 (Wang Leyang, Yu Dongdong. Virtual Observation Method to Ill-Posed Total Least Squares Problem[J]. Acta Geodaetica et Cartographica Sinica, 2014, 43(6): 575-581) (0) [13] Wang L Y, Wen G S, Zhao Y W. Virtual Observation Method and Precision Estimation for Ill-Posed Partial EIV Model[J]. Journal of Surveying Engineering, 2019, 145(4) (0) [14] 王乐洋, 许才军, 鲁铁定. 病态加权总体最小二乘平差的岭估计解法[J]. 武汉大学学报: 信息科学版, 2010, 35(11): 1346-1350 (Wang Leyang, Xu Caijun, Lu Tieding. Ridge Estimation Method in Ill-Posed Weighted Total Least Squares Adjustment[J]. Geomatics and Information Science of Wuhan University, 2010, 35(11): 1346-1350) (0) [15] 王彬, 李建成, 高井祥, 等. 抗差加权整体最小二乘模型的牛顿-高斯算法[J]. 测绘学报, 2015, 44(6): 602-608 (Wang Bin, Li Jiancheng, Gao Jingxiang, et al. Newton-Gauss Algorithm of Robust Weighted Total Least Squares Model[J]. Acta Geodaetica et Cartographica Sinica, 2015, 44(6): 602-608) (0) [16] Xu P L, Liu J N, Shi C. Total Least Squares Adjustment in Partial Errors-in-Variables Models: Algorithm and Statistical Analysis[J]. Journal of Geodesy, 2012, 86(8): 661-675 DOI:10.1007/s00190-012-0552-9 (0)
Regularized Robust Solution for Ill-Posed Weighted Total Least Squares Model
ZOU Shilin1,2     WU Xing2     WANG Fengwei3
1. Faculty of Geomatics, East China University of Technology, 418 Guanglan Road, Nanchang 330013, China;
2. Institute of Survey and Design, East China University of Technology, 56 Xuefu Road, Fuzhou 344000, China;
3. College of Surveying and Geo-Informatics, Tongji University, 1239 Siping Road, Shanghai 200092, China
Abstract: The coefficient matrix of the error-in-variables (EIV) model is ill-posed, and the precision of the coefficient matrix and the observation are not equal, so we derive the regularized solution of the ill-posed weighted total least squares model using the Lagrangian multiplier method. We prove the existing regularized solution of the ill-posed total least squares model to be a special case of the new model. On this basis, we propose the regularized robust solution based on the median method for ill-posed weighted total least squares model, and the effectiveness of the new algorithm is verified by two examples of the first kind Fredholm integral equation and ill-posed trilateration network. The results show that the least squares solution and the total least squares solution have poor accuracies and seriously deviate from the true value due to the influence of ill-posedness and the outliers of the model, while the accuracy of the regularized solution has been improved for weakening the ill-posedness of the model taking into account the errors of coefficient matrix and the observations. On the basis of regularized solution, the regularized robust solution reconstructs the weight matrix using the equivalent weight function, which can effectively resist the influence of gross error and has the highest accuracy.
Key words: ill-posed model; total least squares; regularized method; robust estimation