文章快速检索     高级检索
  大地测量与地球动力学  2018, Vol. 38 Issue (1): 92-96  DOI: 10.14075/j.jgg.2018.01.020

引用本文  

陶武勇, 鲁铁定, 吴飞, 等. 求解球面拟合的改进总体最小二乘算法[J]. 大地测量与地球动力学, 2018, 38(1): 92-96.
TAO Wuyong, LU Tieding, WU Fei, et al. An Improved Total Least Squares Algorithm for Solving Sphere Surface Fitting[J]. Journal of Geodesy and Geodynamics, 2018, 38(1): 92-96.

项目来源

国家自然科学基金(41204003,41374007,41464001);测绘地理信息公益性行业科研专项(201512026);江西省教育厅科技项目(KJLD12077,GJJ13457);中国博士后科学基金(2012M511962);江西省中青年教师发展计划访问学者专项(2012132);江西省远航工程计划(2013)。

Foundation support

National Natural Science Foundation of China, No.41204003, 41374007, 41464001;Commonweal Research Project on Surveying and Geo-Informatics, No.201512026;Science and Technology Project of the Education Department of Jiangxi Province, No.KJLD12077, GJJ13457;China Postdoctoral Science Foundation, No.2012M511962;Special Visiting Scholar Project of Faculty Development of Jiangxi Province, No.2012132;Voyage Project of Jiangxi Province, No.2013.

通讯作者

鲁铁定,博士,教授,主要从事误差理论与测量平差研究,E-mail:tdlu@ecit.edu.cn

第一作者简介

陶武勇,助教,主要从事大地测量数据处理理论与方法研究,E-mail: 781873533@qq.com

About the first author

TAO Wuyong, assistant professor, majors in theory and method of geodetic data processing, E-mail: 781873533@qq.com.

文章历史

收稿日期:2016-12-31
求解球面拟合的改进总体最小二乘算法
陶武勇1     鲁铁定2     吴飞3     李丁3     
1. 江西信息应用职业技术学院测绘工程系,南昌市气象路58号, 330043;
2. 东华理工大学测绘工程学院,南昌市广兰大道418号, 330013;
3. 中国矿业大学环境与测绘学院,江苏省徐州市大学路1号, 221116
摘要:在球面拟合中,系数矩阵和观测向量的误差同源,均来源于球面点坐标。不同位置有相同元素存在,从理论上讲,这些相同元素应该有相同的改正数,并且非线性形式出现在观测向量中。为此,本文推导了一种改进的总体最小二乘算法,可以很好地克服上述问题。通过算例分析发现,本文方法得到的参数估值更为可靠,证明了本文方法的有效性。
关键词球面拟合非线性总体最小二乘线性化总体最小二乘

总体最小二乘同时考虑了系数矩阵和观测向量的误差,理论严密,受到广泛关注。目前,总体最小二乘已经被应用到各个方面。文献[1-2]针对线性回归,推导了总体最小二乘算法,考虑了自变量误差和系数矩阵中常数项;文献[3]推导了结构总体最小二乘,应用于自回归模型中,可以保证不同位置的相同元素有相同的改正数;文献[4]同样对自回归模型推导了一种新总体最小二乘方法,可以起到结构总体最小二乘一样的作用; 文献[5]将总体最小二乘应用到三维基准转换中;文献[6]将混合总体最小二乘应用到三维基准转换,并附加了约束条件;文献[7]推导了三维基准转换的总体最小二乘拟合推估,将求解坐标转换参数和非公共点坐标同时进行;文献[8]推导了附加约束条件的总体最小二乘,可用于大旋转角的三维坐标转换,但计算过于繁琐,存在系数矩阵含有常数和某些元素重复出现问题;文献[9]将多元总体最小二乘用于大旋转角的三维坐标转换,可以避免上述问题;此外,总体最小二乘还被应用到其他方面,针对不同的情况,学者们对总体最小二乘进行了不同的改进。

文献[10-11]推导了非线性总体最小二乘算法,当系数矩阵的元素不再是自变量本身,而是通过自变量计算得到,即是关于自变量的函数,将这种情况下的总体最小二乘平差称为非线性总体最小二乘平差,如GPS高程拟合。但以上都是针对非线性形式出现在系数矩阵的情况,而在球面拟合当中,非线性形式出现在观测向量中,并且系数矩阵中存在无误差的常数项和重复元素。为此,本文推导了一种改进的总体最小二乘算法来克服上述问题。

1 球面拟合

球面拟合是根据一定数量的球面点来求得球心坐标和球半径的过程,球面方程为:

$ {r^2} = {\left( {x - a} \right)^2} + {\left( {y - b} \right)^2} + {\left( {z - c} \right)^2} $ (1)

式中,(abc)为球心坐标,r为球半径,(x, y, z)为球面上点的坐标。

将球面方程展开得:

$ \begin{array}{*{20}{c}} {{x^2} + {y^2} + {z^2} = }\\ {2xa + 2yb + 2zc + {r^2} - {a^2} - {b^2} - {c^2}} \end{array} $ (2)

将式中abcr2-a2-b2-c2看作待估参数,当观测获得n个球面点时,则有:

$ \begin{array}{*{20}{c}} {\underbrace {\left[ {\begin{array}{*{20}{c}} {2{x_1}}&{2{y_1}}&{2{z_1}}&1\\ {2{x_2}}&{2{y_2}}&{2{z_2}}&1\\ \vdots&\vdots&\vdots&\vdots \\ {2{x_n}}&{2{y_n}}&{2{z_n}}&1 \end{array}} \right]}_\mathit{\boldsymbol{A}}\underbrace {\left[ {\begin{array}{*{20}{c}} a\\ b\\ c\\ {{r^2} - {a^2} - {b^2} - {c^2}} \end{array}} \right]}_\mathit{\boldsymbol{X}} = }\\ {\underbrace {\left[ {\begin{array}{*{20}{c}} {x_1^2 + y_1^2 + z_1^2}\\ {x_2^2 + y_2^2 + z_2^2}\\ \vdots \\ {x_n^2 + y_n^2 + z_n^2} \end{array}} \right]}_\mathit{\boldsymbol{Y}}} \end{array} $ (3)

式中,A为系数矩阵,Y为观测向量,X为待估参数。

采用最小二乘求解,得:

$ \mathit{\boldsymbol{\hat X = }}{\left( {{\mathit{\boldsymbol{A}}^{\rm{T}}}\mathit{\boldsymbol{A}}} \right)^{ - 1}}{\mathit{\boldsymbol{A}}^{\rm{T}}}\mathit{\boldsymbol{Y}} $ (4)

观察系数矩阵和观测向量不难发现,球面拟合中系数矩阵亦是由坐标组成的,同样含有误差,此时采用最小二乘求解并不合理,而应采用总体最小二乘求解。但是系数矩阵和观测向量的误差同源,增广矩阵[A Y]中不同位置有相同的元素存在,并且非线性形式出现在观测向量中,这些在求解球面拟合时都是需要考虑的,此时传统的总体最小二乘也不再合适。

2 改进的总体最小二乘算法

根据文献[10-11]非线性总体最小二乘的思路,非线性总体最小二乘问题可以看作是一个双非线性问题,可以通过两次线性化的方法获得非线性总体最小二乘的解:一次是在系数矩阵中的随机元素初值处线性化,一次是在参数估值的初值处线性化。

在球面拟合中,非线性形式出现在观测向量中,因此根据式(2)对观测向量中的随机元素线性化,即在球面点坐标(x0, y0, z0)处线性化,舍弃二阶极小项,整理得:

$ \begin{array}{*{20}{c}} {2{x^0}\Delta x - 2a\Delta x + 2{y^0}\Delta y - 2b\Delta y + 2{z^0}\Delta z - }\\ {2c\Delta z - 2a{x^0} - 2b{y^0} - 2c{z^0} + }\\ {{a^2} + {b^2} + {c^2} - {r^2} = - {{\left( {{x^0}} \right)}^2} - {{\left( {{y^0}} \right)}^2} - {{\left( {{z^0}} \right)}^2}} \end{array} $ (5)

将上式对参数估值线性化,即在a0b0c0r0处线性化,同样舍弃二阶极小项,整理得:

$ \begin{array}{*{20}{c}} {\left( {2{x^0} - 2{a^0}} \right)\Delta x + \left( {2{y^0} - 2{b^0}} \right)\Delta y + \left( {2{z^0} - 2{c^0}} \right)\Delta z = }\\ {\left( {2{x^0} - 2{a^0}} \right)\Delta a + \left( {2{y^0} - 2{b^0}} \right)\Delta b + \left( {2{z^0} - 2{c^0}} \right)\Delta c + }\\ {2{r^0}\Delta r - \left( {{{\left( {{x^0}} \right)}^2} + {{\left( {{y^0}} \right)}^2} + {{\left( {{z^0}} \right)}^2} + {{\left( {{a^0}} \right)}^2} + {{\left( {{b^0}} \right)}^2} + } \right.}\\ {\left. {{{\left( {{c^0}} \right)}^2} - {{\left( {{r^0}} \right)}^2} - 2{a^0}{x^0} - 2{b^0}{y^0} - 2{c^0}{z^0}} \right)} \end{array} $ (6)

假设有n个球面点,则可以得到:

$ \begin{array}{*{20}{c}} {\left[ {\begin{array}{*{20}{c}} {\left( {2x_1^0 - 2{a^0}} \right)\Delta {x_1} + \left( {2y_1^0 - 2{b^0}} \right)\Delta {y_1} + \left( {2z_1^0 - 2{c^0}} \right)\Delta {z_1}}\\ {\left( {2x_2^0 - 2{a^0}} \right)\Delta {x_2} + \left( {2y_2^0 - 2{b^0}} \right)\Delta {y_2} + \left( {2z_2^0 - 2{c^0}} \right)\Delta {z_2}}\\ \vdots \\ {\left( {2x_n^0 - 2{a^0}} \right)\Delta {x_n} + \left( {2y_n^0 - 2{b^0}} \right)\Delta {y_n} + \left( {2z_n^0 - 2{c^0}} \right)\Delta {z_n}} \end{array}} \right] = \\ \left[ {\begin{array}{*{20}{c}} {2x_1^0 - 2{a^0}}&{2y_1^0 - 2{b^0}}&{2z_1^0 - 2{c^0}}&{2{r^0}}\\ {2x_2^0 - 2{a^0}}&{2y_2^0 - 2{b^0}}&{2z_2^0 - 2{c^0}}&{2{r^0}}\\ \vdots&\vdots&\vdots&\vdots \\ {2x_n^0 - 2{a^0}}&{2y_n^0 - 2{b^0}}&{2z_n^0 - 2{c^0}}&{2{r^0}} \end{array}} \right]} {\left[ {\begin{array}{*{20}{c}} {\Delta a}\\ {\Delta b}\\ {\Delta c}\\ {\Delta r} \end{array}} \right] - \\ \left[ {\begin{array}{*{20}{c}} {{{\left( {x_1^0} \right)}^2} + {{\left( {y_1^0} \right)}^2} + {{\left( {z_1^0} \right)}^2} + {{\left( {{a^0}} \right)}^2} + {{\left( {{b^0}} \right)}^2} + {{\left( {{c^0}} \right)}^2} - {{\left( {{r^0}} \right)}^2} - 2{a^0}x_1^0 - 2{b^0}y_1^0 - 2{c^0}z_1^0}\\ {{{\left( {x_2^0} \right)}^2} + {{\left( {y_2^0} \right)}^2} + {{\left( {z_2^0} \right)}^2} + {{\left( {{a^0}} \right)}^2} + {{\left( {{b^0}} \right)}^2} + {{\left( {{c^0}} \right)}^2} - {{\left( {{r^0}} \right)}^2} - 2{a^0}x_2^0 - 2{b^0}y_2^0 - 2{c^0}z_2^0}\\ \vdots \\ {{{\left( {x_n^0} \right)}^2} + {{\left( {y_n^0} \right)}^2} + {{\left( {z_n^0} \right)}^2} + {{\left( {{a^0}} \right)}^2} + {{\left( {{b^0}} \right)}^2} + {{\left( {{c^0}} \right)}^2} - {{\left( {{r^0}} \right)}^2} - 2{a^0}x_n^0 - 2{b^0}y_n^0 - 2{c^0}z_n^0} \end{array}} \right]} \end{array} $ (7)

$\mathit{\boldsymbol{V}} = \left[{\begin{array}{*{20}{c}} {(2{x^0}_1-2{a^0})\Delta {x_1} + (2{y^0}_1-2{b^0})\Delta {y_1} + (2{z^0}_1-2{c^0})\Delta {z_1}}\\ {(2{x^0}_2 - 2{a^0})\Delta {x_2} + (2{y^0}_2 - 2{b^0})\Delta {y_2} + (2{z^0}_2 - 2{c^0})\Delta {z_2}}\\ \vdots \\ {(2{x^0}_n - 2{a^0})\Delta {x_n} + (2{y^0}_n - 2{b^0})\Delta {y_n} + (2{z^0}_n - 2{c^0})\Delta {z_n}} \end{array}} \right], \delta \mathit{\boldsymbol{\xi }} = \left[{\begin{array}{*{20}{c}} {\mathit{\Delta a}}\\ {\mathit{\Delta }b}\\ {\mathit{\Delta }c}\\ {\mathit{\Delta }r} \end{array}} \right]$$\mathit{\boldsymbol{A}}{^0} = \left[{\begin{array}{*{20}{c}} {2{x_1}^0-2{a^0}}&{2{y_1}^0-2{b^0}}&{2{z_1}^0-2{c^0}}&{2{r^0}}\\ {2{x_2}^0 - 2{a^0}}&{2{y_2}^0 - 2{b^0}}&{2{z_2}^0 - 2{c^0}}&{2{r^0}}\\ \vdots & \vdots & \vdots & \vdots \\ {2{x_n}^0 - 2{a^0}}&{2{y_n}^0 - 2{b^0}}&{2{z_n}^0 - 2{c^0}}&{2{r^0}} \end{array}} \right], \mathit{\boldsymbol{L}}{^0} = $$\left[{\begin{array}{*{20}{c}} {{{({x_1}^0)}^2} + {{({y_1}^0)}^2} + {{({z_1}^0)}^2} + {{({a^0})}^2} + {{({b^0})}^2} + {{({c^0})}^2}-{{({r^0})}^2}-2{a^0}{x_1}^0-2{b^0}{y_1}^0 - 2{c^0}{z_1}^0}\\ {{{({x_2}^0)}^2} + {{({y_2}^0)}^2} + {{({z_2}^0)}^2} + {{({a^0})}^2} + {{({b^0})}^2} + {{({c^0})}^2} - {{({r^0})}^2} - 2{a^0}{x_2}^0 - 2{b^0}{y_2}^0 - 2{c^0}{z_2}^0}\\ \vdots \\ {{{({x_n}^0)}^2} + {{({y_n}^0)}^2} + {{({z_n}^0)}^2} + {{({a^0})}^2} + {{({b^0})}^2} + {{({c^0})}^2} - {{({r^0})}^2} - 2{a^0}{x_n}^0 - 2{b^0}{y_n}^0 - 2{c^0}{z_n}^0} \end{array}} \right]$, 则式(7)可以变为:

$ \mathit{\boldsymbol{V}} = {\mathit{\boldsymbol{A}}^0}\delta \mathit{\boldsymbol{\xi }} - {\mathit{\boldsymbol{L}}^0} $ (8)

将各点球面坐标提取出来组成列向量:

$ \mathit{\boldsymbol{l}} = {\left[ {\begin{array}{*{20}{c}} {{x_1}}&{{x_2}}& \cdots &{{x_n}}&{{y_1}}&{{y_2}}& \cdots &{{y_n}}&{{z_1}}&{{z_2}}& \cdots &{{z_n}} \end{array}} \right]^{\rm{T}}} $ (9)

同时用el表示各点坐标误差改正数组成的列向量:

$ \begin{array}{*{20}{c}} {{\mathit{\boldsymbol{e}}_l} = \left[ {\begin{array}{*{20}{c}} {\Delta {x_1}}&{\Delta {x_2}}& \cdots &{\Delta {x_n}}&{\Delta {y_1}}&{\Delta {y_2}}& \cdots \end{array}} \right.}\\ {{{\left. {\begin{array}{*{20}{c}} {\Delta {y_n}}&{\Delta {z_1}}&{\Delta {z_2}}& \cdots &{\Delta {z_n}} \end{array}} \right]}^{\rm{T}}}} \end{array} $ (10)

则有V = B0el,并且

$ {\mathit{\boldsymbol{B}}^0} = {\left[ {\begin{array}{*{20}{c}} {2x_1^0 - 2{a^0}}&0& \cdots &0\\ 0&{2x_2^0 - 2{a^0}}& \cdots &0\\ \vdots&\vdots&\ddots&\vdots \\ 0&0& \cdots &{2x_n^0 - 2{a^0}}\\ {2y_1^0 - 2{b^0}}&0& \cdots &0\\ 0&{2y_2^0 - 2{b^0}}& \cdots &0\\ \vdots&\vdots&\ddots&\vdots \\ 0&0& \cdots &{2y_n^0 - 2{b^0}}\\ {2z_1^0 - 2{c^0}}&0& \cdots &0\\ 0&{2z_2^0 - 2{c^0}}& \cdots &0\\ \vdots&\vdots&\ddots&\vdots \\ 0&0& \cdots &{2z_n^0 - 2{c^0}} \end{array}} \right]^{\rm{T}}} $ (11)

此时总体最小二乘目标函数为elTPlel=min,Pl为坐标观测值的权阵,与式(8)联立得:

$ \left\{ \begin{array}{l} \mathit{\boldsymbol{e}}_l^{\rm{T}}{\mathit{\boldsymbol{P}}_l}{\mathit{\boldsymbol{e}}_l} = \min \\ {\mathit{\boldsymbol{B}}^0}{\mathit{\boldsymbol{e}}_l} = {\mathit{\boldsymbol{A}}^0}\delta \mathit{\boldsymbol{\xi }} - {\mathit{\boldsymbol{L}}^0} \end{array} \right. $ (12)

构建拉格朗日方程:

$ \mathit{\Phi }\left( {{\mathit{\boldsymbol{e}}_l},\delta \mathit{\boldsymbol{\xi }},\mathit{\boldsymbol{\lambda }}} \right) = \mathit{\boldsymbol{e}}_l^{\rm{T}}{\mathit{\boldsymbol{P}}_l}{\mathit{\boldsymbol{e}}_l} + 2{\mathit{\boldsymbol{\lambda }}^{\rm{T}}}\left( {{\mathit{\boldsymbol{L}}^0} + {\mathit{\boldsymbol{B}}^0}{\mathit{\boldsymbol{e}}_l} - {\mathit{\boldsymbol{A}}^0}\delta \mathit{\boldsymbol{\xi }}} \right) $ (13)

对上式求偏导数并令其等于0:

$ \frac{1}{2}\frac{{\partial \mathit{\Phi }}}{{\partial {\mathit{\boldsymbol{e}}_l}}} = {\mathit{\boldsymbol{P}}_l}{{\mathit{\boldsymbol{\hat e}}}_l} + {\left( {{\mathit{\boldsymbol{B}}^0}} \right)^{\rm{T}}}\mathit{\boldsymbol{\lambda }} = 0 $ (14)
$ \frac{1}{2}\frac{{\partial \mathit{\Phi }}}{{\partial \mathit{\boldsymbol{\lambda }}}} = {\mathit{\boldsymbol{L}}^0} - {\mathit{\boldsymbol{A}}^0}\delta \mathit{\boldsymbol{\xi }} + {\mathit{\boldsymbol{B}}^0}{{\mathit{\boldsymbol{\hat e}}}_l} = 0 $ (15)
$ \frac{1}{2}\frac{{\partial \mathit{\Phi }}}{{\partial \left( {\delta \mathit{\boldsymbol{\xi }}} \right)}} = - {\left( {{\mathit{\boldsymbol{A}}^0}} \right)^{\rm{T}}}\mathit{\boldsymbol{\hat \lambda }} = 0 $ (16)

通过整理可以得到:

$ \delta \mathit{\boldsymbol{\xi }} = {\left( {{\mathit{\boldsymbol{A}}^{0{\rm{T}}}}{{\left( {{\mathit{\boldsymbol{B}}^0}{\mathit{\boldsymbol{Q}}_l}{\mathit{\boldsymbol{B}}^{0{\rm{T}}}}} \right)}^{ - 1}}{\mathit{\boldsymbol{A}}^0}} \right)^{ - 1}}{\mathit{\boldsymbol{A}}^{0{\rm{T}}}}{\left( {{\mathit{\boldsymbol{B}}^0}{\mathit{\boldsymbol{Q}}_l}{\mathit{\boldsymbol{B}}^{0{\rm{T}}}}} \right)^{ - 1}}{\mathit{\boldsymbol{L}}^0} $ (17)
$ \mathit{\boldsymbol{\hat \lambda }} = {\left( {{\mathit{\boldsymbol{B}}^0}{\mathit{\boldsymbol{Q}}_l}{\mathit{\boldsymbol{B}}^{0{\rm{T}}}}} \right)^{ - 1}}\left( {{\mathit{\boldsymbol{A}}^0}\delta \mathit{\boldsymbol{\hat \xi }} - {\mathit{\boldsymbol{L}}^0}} \right) $ (18)
$ {{\mathit{\boldsymbol{\hat e}}}_l} = - {\mathit{\boldsymbol{Q}}_l}{\left( {{\mathit{\boldsymbol{B}}^0}} \right)^{\rm{T}}}\mathit{\boldsymbol{\hat \lambda }} $ (19)

式中,Ql= Pl-1

总结计算步骤如下:

1) 采用最小二乘计算球心坐标和球半径的初值a0b0c0r0,组成初值ξ(0)=[a0  b0  c0  r0]T,将坐标观测值作为坐标初值,组成l0=[x10  x20xn0  y10  y20yn0  z10  z20zn0]T

2) 构建A(i)B(i)L(i),分别按式(17)和式(19)计算$\delta {\boldsymbol{\hat \xi} ^{(i + 1)}}$$ \boldsymbol{\hat e}_l^{i + 1}$

3) 用公式${\mathit{\boldsymbol{\hat \xi }}^{(i + 1)}} = {\mathit{\boldsymbol{\hat \xi }}^{(i)}} + \delta {\mathit{\boldsymbol{\hat \xi }}^{(i + 1)}}, {\mathit{\boldsymbol{l}}^{(i + 1)}} = {\mathit{\boldsymbol{l}}^{(i)}} + \mathit{\boldsymbol{\hat e}}_l^{(i + 1)}$计算新的更新值,重新构建A(i+1)B(i+1)L(i+1)

4) 重复第2)、3)步,直到满足收敛条件为止。

5) 计算单位权方差和参数估值的方差:

$ \hat \sigma _0^2 = \frac{{\mathit{\boldsymbol{\hat e}}_l^{\rm{T}}{\mathit{\boldsymbol{P}}_l}{{\mathit{\boldsymbol{\hat e}}}_l}}}{{n - 4}} $ (20)
$ {\mathit{\boldsymbol{D}}_{\hat \xi }} = \hat \sigma _0^2{\mathit{\boldsymbol{Q}}_{\hat \xi }} = \hat \sigma _0^2{\left( {{\mathit{\boldsymbol{A}}^{\left( i \right){\rm{T}}}}{{\left( {{\mathit{\boldsymbol{B}}^{\left( i \right)}}{\mathit{\boldsymbol{Q}}_l}{\mathit{\boldsymbol{B}}^{\left( i \right){\rm{T}}}}} \right)}^{ - 1}}{\mathit{\boldsymbol{A}}^{\left( i \right)}}} \right)^{ - 1}} $ (21)
3 算例分析 3.1 算例1

利用MATLAB模拟数据,取球心坐标真值为a=20,b=30,c=40,球半径真值为r=5。在球面上选取12点,12个点的坐标真值如表 1所示。模拟时在12个点坐标上加入期望为0、中误差为0.1 m、服从正态分布的随机误差。为了比较分析,采用如下两种方案。

表 1 各点坐标真值 Tab. 1 True value of each point coordinates

1) 加权总体最小二乘(WTLS)[12]。系数矩阵的协因数阵和观测向量的协因数阵按下式计算:

$ {\mathit{\boldsymbol{Q}}_A} = \mathit{\boldsymbol{M}}{\mathit{\boldsymbol{Q}}_l}{\mathit{\boldsymbol{M}}^{\rm{T}}},{\mathit{\boldsymbol{Q}}_Y} = \mathit{\boldsymbol{N}}{\mathit{\boldsymbol{Q}}_l}{\mathit{\boldsymbol{N}}^{\rm{T}}} $ (22)

式中,$\mathit{\boldsymbol{N}} = \left[{\begin{array}{*{20}{c}} {2{x_1}}&0& \cdots &0&{2{y_1}}&0& \cdots &0&{2{z_1}}&0& \cdots &0\\ 0&{2{x_2}}& \cdots &0&0&{2{y_1}}& \cdots &0&0&{2{z_2}}& \cdots &0\\ \vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots \\ 0&0& \vdots &{2{x_{12}}}&0&0& \vdots &{2{z_{12}}}&0&0& \vdots &{2{z_{12}}} \end{array}} \right], \mathit{\boldsymbol{M}} = \left[\begin{array}{l} \mathit{\boldsymbol{J}}\\ \mathit{\boldsymbol{K}} \end{array} \right]$J =2I36K = 0 12×36I36为36×36的单位矩阵,012×36为12×36的零矩阵。

2) 本文的总体最小二乘(ITLS)。模拟500次实验,将500次参数估值的平均值列于表 2

表 2 参数估值的平均值 Tab. 2 Mean value of parameter estimation

表 2可以看出,本文方法的计算结果参数估值的平均值与真值之差要小于WTLS计算的参数估值平均值与真值之差,计算结果更为接近真值,说明本文的方法更为可靠,能够得到更好的计算结果。

为了进一步说明本文方法的有效性,现进行如下实验。向坐标加入从0.01 m开始到0.22 m结束、步长为0.03的中误差,每个先验中误差下模拟500次,同样采用WTLS和ITLS进行计算。两种方法得到的参数估值的平均值与真值之差同先验中误差的关系如图 1~4所示。

图 1 a的平均值与真值的差值 Fig. 1 Difference between mean value of a and true value

图 2 b的平均值与真值的差值 Fig. 2 Difference between mean value of b and true value

图 3 c的平均值与真值的差值 Fig. 3 Difference between mean value of c and true value

图 4 r的平均值与真值的差值 Fig. 4 Difference between mean value of r and true value

从图中可以看出,在加入误差较小时,两种算法的计算结果差异不大,随着加入的误差越来越大,两种算法的计算结果偏离真值越远。整体看来,ITLS计算结果要优于WTLS的计算结果,每次计算结果都更为接近真值,特别是球半径的计算结果改进较为明显。因为WTLS没有考虑观测向量中的非线性形式,导致WTLS计算结果不合理,而ITLS则考虑了这一问题,可以得到更为可靠的参数估计。

3.2 算例2

采用表 1中相同的真值,对1~4号点的三维坐标加入中误差为0.1 m的随机误差,5~8号点的三维坐标加入中误差为0.2 m的随机误差,9~12号点的三维坐标加入中误差为0.3 m的随机误差,比较在精度不等的情况下两种算法的计算结果。同样模拟500次实验,计算参数估值的平均值,计算结果列于表 3

表 3 参数估值的平均值 Tab. 3 Mean value of parameter estimation

表 3可以看出,ITLS计算结果比WTLS要好,与真值更为接近,说明在精度不等的情况下ITLS同样有效,更加说明了本文方法的可行性。

4 结语

本文将总体最小二乘应用到球面拟合中,推导了一种改进的总体最小二乘算法。这种算法考虑到了观测向量中的非线性形式,还能保证系数矩阵和观测向量中不同位置的相同元素有相同的改正数,理论更为严谨。通过与加权总体最小二乘算法进行比较发现,本文方法的计算结果更为可靠,可以得到更准确的参数估值。

参考文献
[1]
汪奇生, 杨德宏, 杨根新. 考虑自变量误差的线性回归迭代算法[J]. 大地测量与地球动力学, 2014, 34(5): 110-113 (Wang Qisheng, Yang Dehong, Yang Genxin. Iteration Algorithm of Linear Regression Considering the Error of Independent Variables[J]. Journal of Geodesy and Geodynamics, 2014, 34(5): 110-113) (0)
[2]
汪奇生, 杨德宏, 杨建文. 基于总体最小二乘的线性回归迭代算法[J]. 大地测量与地球动力学, 2013, 33(6): 112-115 (Wang Qisheng, Yang Dehong, Yang Jianwen. An Iterative Algorithm of Linear Regression Based on Total Least Squares[J]. Journal of Geodesy and Geodynamics, 2013, 33(6): 112-115) (0)
[3]
胡川, 陈义, 彭友. 自回归模型参数的结构总体最小二乘估计[J]. 大地测量与地球动力学, 2013, 33(1): 45-47 (Hu Chuan, Chen Yi, Peng You. Parameters Estimation of Auto Regression with Structured Total Least Squares[J]. Journal of Geodesy and Geodynamics, 2013, 33(1): 45-47) (0)
[4]
姚宜斌, 黄书华, 陈家君. 求解自回归模型参数的整体最小二乘新方法[J]. 武汉大学学报:信息科学版, 2014, 39(12): 1463-1466 (Yao Yibin, Huang Shuhua, Chen Jiajun. A New Method for Solving the Parameters of Auto Regressive Model[J]. Geomatics and Information Science of Wuhan University, 2014, 39(12): 1463-1466) (0)
[5]
Fang X. A Total Least Squares Solution for Geodetic Datum Transformations[J]. Acta Geodaetica et Geophysica, 2014, 49(2): 189-207 DOI:10.1007/s40328-014-0046-8 (0)
[6]
葛旭明, 伍吉仓. 三维基准转换的约束加权混合整体最小二乘的迭代解法[J]. 武汉大学学报:信息科学版, 2012, 37(2): 178-182 (Ge Xunming, Wu Jicang. Iterative Method of Weighted Constraint Total Least Squares for Three-Dimensional Datum Transformation[J]. Geomatics and Information Science of Wuhan University, 2012, 37(2): 178-182) (0)
[7]
李博峰, 沈云中, 李薇晓. 无缝三维基准转换模型[J]. 中国科学:地球科学, 2012, 42(7): 1047-1054 (Li Bofeng, Shen Yunzhong, Li Weixiao. The Seamless Model for Three-dimensional Datum Transformation[J]. Sci China Earth Sci, 2012, 42(7): 1047-1054) (0)
[8]
陆珏, 陈义, 郑波. 三维非线性基准转换的加权和加约束总体最小二乘解算模型[J]. 大地测量与地球动力学, 2012, 32(6): 129-134 (Lu Jue, Chen Yi, Zheng Bo. Performing Nonlinear 3D Datum Transformation Using WTLS and CTLS Method[J]. Journal of Geodesy and Geodynamics, 2012, 32(6): 129-134) (0)
[9]
杨仕平, 范东明, 龙玉春. 基于整体最小二乘法的任意旋转角度三维坐标转换[J]. 大地测量与地球动力学, 2013, 33(2): 114-119 (Yang Shiping, Fan Dongming, Long Yuchun. Three-Dimensional Coordination Transformation Adapted to Arbitrary Rotation Angle Based on Total Least Squares Method.[J]. Journal of Geodesy and Geodynamics, 2013, 33(2): 114-119) (0)
[10]
胡川, 陈义. 非线性整体最小二乘平差迭代算法[J]. 测绘学报, 2014, 43(7): 668-674 (Hu Chuan, Chen Yi. An Iterative Algorithm for Nonlinear Total Least Squares Adjustment[J]. Acta Geodaetica et Cartographica Sinica, 2014, 43(7): 668-674) (0)
[11]
Mahboub V, Ardalan A A, Ebrahimzadeh S. Adjustment of Non-Typical Errors-in-Variables Models[J]. Acta Geodaetica et Geophysica, 2015, 50(2): 207-218 DOI:10.1007/s40328-015-0109-5 (0)
[12]
Shen Y Z, Li B F, Chen Y. An Iterative Solution of Weighted Total Least Squares Adjustment[J]. J Geod, 2012, 86(5): 359-367 DOI:10.1007/s00190-011-0524-5 (0)
An Improved Total Least Squares Algorithm for Solving Sphere Surface Fitting
TAO Wuyong1     LU Tieding2     WU Fei3     LI Ding3     
1. Department of Surveying and Mapping, Jiangxi Information Institute of Profession Technology, 58 Qixiang Road, Nanchang 330043, China;
2. Faculty of Geomatics, East China Institute of Technology, 418 Guanglan Road, Nanchang 330013, China;
3. School of Environment Science and Spatial Informatics, China University of Mining and Technology, 1 Daxue Road, Xuzhou 221116, China
Abstract: In sphere surface fitting, the errors in the coefficient matrix and observation vector are from the same place, which is the coordinate of the spherical point.The same elements exist in different places.These same elements ought to have the same corrections.The linear form appears in the observation vector.Therefore, an improved total least squares algorithm is deduced which can overcome the above problems.Through analysis of examples, we discover that the parameters obtained from the method in the paper are more reliable.This demonstrates that the method is feasible and effective.
Key words: sphere surface fitting; nonlinear total least squares; linearization; total least squares