.3
GROSS ERRORS LOCATION BY TWO STEP ITERATIONS METHOD
Prof.
Wang Renxiang
Xian Research Institute of Surveying And Mapping, China , Communssion III
ABSTRICT
In consideration of the capability and relibility about
localizing gross errors are decreased by correlation of
residuals seriously. From the strategical point view,
the iterated weight least squares method is developed to
so-called ' Two step iterations method '. In the first
step of iterations, the observational weight is
calculated by selected weight function in an usual way.
In the second step, we start with statistical test and
analysis of residual correlation. Based on convergence
in the first step, obtain the possible gross error
observation(s) and weighted zero to its. Then the
second step iteration is performed. After that, gross
errors localization is done by regirous statistical test
according to the standardized residuals and with due
regard for the magnitude of so-called ’ weighted zero
residual’. The capability and relibility of localizing
gross errors are improved by two step iterations method.
The paper give some examples with simulated data for
comparasion of the results about gross errors location
by different step iterations methods.
KEYWORDS: Gross Error Location, Standardized residual,
Weighted zero residual, Qvv.P matrix.
INTRODUCTION
Gross errors localizing by iterated weight least squares
method has been investigated for a long time. One of the
key problems of this method is to select weight function
There are many weight functions proposed by different
authors in present application. Every weight function
has its own properties. Among these functions, the types
of function, the parameters and the statistical
quantities as well as the critical values are somewhath
different to each other. However a common property is
that the function is an inverse measure of residual in
absolute. Therefore the magnitute of main diagonal
element relating the observation with rather large
residual in absolute in Qvv.p matrix will be increased,
aftre iteration with the weight function and will be
capable of localizing gross errors (Wang Renxiang, 1986a)
The present author points out in (Wang Renxiang, 1988b,
appendix) that gross errors localization is unreliable by
iteration with weight function, when some residuals are
of strong correlation. The paper has proposed an idea
so-called ' Two step iteration method ’ in order to
improve the capability and reliability about localizing
gross errors.
1. THE TWO STEP ITERATIONS METHOD
The two step iterations method is proposed based on the
properties of so-called weighted zero residual (
appendix--3.) and the "'cheking correlation of residual
programm '( Wang Renxiang 1998b ). From the strategical
point of review, the iterated weight least squares
method has been contrived in two step iterations. The
first. step is to perform least squares iterations with
weight function untill convergence. The second step is
to analyse the correlation of residuals in which the
standardized value is large than the critical value.
Because of the first step iterated convergence, the
searching areas of gross error observations are limlted
in the observations in which the standardized residual
is of large or strong correlation with another large
standardized residual.
1.1 The First Step Iterations
In a general way, all the weight functions used in
iterated least squares method or robust estimate method
can be taken in the first step iterations. However the
present author emphasizes that standardized residuals
have to be used in every iteration at least in the last
one for statistical test. Therfore Qvv.P matrix should
be calculated in every iterations. The papers (Shan Jie
1988,Wang Renxiang 1990 in Chinese, appendix ) have give
the fast recurive algorithm for computation of Qvv.P
matrix. The time comsuming for calculatng Qvv.P matrix
have been overcome. As an experiment in this paper,
the author gives a weight function modified from ( Wang
Renxiang 1989c) and used in this step iterations as
fellows
I. i or CL
p = 1/,8 ; à «C for 1,2 iteration
lb 5 i > C after 2 iterations
where
a, Pl >}
a= | 80 2 50
^i , eleswhere
; | vi ;
i = -—— = enlarged residual
80 Ki
7c 4 = standardized residual
89 K2
Ki “VE = enlarged factor
Kf Ez standardized factor
€ 22.0, a = 2.5, b = 3.0--4.0
1.2 The Second Step Iterations
In-the first step, the mistakes of localizing gross
errors are from two major circumstances. The first is
the gross error that can not be detected by statistical
test with the standard critical value, because the
magnitute of main diagonal element related to gross
error observation in Qvv.P matrix and correspondene to
the main component coefficient of the standardized
residual MCCV ( appendix----3.2 ) are too small. It
is impossible to overcome by any iterationl method
determined by the design matrix. The second, gross error
revealed in the standardized residual is - dispersed by
the correlation coefficient of residuals and make wrong
decision with statistical test. This problem can
possibly be overcome by disassembling the correlation
of residuals. The properties of weighted zero residual
( appendix----3) as fellows plays important part in this
discussion.
1.2.1 Gross error can be revealed in its weighted
zero residual completly.
1.2.2 After iteration, any two observations are