a
pa
0
e
4. Failure with the gross error case c of arrangement D,
where the test criterion wj = 3.15 of the x'-observation of
point no. 2 at image 1 remained slightly under the critical
value of c(wi4) = 3.29
Analysing these failures it becomes evident that in 1. there is a-priori no
chance for detection at all (zero variance problem). In 4. the situation was met,
where the actual gross error (30um) is smaller than the minimal gross error which
can be detected with the probability 8. - 0,8 (see Table 4). Setting vx? - - 32.4ym
in version D the corresponding test criterion exceeds the critical value, thus
leading to a correct rejection.
The false rejection of X^ at image 3 in case Ba and of X9 at image 3 in case Bb
is probably a result of shiftings caused by correlations. Actually the presence
of correlations disturbs significantly the statistical foundation of the gross
error problem, so that correlations should be subject to future studies. The
correlations become particularly of interest in a double sense:
- For the rejection procedure of data-snooping. The answer to the question
whether only max (wi) or all wi which exceed the critical value should
be rejected depends on the size of correlations.
- For the localization of gross errors the correlations are of decisive
importance. Consider our practical examples: In the cases Ad, Bd the
localization of VY, is impossible since the residuals Vy (image 1)
and Vy) (image 2) resp. Vy (image 3) and Vy? (image 4) are perfectly
correlating. Strong correlations exist also in case Cc between the
x'-residuals of a non-control point in all three images. Here we obtain
for the correlation coefficients of the x'-residuals of point no.2:
Fux [31v [0] = - 0.960 . Fux I3] vis [4] 0.956 ° vx [0] vx e [4] = - 0.960
(image number in brackets)
with the result that the gross error VX5 cannot be localized.
Experiences gained so far with the data-snooping technique are very encouraging,
expecially if its results are compared with conventional, non-statistical methods
of gross error detection.
From all 23 gross error cases of our examples (the 2 zero variance cases are ex-
cluded) we obtained by conventional methods and by data-snoopina:
lvl» 3o: 15 correct rejections, 12 incorrect rejetions
|v|> 36: IO correct rejections, "5 incorrect rejections
data-snooping: 19 correct rejections, 2 incorrect rejections
So it becomes an important and promising task to do further investigations in
this field of research to arrive lately at systems for point determination pro-
viding for both accuracy and reliability.