WinFrog User’s Guide - Appendix C – GPS/MX 9400
Page 16 of
22
This dialog box provides you with the ability to enable or disable statistical w test and F
tests. By default, both of the w-test and the F-test are turned off.
w-Test:
The w-test entails normalizing the residuals of a GPS pseudorange solution and
testing these against a 99% confidence limit for outliers. This limit is 2.576. If outliers
are found, they are removed and the solution is re-executed excluding that satellite
data pertaining to the outlier. If more than one satellite is found to produce a residual
outlier, only that satellite with the largest normalized residual is eliminated. This
continues until no outliers are present
or
until the exclusion of any more data would
result in insufficient data for a solution (four satellites for 3D and three for 2D). Since
the normalized residuals tend towards being equal with reduced redundancy (as the
number of satellites used in the solution approaches the minimum required), it is
unlikely that WinFrog will ever have to stop the exclusion of data due
to insufficient
satellites. Nonetheless, WinFrog still checks for this condition.
Caution should be exercised when using this option. If the vessel is too far away
from the selected reference stations and/or there are few satellites in common with
the vessel and the reference station(s) the w-test option may eliminate data to the
point where the solution approaches the minimum required.
F-Test:
The F-test is a check of the unit variance of the GPS pseudorange solution. This
confirms the validity of the model used for the solution and the weighting of the
observations used. Note that it is only a confirmation check, no data or solution
results are thrown out based on the results of the test.
The unit variance is the sum of the weighted, squared residuals, divided by the
degrees of freedom (number of redundant measurements) in the solution. The F-
Test should result in
unity. If the unit variance is consistently different from unity, it
indicates that there may be a problem with the stochastic model used, or an
unmodeled bias in the data.
