Abstract
Reliability analysis is inseparably connected with the formulation of failure scenarios, and common test statistics are based on specific assumptions. This is easily overlooked when processing observation differences. Poor failure identification performance and misleading pre-analysis results, mainly meaningless minimum detectable biases and external reliability measures, are the consequence. A reasonable failure scenario for use with differenced GNSS observations is formulated which takes into account that individual outliers in the original data affect more than one processed observation. The proper test statistics and reliability indicators are given for use with correlated observations and both batch processing and Kalman filtering. It is also shown that standardized residuals and redundancy numbers fail completely when used with double differenced observations.
Similar content being viewed by others
Notes
Center for Orbit Determination in Europe, http://www.cx.unibe.ch/aiub/igs.html
On larger networks, certain baseline selections may be more favorable in terms of ambiguity resolution than others, but as long as only float solutions are considered, or once the ambiguities have been fixed, a transition to another linear combination of the same UD does not change the results.
For example, for the epoch 8:00, this is the length of the longest line emanating from A in Fig. 6.
A more efficient procedure would always search for a maximum set of linear independent DD, thus occasionally introducing DD from additional ‘baselines’ into the adjustment.
MathWorks, http://www.mathworks.com
References
Baarda W (1968) A testing procedure for use in geodetic networks, vol 2–5, new edn. Geodetic Commission, Delft, Netherlands
Caspary WF (1988) Concepts of network and deformation analysis. Monograph 11, School of Surveying, The University of New South Wales, Sydney, Australia
Koch KR, Yang Y (1998) Robust Kalman filter for rank deficient observations models. J Geod 72:436–441
Pope AJ (1975) The statistics of residuals and the detection of outliers. Department of Commerce, NOAA, National Ocean Survey, Technical Report NOS 65 NGS 1, Rockville, Maryland
Rothacher M, Springer T, Schaer S, Beutler G (1997) Processing strategies for regional GPS networks. In: Brunner FK (ed) Advances in positioning and reference frames, Proc. IAG Scientific Assembly, 3–9 September 1997, Rio de Janeiro, pp 93–100
Rousseeuw PJ, Leroy AM (1987) Robust regression and outlier detection. Wiley, New York
Sachs L (1997) Angewandte Statistik, 8th edn. Springer, Berlin Heidelberg New York, 884 pp
Salzmann MA (1990) MDB: a design tool for integrated navigation systems. In: Schwarz KP, Lachapelle G (eds) Proc. Kinematic Systems in Geodesy IAG Symposia, vol 107, Banff, Canada, 10–13 September 1990, pp 218–227
Schaffrin B (1997) Reliability measures for correlated observations. J Surv Eng 123:126–137
Teunissen PJG (1998a) Quality control and GPS. In: Teunissen PJG, Kleusberg A (eds) GPS for geodesy, 2nd edn. Springer, Berlin Heidelberg New York, pp 271–318
Teunissen PJG (1998b) Minimal detectable biases of GPS data. J Geod 72:236–244
Teunissen PJG, Salzmann MA (1989) A recursive slippage test for use in state-space filtering. Manuscript Geod 14:383–390
Wang J, Chen Y (1994) On the reliability measure of observations. Acta Geod Cartogr Sinica 23:42–51
Acknowledgements
This work was supported by the Austrian Science Fund under research grant J2284-N04. I also thank F.K. Brunner, S. Verhagen, B.H.W. van Gelder, and an anonymous reviewer for their valuable comments on a draft.
Author information
Authors and Affiliations
Corresponding author
Additional information
An erratum to this article can be found at http://dx.doi.org/10.1007/s10291-004-0106-6
Appendices
Appendix A: Resume of statistical reliability testing
This resume largely follows Teunissen (1998a), except for the notation and comments.
Batch processing
Mathematical model
Let the parameter estimation be performed in a linearized Gauß-Markov model of full rank, with the n-vector y of reduced observations, m-vector ξ of parameters, Jacobi matrix A and n-vector e of residuals:
E{...} denotes the expectation operator, and D{...} the dispersion. The best linear unbiased estimate of the parameters, and the predicted residuals are obtained from
with
The diagonal elements r i of R are called redundancy numbers because of
Failure detection
Failure detection within the model Eq. 8 is based on the null hypothesis
and the test statistic
If H0 is true, and if the observations are normally distributed T 0 is χ2-distributed. H0 is therefore rejected on the α confidence level i.e., a failure is assumed, if \( T_{0} > \chi ^{2}_{{n - m,1 - \alpha }} . \) This test is often called global model test (GT). It can be performed if n−m≥1.
The variance covariance matrix (VCM) of the observations has been assumed known above (Eq. 8). If only the cofactor matrix can be established, Eq. (15) cannot be used. In this case either a prior estimate of the variance factor is required for scaling the cofactor matrix—the test can then be based on a Fisher distribution with two finite degrees of freedom—or the GT must be omitted. If the variance factor can only be estimated from data at hand, the GT cannot be performed.
Failure identification
The identification of a specific failure requires that H0 be tested against a specific alternative hypothesis
C models the influence of the q bias terms δ on the n observations. It is assumed that
-
1.
δ is a fixed (be it unknown) bias
-
2.
The observations are normally distributed in both cases (H0 and H a ).
Failure identification is performed within a set of possible scenarios, each of which must be expressed in terms of a matrix C (Eq. 16) and statistically tested against H0. If one of these tests ends in favor of H a , the related failure is identified. Clearly, it is essential that the scenarios tested are the likely ones.
Computationally these local tests rely on estimating the respective bias terms δ as additional parameters and checking for their statistical significance. The test statistic for the k-th scenario is
It is χ2-distributed with q degrees of freedom. A redundancy of at least 2 is required for identification of a single outlier.
Simplified testing
In addition to the above assumptions 1 and 2, it is usually assumed that
-
3.
Only one observation is contaminated in the sense of Eq. (16), and
-
4.
The observations are uncorrelated i.e., Σ is diagonal.
In this case, the contamination is described by a scalar bias, each of the n observations is tested individually for this bias, and the test statistic can be simplified to yield the well-known standardized residuals
Here, the index k denotes the respective observation, e k , σ k and r k the residual, a-priori standard deviation, and redundancy number (see Eq. 13). T k is normally distributed according to
So, H0 is rejected in favor of H ak , if
with a suitably chosen probability α 0 of a type 1 error. Following Baarda’s (1968) suggestion, α 0 is tied to the type 1 error probabilities (α) and the power γ of the GT by the condition
This relation is based on the idea that a specific failure which is detected with the probability γ by the GT should yield a value T k greater than the threshold with the same probability. This is reasonable, once it has been assumed that there is only one outlier. Given the degree of freedom f=n–m (determined by the functional model), and two of α, γ and α 0, the third one must be computed by evaluating Eq. (21). Note that this approach is totally different from the one suggested by Pope (1975) for the case where no GT is performed.
Testing correlated observations
If assumption 4 is invalid i.e., if the observations are correlated (Σ is not diagonal), the above simplifications are not possible. Rather, in this case, the value of
must be used to check for a single outlier in the k-th observation, where η (k) is the k-th canonical unit vector. It is readily verified that Eq. (22) equals Eq. (18) if Σ is diagonal. Still, T k is normally distributed, although with a different noncentrality in the case of H ak
Testing observation differences
When processing differenced observations, assumption 3 does not generally hold, see main text above. In order to revalidate the assumption of a scalar bias term rather than a vector—and thus allow for a simple identification strategy—one must test for failures at the level of the orginal i.e., undifferenced observations (UD). This is indeed possible, by taking the transformation from UD to DD into account, even when processing e.g., double differenced observations (DD)
Here, D is the linear double-difference operator in matrix form. The k-th column d (k) of this matrix shows how the k-th UD contributes to all the DD. Correspondingly, it reveals how a scalar outlier δ k in the k-th UD contaminates all DD:
The test statistic considering each individual UD as a possible outlier is then readily obtained from Eq. (17) as
where n UD is the number of UD involved, and \( {\mathbf{\Sigma }},\;{\mathbf{ \ifmmode\expandafter\tilde\else\expandafter\~\fi{e}}},\;{\text{and}}\;{\mathbf{\Sigma }}_{{{\mathbf{ \ifmmode\expandafter\tilde\else\expandafter\~\fi{e} \ifmmode\expandafter\tilde\else\expandafter\~\fi{e}}}}} \) are the same as above i.e., they refer to the DD model. Again, each of the T k is normally distributed, specifically
Statistical reliability
In the design phase of a network and for assessing the probability of correct failure detection, the statistical reliability is investigated. For each specific scenario formulated in terms of a matrix C k , it addresses (1) the minimum length of δ k (minimum detectable bias) which allows for a certain minimum probability of failure detection, and (2) the effect of an undetected bias on the estimated parameters. For practical purposes, only the latter may be of interest. However, the former is needed for the computations.
Simplified reliability analysis
In the simplified model of uncorrelated observations and individual outliers, the well-known equation for the MDB is
where the noncentrality \( {\sqrt {\lambda _{0} } } \) is computed from α 0 and γ 0 using Eq. (21).
The influence of an outlier on the estimated parameters is proportional to the size of the outlier, see Eq. (9). So, small outliers usually produce small effects. Large outliers produce serious effects but are easily detected. The most dangerous outliers are the ones with values close to MDB, which may just go unnoticed and yet have a significant influence. So, only their effect on the estimated parameters is computed to quantify the external reliability
Note that Eq. (29) produces an error vector for each of the MDB values i.e., for each possible individual outlier. With few observations, these vectors can be plotted (see e.g., Fig. 3) to give a clear idea, of what the effect of undetected failures on the estimated position might be, and which failures are most detrimental.
Reliability analysis of correlated observations
If the observations are correlated, the redundancy numbers are no longer bound to the interval [0, 1] and may even be negative (Wang and Chen, 1994). No change of Eq. 29 is necessary, but obviously, Eq. 28 cannot hold any more. The correct expression is given by (Schaffrin 1997) in a formally equivalent notation to Eq. 28 using normalized redundancy numbers. In a computationally less expensive form it reads:
This equation can also be verified by comparison with Teunissen (1998a; p. 286) considering C≡η (j).
Reliability analysis of observation differences
If we consider the individual failure to occur on the level of the UD, the MDB and the external reliability must be computed from
Appendix B: Kalman filter
The failure detection and identification in the scope of a Kalman filter can be performed formally equivalent to the above procedure. The main difference is that testing is possible at the level of individual epochs (local test) or the level of several epochs (global test), and can be restricted to failures in the observations or also include failures in the predicted states. Here, only local testing of the observations is considered. Furthermore, only the differences with respect to Appendix A are outlined. More information can be found e.g., in Teunissen (1998a) and Koch and Yang (1998).
Mathematical model
The reliability checking is based on the innovations
computed from the reduced observations y t at epoch t, the Jacobi matrix A t , and the predicted state vector \(\tilde{\xi }_{{\user2{t}}} .\) This particular notation is chosen to underline the formal correspondence to the equations in Appendix A.
Failure detection
As above, it is assumed that the innovations are normally distributed. The overall model test for failure detection at the epoch level is performed using
If H0 is true, T 0 is χ2-distributed with n t degrees of freedom, where n t is the number of observations at epoch t.
Failure identification
For failure identification, each of the failure scenarios considered must be expressed as an alternative hypothesis in terms of a matrix C and a bias δ
H0 is tested against H ak using the test statistic
Simplified testing
If the scenario investigated is a single outlier contaminating the k-th uncorrelated observation, the test statistic is a function of the corresponding residual and a-priori standard deviation
Testing correlated observations
For a single outlier among correlated observations, the test statistic is
Testing observation differences
If the scenario investigated is a single outlier contaminating the k-th undifferenced observation, the test statistic is
where n UD,t is the number of undifferenced observations of this epoch, and d (k) is the k-th column of the matrix difference operator (see Eq. 24).
Statistical reliability
Simplified reliability analysis
The MDB and the external reliability of the k-th uncorrelated observation are
where K t is the Kalman gain matrix at epoch t.
Reliability analysis of correlated observations
The minimum detectable bias of the k-th correlated observation is
Its external reliability can be computed using Eq. 46.
Reliability analysis of observation differences
The MDB and external reliability of the k-th undifferenced observation can be computed from
Rights and permissions
About this article
Cite this article
Wieser, A. Reliability checking for GNSS baseline and network processing. GPS Solutions 8, 55–66 (2004). https://doi.org/10.1007/s10291-004-0091-9
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10291-004-0091-9