In order to better support the Telemetry monitoring of both current and future missions at EUMETSAT, flight control teams have developed semi-supervised Outlier Detection algorithms which compliment traditional monitoring techniques such as Out of Limit Monitoring. This Outlier Detection has now been in use for several years. Through the process of developing and utilising the Outlier Detection algorithms, it has been found that the main challenges are not related to the choice, tuning or development of the algorithm. Instead, they are related to the more practical aspects surrounding the presentation of results, choice of parameters to be monitored, filtering of results and the labelling of Nominal Data. Moreover, these practical problems are largely algorithm independent and so once an adequate algorithm was developed, it was realised that chasing a better algorithm before these practical had been addressed would simply lead to a delay in the use of this technique. This paper discusses the Outlier Detection algorithm and its development, but the main focus is on the lessons learned from this process which led to the identification of practical problems and solutions to them. In particular, the Outlier Detection results are presented in a way which allows users to understand why outliers have been flagged, which minimises the “Black Box” effect. This, however, implies limiting the number of dimensions used when performing Outlier Detection. Another issue observed was floods of outliers being reported during non-nominal operations or anomalies. Not only does this mask real issues, but being since Outlier Detection is a relatively new technique, it also serves to erode trust in the process. This has mainly been resolved by a conditional Outlier Detection method which has been developed by categorising parameters hierarchically as Influencers and Followers, and also by careful selection of parameters. Finally, the main sticking point for any semi-supervised Outlier Detection techniques is keeping up to date with the labelling of Nominal Data. Again, this is partially addressed through careful selection of parameters in order to avoid having too many datasets to maintain, however it is the one area where development is still on-going within EUMETSAT, and so this paper describes the future plans to facilitate the labelling of Nominal Data.
From Theory to Practice: Operational Implementation of Telemetry Outlier Detection at EUMETSAT
Springer Aerospace Techn.
2022-03-16
25 pages
Article/Chapter (Book)
Electronic Resource
English
EUMETSAT Multi-Mission Administrative Message Goes Operational
British Library Conference Proceedings | 2012
|FIRMS & INSTITUTIONS - EUMETSAT
Online Contents | 2000
AUTOMATED COMMANDING FOR EUMETSAT
Online Contents | 1995
|