Practical Aspects of Anomaly Detection Algorithms in Satellite Operations
Grzegorz Adamski, Julian Spencer-Jones, Gilles Kbidy · 2018 SpaceOps Conference · 2018
Telenor Satellite Broadcasting and L3 T&RF have worked on a joint project to apply big data technologies and modern machine learning/AI tools and analyze data from the existing satellite fleet.One of the primary goals of monitoring spacecraft telemetry is to detect any anomalous vehicle behavior, which in turn allows the operators/engineering team to employ mitigation strategies.The issue is that modern satellites downlink hundreds or even thousands of data points per second to report their state of health and status of on-board components, while the number of different measurands they report on is frequently measured in tens of thousands.Classical anomaly detection, with pre-defined nominal/warning/critical ranges allow the detection of the most obvious situations, however the sheer size of the databases operators deal with make a very fine grained definition impractical.On top of that, simply looking at value ranges might not give the operator a full view of the satellite behavior.Using a data driven approach, with a machine learning-based outlier detection mechanism will likely provide a better detection of out of family situations.Considering however that these algorithms are largely unaware of the underlying physical behavior, they may be prone to either a large degree of false positives or false negatives.Considering that users exposed to large number of false alarms will most likely get de-sensitized to them, it's imperative that the number of incorrect reports are minimized.At the same time, pushing the reporting threshold too low may lead to the operators missing significant spacecraft events, which may be detrimental to the vehicle health.A joint Telenor/L-3 Engineering group will work using data collected from Telenor satellite fleet using InControl Nebula™ archiver to define and train anomaly detection models.Considering the need to define proper model thresholds that will correctly balance the need to predict incorrect behavior and at the same time minimize the number of false positives, we will attempt to use data from other sources to fine tune models to allow them to be better suited for practical operational use. I.