Tommy Fryer, CGI; Robert Arthur, CGI; Paul Walker, CGI
Keywords: Kalman, Maneuver, operational
Abstract:
In order to monitor the health of our Kalman Filter, which may be affected by anomalous observations, changes in motion or poorly tuned process noise, we define a statistic that is able to report adverse conditions. In this paper we assess techniques for detecting manoeuvres and discriminating them from observation anomalies using this metric.
We assess the performance of three classes of statistical tests to detect deviation from the null hypothesis which is that nothing has changed from a baseline: Cramer-von-Mises Chi-Squared test, two-sample tests against an empirical distribution function and testing the sample mean and variance against baseline empirical values. For synthetic models with realistic process noise, we find that impulse manoeuvres of a typical size can be detected a short time after they occur, we propose to distinguish them from anomalies by rejecting anomalous Kalman Updates and accruing evidence of filter anomaly from more than one observer. For impulses on an otherwise inert object, we find that the Cramer-von-Mises Chi-Squared test performs best, requires the least computation, requires no baseline data and is as robust to excess process noise as the empirical distribution function tests.
For small continuous thrusts where it is not possible to identify a change in the anomaly metric from the baseline thrust in a single pass, we are able to identify the change by accruing the statistics over multiple passes. The most promising test was the Bootstrap sample variance test which was most robust to sub-optimal process noise and able to detect changes even when baselined against a continuously thrusting object. The Empirical Distribution function tests performed less well but were much improved over the Chi-Squared tests when process noise was not tuned well or when the baseline period includes thrust.
Date of Conference: September 19-22, 2023
Track: Astrodynamics