Human and Machine Diagnosis of Orbit Determination Errors From Filter Arrays

Deep space orbit determination requires careful modeling of spacecraft and observation dynamics in order to fit radio tracking measurements. Unexpected spacecraft events, corrupted tracking data, or mis-modeling of the dynamics can result in an incorrect orbit estimate with potentially catastrophic consequences. The integrity of an orbit estimate is typically verified using an array of different filter set-ups. If the solutions of two or more filters become statistically inconsistent, this can indicate a flawed model and an inaccurate baseline solution. Since each filter is sensitive to different errors or anomalies, the pattern of filter comparisons provides a fingerprint with which to diagnosis the error. By first simulating the response of a filter array to a catalogue of possible orbit determination mis-modeling errors for the Sample Retrieval Lander, we train a Long short-term memory (LSTM) neural network to perform this diagnosis autonomously. We also develop a second autoencoder neural network for cataloging the unique filter comparison fingerprint of each mis-modeling scenario. We show that the latent space of the fingerprints provides insight into and verification of the LSTM’s decisions, as well as helps develop intuition for the patterns that can be used to train human navigators, prepare operational readiness tests, and design orbit determination filters.