Diagnosing Mismodeling Errors in Orbit Determination Using Autoencoders

Flawed orbit determination models or anomalous spacecraft activity can degrade the accuracy of a spacecraft's orbit estimate with potentially catastrophic consequences. Cases of mismodeling are often detectable through patterns they elicit in tracking data residuals. These signatures currently require expert investigation to diagnose a precise cause such as an unknown acceleration on the spacecraft. We simulate a dataset of orbit determination mismodeling errors for the Sample Retrieval Lander and develop an autoencoder neural network that clusters anomalous range and Doppler residuals in a latent vector space. The clusters permit both the detection and diagnosis of mismodeling errors. Reconstructions of the residuals from latent vectors also allow known types of anomalies to be distinguished from novel cases. We describe the architecture and features of the model and apply it retroactively to residuals from the InSight mission to diagnose a missing station location correction.