Machine Learning Algorithms for Orbit Region Classification: A Case-Study from Cassini

Abstract:

Significant effort is currently expended by science teams to manually inspect data for short-term “anomalous” science events or classifying magnetospheric regimes along the spacecraft’s trajectory, such as when a spacecraft is encountering the solar wind versus within a planetary magnetosphere. While much of this effort is currently done by scientists in post-processing, recent advancements in machine learning (ML) could allow for such operations to be automated or even ported to the spacecraft flight software. Utilizing ML to automate data interpretation on-board the spacecraft could allow for science operations to be completed autonomously, such as payload queuing based on the spacecraft’s environment or prioritization of data to be downlinked. With such optimization on-board, the science return from a mission could be dramatically improved, particularly for missions that are telemetry-constrained. Here we present one such case study for ML implementation – orbit region identification (magnetosphere versus magnetosheath versus solar wind) around Saturn for the Cassini mission. Using a list of bow shock and magnetopause crossings for the Cassini mission as our training labels, we explore the classification capabilities of various ML algorithms and data sets. Data sets included combinations of magnetic field and energetic charged particle data from the Charge Mass Spectrometer (CHEMS) and Low-Energy Magnetospheric Measurement System (LEMMS) instruments. Across the various machine learning methodologies surveyed, we found the best classification performance from recurrent neural networks utilizing solely magnetometer data, with a maximum accuracy of ~92% on an unseen test set. Simpler ML approaches, such as Gaussian Mixture Models (GMMs), were able to approach a similar accuracy with far less training data when magnetometer data was supplemented with CHEMS and LEMMS data.