Addressing magnetic interference in small spacecraft with machine learning

Magnetic fields are ubiquitous in spaceand their structure and dynamics govern the motion of plasmas and their interactions throughout the heliosphere . For this reason, having accurate measurements of magnetic fields in space provided by sensitive magnetometers is key to understanding any space environment. Magnetic fields are also produced by spacecraft subsystems, including other instruments, attitude control systems , power systems and solar panels. These fields contribute a disturbance, or noise, to the data provided by magnetometers, thus affecting the quality of the measurements. Traditionallymagnetometers have been placed o n booms with sufficient length to reduce the magnitude of these disturbances to a level that, combined with the sensitivity of the particular magnetometer being used, allowfor the signals of interest to be detected. This solution, although functional, adds complexities to the design and build of a spacecraft, as well as complex mechanical systems that could fail during deployment. The growth and popularity of CubeSats and small satellites makes it possible to simplify the process of designing, building, integrating and launching powerful instruments to study different phenomena. In this context, the added costs and design challenges of a long boom mean that the inclusion of science-grade magnetometers can be compromised. Recent developments in machine learning allowthe processing of complex streams of data in on-board computers, opening a new world of possibilities in terms of data analysis. By using a novel machine learning method of simple regret minimization for contextual bandits, data returned by multiple distributed magnetometers inside a small spacecraft are used to identify the magnetometer that returns the cleanest data based on information regarding the status of relevant subsystems at any given time. The algorithm works by a combination of an exploration (when a learning process takes place based on a known external magnetic field) and an exploitation (when the field is determined by the algorithm itself) phases. A validation of the algorithm in a lab setting is presented, as well as using telemetry data from the GEO-CAPE ROIC In-Flight Performance Experiment ( GRIFEX) and the Tande mBeacon Ex periment ( TBEx) CubeSats.