Dipole Fitting in Unit-Level Spacecraft Equipment with Deep Neural Networks

Magnetic Dipole Modeling (MDM) is a well- established method for fitting the spatial magnetic signature generated by magnetic sources into a model comprised of magnetic dipoles. In this paper, we study the dipole fitting problem using Machine Learning (ML) techniques. Simulated data were used to estimate the dipole parameters that can accurately reconstruct the measured magnetic field. We applied our methodology in two widely used measurement facilities, namely the Magnetic Coil Facility (MCF) and the Multi Magnetometer Facility (MMF). To solve this regression problem, Artificial Neural Networks (ANNs) were considered. Simulation results showed that ANNs significantly outperform other benchmark ML methods in terms of the achieved accuracy of the estimated dipole parameters. Finally, we observed that MMF model shows enhanced field reconstruction accuracy with reduced position estimation loss (~70% lower position error), magnetic moment estimation loss (~55% lower moment error) and model complexity (33.3% lower complexity) compared to the MCF model.