Autonomous Fault Detection and Prognosis for Spacecraft Reaction Wheels Using Sparse Identification

This paper presents a physics-based data-driven approach for predicting the remaining useful life (RUL) of a spacecraft reaction wheel (RW). Our method combines a physics-based model with a data-driven regression and machine learning technique known as the Sparse Identification of Nonlinear Dynamics Systems with Control Input (SINDYc). This approach is used for fault detection and RUL prediction of the RW. For fault detection, we predict the states and health index (HI) parameters of the RW, with the coefficients of output torque and viscous friction selected as the HI parameters. To estimate the RUL, we analyze the trends of these HI parameters over time, predicting when the failure threshold will be crossed. We demonstrate that the proposed method is more effective and suitable for autonomous onboard applications compared to existing methods, such as long short-term memory recurrent neural networks. SINDYc also provides explicit identification of the underlying RW governing equations while requiring orders of magnitude less computational effort to produce a model. We establish the robustness of this methodology by confirming only slight degradation in model accuracy when fitting to noisy data.