Machine Learning Based Fault Detection and Diagnosis for the Actuators of a Satellite

The purpose of this work is to perform fault detection and diagnosis regarding the reaction wheels of a satellite orbiting around the Earth using machine learning different algorithms. The difficulties with respect to the application of machine learning in other fields are given by the complexity of the dynamics of the satellite, that is generally non-linear and timevarying. Also, the presence of perturbations of many kind, as well as of the change of maneuvers, provokes a rapid decrease of the accuracy of the predictions made by the model, which lead to results that are not adequate to the requirements of reliability of the FDI system imposed by the aerospace sector. For this reason, the objective of this thesis is to obtain the highest value possible of the accuracy in case of change of maneuvers, by using different types of machine learning models, and to test the robustness of the algorithms in presence of disturbances that may lead to small variations in the dynamics of the satellite. The algorithms applied in this work include support vector machine, naive-bayes, and random forest, as well as the application of different techniques of feature selection.