Machine Learning Applied to Control Moment Gyroscope Fault Diagnosis
A novel applicationof data-driven machine learning methods for fault isolation of nonlinear systems with a case-study for an in-orbit closed-loop controlled satellite with control moment gyros is presented. High-fidelity models of thesatellite are developed to provide an abundance of data for both healthy and various faulty conditions. This data is then used as input for the proposed data-driven fault isolation. Statistical features such as the maximum, minimum and number of peaks are calculated to describe the time series. Then, the most meaningful features are extracted using recursive feature elimination. The fault isolation employs a random forest classifier to identify the fault scenario, and a multioutput regressor with gradient boosting to estimate the actuator’s effectiveness. Results of the classified faulty condition are then reported using 5-fold cross-validation. A comprehensive comparison of the performance for different ensemble architecture as well as sensitivity analysis for the hyperparameters of each model is provided.
Keywords - Attitude determination and control system (ADCS), fault detection and isolation (FDI), spacecraft, satellite, ensemble machine learning, control moment gyro.