An Online Machine Learning Paradigm for Spacecraft Fault Detection

Anomaly and threat detection is critical in space systems and the aerospace industry as missions become more complex requiring higher levels of autonomy. Threat Detection, Isolation, and Recovery (TDIR) involves the identification and prognostics relating to threats in the given system; these threats include subsystem failures, external environment disturbances, and cybersecurity attacks, all capable of endangering the normal operations of the spacecraft. Often, these TDIR architectures for space systems are implemented on the ground analyzing telemetry data due to their computationally expensive designs and limited spacecraft processing bandwidth. This decoupling of TDIR and spacecraft operations presents a gap in autonomy where recovery operations are delayed and nominal operations are disrupted until the ground segment can identify the source of the problem. This paper presents an Incremental Online Machine Learning (IOML) method to address threat detection. This IOML utilizes Support Vector Regression (SVR), a supervised machine learning technique with proven excellent learning and prediction capabilities. Principle Component Analysis (PCA) is used to select the needed features providing a more robust representation of system performance and also reduces training and computation requirements through data dimensional reduction. This machine learning paradigm is applied in simulation to a generic spacecraft power system model to evaluate the fault detection rate and accuracy and its ability to adapt in real time to new system data and performance. The proposed IOML architecture is capable of accurately detecting system faults using dual Incremental Online Support Vector Machines (IO-SVM).