The United States Space Force needs a technical solution for rapidly and effectively identifying the cause of detected satellite anomalies. Recognizing that the satellite fault identification problem is inherently a Time Series Classification (TSC) problem that attempts to map time series telemetry data to specific onboard fault classes, we will conduct a comprehensive comparative assessment of a representative, but exhaustive, selection of TSC algorithms. Algorithms we will consider range from traditional rule-based frameworks to formal methods, to modern analytical techniques from Artificial Intelligence and Machine Learning (AI/ML). Selected algorithms will also cover different approaches to TSC ranging from: single- vs. multi-dimensional, to model-based vs. model-free, to supervised vs. unsupervised TSC approaches. In order to effectively select, design, and assess TSC algorithms for the satellite fault identification problem, we will also identify a taxonomy of onboard satellite faults that fully capture the key characteristics of such faults and key performance evaluation metrics and criteria for the TSC algorithms considered in this effort. Based on the comparative study, a subset of TSC algorithms will be down-selected for further evaluation in a limited scope, but non-ambiguous, spacecraft Attitude Determination and Control System that exhibits a set of possible faults that range from mechanical, electrical, sensor, software and cyber, to those induced by weather (e.g., single event upsets). Finally, we will also test the down-selected algorithms against historical flight telemetry data that will be delivered as Government Furnished Information.
