Time Series Classification for Satellite Fault Detection

Satellite fault detection is critical for ensuring the reliability, safety, and overall success of satellite missions. Satellite systems that include an autonomous fault detection system could reduce the reliance on continuous ground-based monitoring, allowing for quick and efficient assessment of telemetry without expensive manual intervention --- thereby increasing operational efficiency. Multiple time series classification (TSC) algorithms are evaluated for pre-processing satellite data, facilitating an efficient integration with a downstream probabilistic fault detection engine. These algorithms are tested on simulated satellite data generated using the Basilisk satellite simulator, with models individually trained to detect specific fault types (thruster leaks or blockages, solar panel deployment failure, etc.). The TSC algorithms are evaluated based on their computational time, accuracy, and ability to classify the data into different fault categories. The results of this study show the MiniROCKET transformation, with a multi layer perceptron back end, to be the most effective balance between computational efficiency and classification performance across all datasets.