UNSUPERVISED CLUSTERING OF ANOMALOUS SAMPLES OF TIME-SERIES DATA USING ANOMALY-DETECTION-CAPABLE STATISTICAL FEATURES

Abstract

Dependable data quality in satellite telemetry is critical for the reliability of space missions. CubeSats such as the University of Kansas’ KUbeSat-1 encounter various anomalies from harsh space environments and hardware degradation. Traditional threshold-based anomaly detection models fail to differentiate error sources within telemetry data. Despite existing methods, there remains a need for a machine learning based model that systematically categorizes telemetry anomalies. This study proposes an unsupervised learning approach that leverages historical telemetry data to recognize characteristic anomaly signatures. The model clusters segments into distinct behavioral groups using a KMeans algorithm, which are assessed through dimensionally reduced visualizations and silhouette scores. This pipeline enhances CubeSat reliability by supporting scalable, telemetry monitoring frameworks that enable smarter design choices and resource allocation for future small satellite missions.