Anomaly Detection in Spacecraft Telemetry Using Vision Transformer and Caps-LSTM

The successful completion of space missions increasingly relies on monitoring spacecraft health, given their ambitious and intricate nature. Assessment of spacecraft telemetry heavily depends on anomaly detection, which enables engineers to rapidly identify irregular and unexpected behavior in spacecraft data. Conventional anomaly detection models often struggle to handle the time-dependent, non-stationary, non-linear, high-dimensional, and complex nature of spacecraft telemetry data. Hence, this work presents a novel deep learning model for anomaly identification in spacecraft telemetry. Input data is collected from public repositories including the Mars Science Laboratory (MSL) rover and Soil Moisture Active Passive (SMAP) satellite. Data cleansing is performed using linear interpolation for missing data handling and noise reduction, followed by feature extraction using Dynamic Time Warping. The Vision Transformer and Capsule Network-based Long Short-Term Memory (VT and Caps-LSTM) model is proposed for anomaly detection in spacecraft telemetry. Implemented in Python, the proposed work demonstrates superior efficiency compared to conventional approaches, achieving 97 % accuracy, 96 % precision, and a 98 % F1-score at the 100th epoch.