Spacecraft Anomaly Detection and Relation Visualization via Masked Time Series Modeling

Anomaly detection (AD) refers to find patterns in time series data that do not behave expectedly. Current state-of-the-art anomaly detection method, based on reconstruction error generated by LSTM sequence modeling. Recently, the remarkable improvement achieved by BERT model in language translation demonstrated that transformer is superior to LSTM models, due to its extracting relations ignoring distance. In this paper, we propose a transformer-based architecture, Masked Time Series Modeling, modeling data stream. We compared the performances of our method with state-of-the-art AD methods on challenging public NASA telemetry dataset. The experiment results demonstrated our method saves about 80% time cost because of parallel computing compared with LSTM methods and achieves 0.78 FI point-based score. Moreover, we visualize anthropogenic anomalies through attention score matrix.