Abstract: Real time anomaly detection of spacecraft telemetry data is critical for space mission. Previous methods mostly model the regularly sampled time series data with low missing rates. However, spacecraft telemetry time series data has the characteristics of large dimensions, many noises, high missing rates, irregular sampling intervals, and it is thus more difficult to conduct anomaly detection tasks. An Irregularly sampled Multivariate time series Anomaly Detection (IMAD) algorithm is proposed to model the irregularly sampled multi-dimensional spacecraft time series data with missing values. First, Gated Recurrent Unit with trainable Decays (GRU-D) to model missing values and irregular sampling intervals. Furthermore, the variational autoencoder is used to model the randomness and learn the distribution of normal time series data. Finally, the adaptive threshold determination method based on the extreme value theory is adopted to determine the appropriate threshold for anomaly detection. The experimental results on two real-world spacecraft telemetry datasets show that IMAD outperforms state-of-the-art anomaly detection algorithms. Experiments demonstrate that IMAD is robust to the changes of missing rates, parameters and datasets.

Key words: anomaly detection, time series data, irregular sampling, variational autoencoder, GRU-D