Anomaly Detection for Multivariate Telemetry Series of Satellite with Improved HVAE

Spacecrafts, as import devices for exploration of outer space, whose performance and health status is mainly reflected by telemetry data. Effective anomaly detection for telemetry data can improve the operational safety and reliability of spacecraft in orbit. While the complex temporal characteristics and spatial correlations of multivariate telemetry sequences bring great challenges for typical methods. In this work, an anomaly detection method based on imporved Hierarchical Variational Auto-Encoder (HAVE) is proposed. Where the HVAE with two random hidden variables is employed to model the normal patterns of multivariate telemetry sequences, each of which learns temporal correlation and variables correlation respectively. Moreover, the Self-Attention mechanism based on the gated recurrent unit neural network is integrated to augment the model's ability to handle long-term series dependencies between telemetry data and control commands. The experiments are performed on public telemetry datasets of SMAP and MSL from NASA spacecrafts. These experimental results verify its effectiveness and applicability compared with other related methods.