Discrete Features Enhancement Based Online Anomaly Detection for Satellite Telemetry Series

Telemetry data, collected from different onboard sensors and transmitted via a telemetry link, are the only basis for monitoring the health status and failure of on-orbit spacecraft. The capability of anomaly detection for telemetry series should be further enhanced in ground stations. Owing to the advantages of temporal modeling, dynamic threshold generation, online application, and strong interpretability, probabilistic prediction methods have been designed to realize anomaly detection of telemetry data; however, they face the risks of false alarms for isolated normal points and missing alarms for continuous anomalies. In this case, an improved method is proposed to promote the detection ability of probability prediction methods for collective anomalies in the telemetry data. First, effective dynamic thresholds are derived by the modified probabilistic prediction model with better prediction confidence levels. Then, we design a discrete method based on equal-width discretization and statistical analysis. In detail, each prediction error is divided into several intervals corresponding to different abnormal levels. In this way, the quantitative characterization ability for samples is enhanced. Finally, based on Markov chain and majority voting integration, discrete feature enhancement at multitime scales referring to multistep and multiwindow is implemented to realize temporal modeling of multistep prediction features. The experiments on simulated datasets and real telemetry data verify its effectiveness and applicability compared to other methods of labeling anomalies.