Telemetry series, generally acquired from sensors, are the only basis for the ground management system to judge the working performance and health status of orbiting spacecraft. In particular, anomalies within telemetry can reflect sensor failure, transmission errors, and the major faults of the related subsystem. Therefore, anomaly detection for telemetry series has drawn great attention from the aerospace area, where probability prediction methods, e.g., Gaussian process regression and relevance vector machine, have an inherent advantage for anomaly detection in time series with uncertainty presentation. However, labelling a single point with probability prediction faces many isolated false alarms, as well as a lower detection rate for collective anomalies that significantly limits its practical application. Simple sliding window fusion can decrease the false positives, but the support number of anomalies within the sliding window is difficult to set effectively for different series. Therefore, in this work, fused with the probability prediction-based method, the Markov chain is designed to compute the support probability of each testing series to realize the improvement on collective anomaly mode. The experiments on simulated data sets and the actual telemetry series validated the effectiveness and applicability of our proposed method.
Keywords: Gaussian process regression; Markov chain; collective anomalies; false positive; probability prediction; relevance vector machine; telemetry series.