Anomaly Detection for Satellite Telemetry Series with Prediction Interval Optimization

Anomaly detection is an essential task enabling condition-based maintenance (CBM), fault diagnostics and system prognostics of complex satellite systems. With the advantages of requiring no labels and strong online detecting performance, some prognostics methods have been applied to detect anomalies among satellite telemetry series. Especially, the probability prediction models, e.g., Gaussian Process Regression and Relevance Vector Machine, are more suitable to identify these discords with the ability of providing uncertainty presentation. However, the prediction interval (PI) is derived by the optimization on the prediction-related index which is not constructed for the application of anomaly detection, and its performance will be discounted in this case. Thus, this paper introduces the detection-related index, i.e., coverage-width-based criterion (CWC), to optimize the model parameters which focus on the width of PI, as well as the coverage probability of PI. And the experiments on different types of telemetry series validate the improvement of our proposed framework on PI. Finally, the experiment on actual satellite telemetry series with abnormal points validate the effectiveness and applicability of our proposed method.