Spacecraft Multivariate Time Series Anomaly Detection in the Presence of Non-Anomalous Spikes

Detecting anomalies in spacecraft telemetry is important due to the extreme operational environment of these systems. Detecting anomalies can serve as early warnings for potential system failures. Typically, a spacecraft system contains hundreds (or thousands) of telemetry channels. There is a need for automated anomaly detection to enhance efficiency and accuracy. Deep learning reconstruction-based anomaly detection methods for multivariate time series have been extensively studied. In these methods, a network is used for time series reconstruction and anomalies are detected based on the error sequence between the reconstructed time series and the original time series. However, many approaches cannot accurately reconstruct abrupt changes, such as spikes, which may result in false or missed detections. In this paper, we examine reconstruction errors for anomaly detection. A Long Short- Term Memory (LSTM) autoencoder network is used for multivariate time series reconstruction. The error sequence between the reconstructed time series and the original time series is obtained. Then, the period of the error sequence is estimated to remove reconstruction errors caused by abrupt changes in the time series. A dynamic threshold is used to detect the anomalies in the error sequence. We examine the performance of the proposed method by using a synthetic dataset. The dataset is characterized by having non-anomalous spikes, large periods and correlation among channels.