An LSTM Autoencoder-Based Framework for Satellite Telemetry Anomaly Detection

This paper proposes an approach for detecting and identifying anomalies in satellite telemetry based on multi-dimensional time series analysis. Firstly, the long short-term memory-based autoencoder (LSTM-AE) model is established to reconstruct the normal operating data by simultaneously capturing the nonlinear spatial dependency among different telemetry variables and the temporal dependency in each telemetry variable. Then, the anomaly score is derived from the reconstruction residual based on the Mahalanobis distance. An adaptive threshold estimation method based on random forest regression algorithm is developed to identify anomalous telemetry data samples. The effectiveness of the proposed method is verified by a case study using real-world satellite monitoring telemetry variables.