A Multivariate Anomaly Detector for Satellite Telemetry Data Using Temporal Attention-Based LSTM Autoencoder

Telemetry anomaly detection is a prominent health condition monitoring task that plays an increasingly crucial role in identifying unexpected events and improving satellite’s overall reliability. Since a component or subsystem of a satellite contains multiple telemetry parameters, it is essential and practical to develop a multivariate telemetry anomaly detection framework. However, the complex nonlinear correlations among different telemetry parameters and the temporal dependency hidden in each telemetry parameter pose significant challenges. In this study, a temporal-attention-based long short-term memory autoencoder (TA-LSTM-AE) anomaly detector is proposed for detecting multivariate telemetry anomalies. Initially, a TA-LSTM-AE model is established to learn the latent representation of the correlations among the monitored telemetry parameters and the temporal dependency within each parameter. A temporal attention mechanism is applied to enhance the long-term temporal dependency modeling ability of the model. Subsequently, the Mahalanobis distance of the reconstruction error is defined as the anomaly score. An adaptive thresholding approach is specifically designed considering the variation in the learned latent representation by balancing the tradeoff between the detection reliability and accuracy. A critical parameters identification strategy is also presented for recognizing the telemetry parameters that contribute significantly to the triggered multivariate anomalies. Finally, the proposed detector is applied to detect anomalies in multivariate telemetry data collected from a real-world satellite. The experimental results verify the capability of the proposed detector in multivariate telemetry anomaly detection.