Deep Metric Learning for Abnormal Rotation Detection of Satellites from Irregularly Sampled Light Curve

Abstract

In recent years, satellites have become an indispensable infrastructure in our lives. The number of satellites is increasing yearly and becoming increasingly active. To use satellites safely, it is crucial to manage them and detect the anomaly of satellites as much as possible. However, it currently takes skilled operators to detect an anomaly, and it is difficult for even skilled operators to detect the anomaly early without the telemetry data in cases such as an abnormal rotation. To address these challenges, we tested the feasibility of using deep metric learning for early anomaly detection from the irregularly sampled light curve. One of the characteristics of a light curve is unequally spaced because the optical sensor on the ground can only observe the subject at night and not when the weather is terrible. Given an irregularly sampled light curve, our model employs a long short-term memory (LSTM) unit of encoding the temporal dynamics and learns the embedding on the feature space using triplet loss. Then, an anomaly score is calculated based on pairwise distances between segments from the learned embedding in the feature space. With actual data from the satellite being operated, we showed the effectiveness of our model and the feasibility of early anomaly detection. Also, by exploring learned embedding in the feature space, we show that our model could capture the continuous state of the satellite.