DATA-DRIVEN ANOMALY DETECTION FOR RESIDENT SPACE OBJECTS USING AUTOENCODER WITH BINARY CLASSIFICATION

To detect anomaly for RSOs accurately and timely is critical to protect the long-term sustainability of space activities, including Space Situational Awareness (SSA) and Space Traffic Management (STM). In this paper, we explore a new data-driven framework based on deep autoencoder for RSOs’ anomaly detection. A novel two-input autoencoder model is proposed to identify whether the tracks belong to the same orbit or not. An in-house simulation-based space catalog environment is used for experiments and analysis. We compare the proposed model with Principal Component Analysis (PCA) method in classification and the results show that the proposed method achieves higher accuracy in identifying whether two tracks are from the same orbits or not than the classical PCA method. Furthermore, the proposed method is also robust to noise data with high accuracy.