Applying Deep Learning to Anomaly Detection of Russian Satellite Activity for Indications Prior to Military Activity

Abstract

We apply deep learning techniques for anomaly detection to analyze activity of Russian-owned resident space objects (RSO) prior to the Ukraine invasion and assess the results for any findings that can be used as indications and warnings (I&W) of aggressive military behavior for future conflicts. Through analysis of anomalous activity, an understanding of possible tactics and procedures can be established to assess the existence of statistically significant changes in Russian RSO pattern of life/pattern of behavior (PoL/PoB) using publicly available two-line element (TLE) data. This research looks at statistical and deep learning approaches to assess anomalous activity. The deep learning methods assessed are isolation forest (IF), traditional autoencoder (AE), variational autoencoder (VAE), Kolmogorov Arnold Network (KAN), and a novel anchor-loss based autoencoder (Anchor AE). Each model is used to establish a baseline of on-orbit activity based on a five-year data sample. The primary investigation period focuses on the six months leading up to the invasion date of February 24, 2022. Additional analysis looks at RSO activity during an active combat period by sampling TLE data after the invasion date. The deep learning autoencoder models identify anomalies based on reconstruction errors that surpass a threshold sigma. To capture the nuance and unique characteristics of each RSO an individual model was trained for each observed space object. The research made an effort to prioritize explainability and interpretability of the model results thus each observation was assessed for anomalous behavior of the individual six orbital elements versus analyzing the input data as a single monolithic observation. The results demonstrate not only statistically significant anomalies of Russian RSO activity but also details anomalous findings to the individual orbital element.