Jun Yoshida, NEC Corporation; Ryosuke Nakazawa, NEC; Naoki Yoshinaga, NEC Corporation; Taichiro Sano, NEC; Katsuaki Taya, NEC Corporation; Masatoshi Ebara, NEC Corporation; Ryosuke Togawa, NEC Corporation;
Keywords: Machine Learning, Deep Metric Learning
Abstract:
Recently, there has been increased emphasis on satellites used for aspect of both our life and security. NEC operates remote sensing satellite ASNARO2, which was developed and operated in-house. On the other hand, an increase in the number of space objects (e.g. satellite constellation or debris), anti-satellite weapon (ASAT), the satellites with the purpose of rendezvous, and such objects have led to the increasing risk of a collision and an attack when outer space is utilized. There are a considerable number of several conjunctions in operational satellites; dozen conjunctions which is received from Combined Space Operations Center (CSpOC) as a Conjunction Data Messages (CDM) have occurred in ASNARO2 over a year. The recognition of such objects is essential in dealing with a risk in terms of security, therefore it is necessary to identify such objects from a large amount of space objects as soon as possible. Currently, NEC has managed orbits of space objects using SSA Software Suite (SSS) system developed by company COMSPOC, which is based on observation data of some optical telescopes and radars. However, due to the lack of sensor and human resources associated with increased space objects, it is difficult to monitor a large amount of them. Therefore new tools and frameworks are needed to solve these operational problems to improve the capability of Space Situational Awareness (SSA) and Space Domain Awareness (SDA).
This study set out to asset the notable objects detection, which is the dynamic object differs from its normal orbit for some reason, such as a malfunction or collision, based on two line element (TLE) of Geostationary Orbit (GEO), which is published in SpaceTrack.org. This study tested whether it could identify the difference between normal and notable objects from TLE. To date, there has been no detailed investigation of that feasibility. For this assessment, this study leverage the model NEC developed, which model obtains embedding of similarity between data points considering temporal information. This model is composed of the combination of time series and metric learning using Deep Neural Network (DNN). Metric Learning has been successfully used in various field, including object identification, natural language processing, speech recognition and medical diagnosis. Metric Learning learns data similarity on a distance basis and can perform similarity discriminations on unknown data. Furthermore, metric learning has achieved remarkable results in recent years when combined with DNN. This model learns mapping from TLEs to embedding in feature space, where similar TLEs are located nearby and dissimilar ones is far away, based on the unsupervised manner. To create the notable objects detection model in unsupervised manner, the model is trained on normal object TLEs that has been confirmed by expert and labeled using pseudo-labeling based on a distance. In experiments, by comparing pairwise distances between TLEs embedding of normal and notable objects, the model succeeded in detecting differences between normal and notable objects from TLEs. While this study confirms few GEO objects, it did partially substantiate the effectiveness of the model NEC developed and the feasibility of the notable objects detection based on TLE. These findings could be used to help SSA/SDA operation. Further research could also be conducted to determine the effectiveness and feasibility of Low Earth Orbit (LEO) objects.
Date of Conference: September 19-22, 2023
Track: Machine Learning for SDA Applications