Session
Technical Session 1: Mission Operations and Autonomy
Location
Utah State University, Logan, UT
Abstract
Despite its recent disruption to many engineering fields, incorporating machine learning into satellite operations is difficult due to the associated performance restrictions with the relatively small amount of available data during a satellite’s operating lifetime; a satellite needs to be operating for several months to years in order to grow a dataset able to be used in machine learning. That is the reason behind its historically limited application to active satellite operations, whereas a posteriori applications to past missions can be found in the literature. Recently however, the eruption of mega-constellations of satellites, whose hardware, software, and orbit are almost identical, presents the opportunity to utilize a much larger dataset, even if the lifespan of the individual satellite is limited. This is the case of the Planet Scope constellation, which consists of hundreds of “Dove” satellites that are regularly replenished with new launches. The large number of similar satellites, combined with an automated and thorough metric retrieval system, enables the generation of large matrices of data that can be used as machine learning features, aggregated daily from telemetry, ephemeris, software logs and configuration. On top of these features, detailed operator knowledge of the system is included in the dataset, in the form of human-provided labels that are historically recorded and used for training purposes. Through the use of this novel dataset, this paper presents an implementation of machine learning to autonomously detect anomalies on active satellites. This early anomaly detection system is currently used in operation to identify satellites that require further operator intervention, as well as to support the diagnosis process.
Machine Learning for Early Satellite Anomaly Detection
Utah State University, Logan, UT
Despite its recent disruption to many engineering fields, incorporating machine learning into satellite operations is difficult due to the associated performance restrictions with the relatively small amount of available data during a satellite’s operating lifetime; a satellite needs to be operating for several months to years in order to grow a dataset able to be used in machine learning. That is the reason behind its historically limited application to active satellite operations, whereas a posteriori applications to past missions can be found in the literature. Recently however, the eruption of mega-constellations of satellites, whose hardware, software, and orbit are almost identical, presents the opportunity to utilize a much larger dataset, even if the lifespan of the individual satellite is limited. This is the case of the Planet Scope constellation, which consists of hundreds of “Dove” satellites that are regularly replenished with new launches. The large number of similar satellites, combined with an automated and thorough metric retrieval system, enables the generation of large matrices of data that can be used as machine learning features, aggregated daily from telemetry, ephemeris, software logs and configuration. On top of these features, detailed operator knowledge of the system is included in the dataset, in the form of human-provided labels that are historically recorded and used for training purposes. Through the use of this novel dataset, this paper presents an implementation of machine learning to autonomously detect anomalies on active satellites. This early anomaly detection system is currently used in operation to identify satellites that require further operator intervention, as well as to support the diagnosis process.
