Utilizing Machine Learning Methods for Classifying Telemetry of Human Spaceflight Systems

Kurzfassung

All past and present crewed outposts in space, most prominently the International Space Station (ISS), rely on one particular mode of operation: real-time operation ('ops' for short). Here, the vehicle, and it's subsystems, are almost entirely monitored and controlled from ground. The astronauts on-board instead focus their time and effort on achieving the scientific mission objectives. However, within the next decade, almost all international space agencies have defined goals to achieve human presence around and on the Moon, as well as prepare for deep space missions beyond the Moon, and towards Mars. For such missions, it is obvious that, with increasing distance from earth, the signal delay between a vehicle and a ground station also increases. As a consequence, beyond a certain distance, a ground controller is no longer capable to monitor and control the vehicle in real time. Therefore, a new operational concept is needed, shifting the responsibility of evaluating and classifying real time data of the on-board systems. Shifting the responsibility to crew is not feasible, since available crew size and crew time are limited resources on-board a crewed spacecraft. So, in order to enable the crew to still achieve the mission objectives, without significantly increasing their duties or size, a shift is needed to an intelligent on-board system instead. This paper will examine how this can be achieved by analyzing the data that is generated on-board the Columbus Module of the ISS, as well as introducing different Machine Learning (ML) methods used for classification. For that purpose, we first highlight how time series data is generated on-board, how it's arriving at the ground controllers' console, as well as what properties the data has. Afterwards, we focus on the different ML methods capable of dealing with the given data. To validate and compare the different algorithms and preparation methods we classified nominal and unexpected operating states of the Columbus module using archive data from the Columbus Training, Qualification and Verification System (TQVS). Reports from an anomaly database were used to map unexpected states and a mission timeline was used to map nominal states. Our research demonstrates the benefits and drawbacks of different supervised machine learning algorithms in the context of telemetry classification and system monitoring in Human Spaceflight. On top of that, this work offers valuable insights for further developments on smart systems in the context of human deep space exploration.