State of the Art of Machine Learning Techniques for Spacecraft Applications

Any kind of spacecraft, manned or unmanned, requires a certain level of controllability which is usually higher than most other applications, due to the strict requirements of a space-based mission, along with the inaccessibility of a hands-on approach on the hardware. Spacecraft operations need therefore a high grade of precision in order to accomplish their goals, justifying the enormous cost of not only launch but many other aspects of a spacecraft development and production. This means that, whether in nominal operation mode or after changes due to external inputs, the behavior of the machine must be closely controlled, with a strict policy of decision-making apt to maximize efficiency of operations and at the same time safeguarding the spacecraft from potential failures. This paper analyzes those instances where this control is handled not by pilots, nor flight engineers nor ground station workers, but by software that is capable to tackle different scenarios from the experience learned throughout a prior training process. This search for autonomy is crucial in the new developments that are boosting the space industry in what has been called the “New Space Economy”. An important focus is the creation and management of Mega-constellations, fleets of spacecraft that create a network for all kind of operations. The elaborate will analyze cases and studies regarding Machine Learning for purposes helpful for spacecraft fleets, with the overall satellite management, on orbit operations and formation flying being the three main focal points.