Design to the Edge: Perspectives of AI and Estimation Techniques in Autonomous Spacecraft

Artificial Intelligence (AI), in its different categories, is expected to remarkably affect space technology. This is already visible in the ground segment, especially in observation scheduling and data processing (images), but significant applications are envisaged also at the platform level, i.e. in the space segment. First, the field of failure identification and detection greatly benefits from these increased capabilities combining expertise and computation. Not only the diagnosis step, but also the more demanding prognosis one can be carried out continuously all over the bus functionalities. This monitoring will lead to an effective onboard autonomous failure management, defined as the adoption of the best solution to work in presence of non-nominal behaviour, without ground controllers’ intervention, outages nor delays. Even more remarkable, a breakthrough improvement can be allowed by the autonomous adaptation of the onboard systems to optimally perform in nominal, i. e. not plagued by failures, operations. Likewise the relaxed stability concept introduced in fly-by-wire aircraft, new operations envelopes can be approached. The combination of effective Bayesian estimation techniques (built on the basis of multiple filter algorithms) with learning capabilities of AI opens the possibility of an educated, continuous knowledge of the state of spacecraft subsystems together with its (estimated, yet credible) evolution in time. Subsequent operations can be therefore defined neglecting conservative safety allowances. This is a “design to the edge”, always maximising the performance on the basis of an up-to-date evaluation of the critical conditions. So, not anymore nominal conditions and pre-designed behaviour, yet the best exploitation of available resources every time. Notice that similar approaches are currently attempted in state-of-the art technological ventures as racing cars and high-performance sailing. The paper aims to investigate these topics by recalling some important characteristics of the ingredients involved. Specifically, the role of multiple model estimators, evolution of the Kalman filter algorithm, is discussed. It is deemed that multiple model filters’ introduction in AI deep learning architectures could approximate the behaviour of the control engineers in charge of a mission at the ground control centers, and indeed, once computational capabilities will allow its fielding on board, lead to full autonomy. Some possible applications in the astronautics field are listed, and their relevance shortly explained.