ESA GNC Conference Papers Repository

The consolidation of the artificial intelligence (AI) field has resulted in a paradigm shift towards data-driven machine learning (ML) tools for modelling, design, analysis, verification and validation. There is much interest in the space community in using these AI/ML methods with the aim to improve the performance, robustness and/or capabilities of the current (traditional and advanced) modeling and design approaches, but there are not yet many feasibility studies, and much less applications to systems of sufficient fidelity, to guarantee the successful transfer of the AI/ML methods to industrial operability. Addressing the aforementioned lack of studies, ESA released in 2020 a call for proposals to study the use of AI techniques for GNC design, implementation, and verification. This article presents results from one of the resulting projects (see details on the project and consortium below). Specifically, the article shows the results used to demonstrate the feasibility of a data-driven ML technique used to identify the most relevant parameters of a launch vehicle during the atmospheric ascent. The ML technique used is based on sparse regression techniques [1] in conjunction with compressed sensing, and it allows identification of linear or nonlinear systems from measurement data alone. The algorithm exploits sparsity-promoting techniques and machine learning with a library of possible candidate functions to identify the governing equations of systems characterized by relatively few non-zero terms. In a first phase of the project, the one presented in this article, the algorithm was applied to a simplified model of a launcher during atmospheric ascent (a well-known 2nd order nonlinear transfer function) in order to demonstrate its feasibility, performance, robustness, and shortcomings. This phase is critical to assess whether the algorithm is capable of being used subsequently for the nonlinear benchmark as well as to gather experience and knowledge on its tuning and onboard implementation capability. The aim of this feasibility application was to identify the two most relevant rigid-body rotational launcher parameters, commonly known as a6 and k1. These parameters are of particular interest to flight mechanics and control, as they are directly linked to the controllability and stability of the vehicle. The analysis of the proposed compressed sparse identification approach is presented in incremental steps of complexity in order to build confidence and gain insight on the process: starting with the case of constant dynamics, and gradually building up the complexity of the identification approach by considering a windowing compressed estimation, and finally a real ascent-flight, time-varying profile for the launcher dynamics. The results show that the proposed windowed compressed sparse identification approach can correctly identify the time-varying dynamics of the launch vehicle for a rotational parameter variation profile extracted from a real mission using nominal and dispersed (i.e uncertain) scenarios. This work is part of the project “Artificial intelligence techniques for GNC design, implementation, and verification” funded by ESA contract No. 4000134108/21/NL/CRS and participated by Deimos Engenharia, Deimos Space, INESC-ID, Lund University and TASC. References [1] S. L. Brunton, J. L. Proctor, and J. N. Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems, Proc. Natl. Acad. Sci. USA 113, 3932 (2016).