Higher, faster, and further on the one hand, but also safer and more efficient on the other hand – this trend demands a great deal of today’s and future engineers. In addition to the challenging maxim, economic aspects lead to the desire to reduce time to market of products required to master today’s challenges, which pushes society away from physical towards virtual prototyping, i.e., the use of digital twins.
Multibody system dynamics simulations are powerful tools to realistically analyse real-world, multiple-component devices in their intended operating environment. However, simulations of systems of engineering relevance include often numerous components and degrees of freedom, especially when the bodies’ flexibility needs to be taken into account. This computational burden has to be alleviated with model order reduction approaches in most cases for these simulations to show potential for real-time applications, e.g., in the control of humanoid robots. Furthermore, the inherently non-linear governing equations are often (at least partially) unknown, e.g., contact and damping terms, and cannot be easily reduced with classical projection-based reduction techniques. These problems may be addressed with data-driven approaches.
This thesis will, therefore, focus on the combination of data-driven and classical approaches for non-linear model order reduction and the identification, i.e., “learning”, of the dynamics of mechanical systems.