Machine learning based digital twin for dynamical systems with multiple time-scales

Abstract Digital twin technology has a huge potential for widespread applications in different industrial sectors such as infrastructure, aerospace, and automotive. However, practical adoptions of this have been slower, mainly due to lack application-specific details. Here we focus on digital framework linear single-degree-of-freedom structural dynamic systems evolving two operational time scales addition its intrinsic time-scale. Our approach strategically separates into components – (a) physics-based nominal model data processing response predictions, (b) data-driven machine learning the time-evolution system parameters. The is system-specific selected based problem under consideration. On other hand, generic. For tracking multi-timescale evolution parameters, propose exploit mixture experts model. Within model, Gaussian Process (GP) used expert primary idea let each track parameters at single hyperparameters ‘mixture using GP’, an efficient that exploits expectation-maximization sequential Monte Carlo sampler used. Performance illustrated dynamical with stiffness and/or mass variations. found be robust yields reasonably accurate results. One exciting feature proposed capability provide reasonable predictions future time-steps. Aspects related quality quantity are also investigated.