Uncertainty quantification in real-time parameter updating for digital twins using Bayesian inverse mapping models

Abstract To reach its full predictive potential, a digital twin should accurately represent its physical copy in the real world during its entire service lifespan. To this extent, model parameter updating is employed using the inverse mapping parameter updating method. This method enables physically interpretable parameters to be updated for a wide range of (nonlinear) dynamical models. It does so by inferring, in real-time, parameter values based on features extracted from measured data. This is achieved by using a regression model (e.g., based on machine learning) that is trained in an offline phase on simulated data. To allow for the quantification of the uncertainty in the inferred updating parameters, this paper proposes to use probabilistic Bayesian neural networks. This type of neural network infers probability distributions for the updating parameter values instead of the point estimates obtained when employing traditional (deterministic) artificial neural networks. Consequently, the user gains insight in how certain the neural network is about the inferred parameter values and can adjust his/her decision-making based on this knowledge.Besides the features of measured responses, additionally, known so-called `input parameters' that characterize the performed experiments (e.g., specific settings on the physical setup) can be provided as input to the neural network. This allows for a broader applicability of the trained neural network.To validate the proposed methodology, it is applied, using both simulated and real measurements, to a real-world mechanical ventilation system.