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

Abstract To achieve its full predictive potential, a digital twin must consistently and accurately reflect its physical counterpart throughout its operational lifetime.To this end, the inverse mapping parameter updating method enables physically interpretable parameter values to be updated, in real-time, for a wide range of (nonlinear) dynamical models using features extracted from measured response data. This paper proposes to extend this method by employing a probabilistic Bayesian neural network, which is trained offline using simulated data, to infer, again in real-time, probability distributions for the updating parameter values instead of (traditionally obtained) point estimates. As a result, the user obtains a quantification of the (un)certainty, providing insight into the degree of trust to be placed in the updated parameter values, which supports the decision-making process for which the digital twin is used. Additionally, it is proposed to include so-called ‘input parameters’ (that characterize the specific settings on the physical setup) as inputs to the neural network to allow for a broader applicability of the updating method. To validate the proposed methodology, it is applied, using both simulated and real-world measurements, to a medical mechanical ventilation system, in which information about uncertainty in the inferred parameter values is important. Parameter values of this system and their uncertainties are shown to be inferred with sufficient accuracy.