An approach for adaptive model performance validation within digital twinning

Abstract

The validation of the operationality of models is considered a crucial step in the model development process. Recent developments in Digital Twinning (DT) enable the online availability of operational data from the physical asset required for operational validation. The benefits of DT in situations where operational validation has formed a basis for model adaptation has also been demonstrated. However, these benefits within DT have not been fully utilized due to the lack of an approach for benchmarking the required quantity, quality and diversity of validation data and performance metrics for online model validation and adaptation. Therefore, there is a need for a framework for benchmarking validation data and metrics requirements during model validation in different domains. An approach for benchmarking the required quantity, quality and variability of validation data and performance metric(s) for online model adaptation within DT is proposed. The approach is focused on addressing the problem of parameter(s) uncertainty of a predictive model within its uncertainty boundary. It involves generating virtual test models, a primary and another reference model for the performance evaluation of one compared to the another with the benchmarked validating data and metrics within DT. This process is repeated until the dataset and/or metric(s) are promising enough to validate primary model against the reference model. The proposed approach is demonstrated using BEASY - a simulator designed to predict protection provided by a cathodic protection system to an asset. In this case, a marine structure is the focus of the study, where the protection potentials to prevent corrosion are predicted over the life of the structure. The algorithm(s) for the approach are provided within a Scientific Software (MATLAB) and integrated to the simulator-based cathodic-protection model.