Active learning in grey-box models for near-real-time online monitoring of dynamic processes

Miriam B. Dodt, Augustin Persoons, Matthias G.r. Faes, David MoensZenodo 2022 · DOI · OpenAlex

Abstract

Near-real-time monitoring of dynamical processes in a production line can aid greatly in increasing the<br> products quality and reliability, leading to an overall extended lifetime and reduced costs. It enables timely<br> intervention in case of diminishing quality or necessary parameter adjustment. In order to do so automatically,<br> a model that can continuously update and evaluate itself and its future predictions with high reliability<br> is of utmost importance. A so-called digital twin is well suited to perform this task, as its digital model is able<br> to update and estimate the future development as well as compute necessary countermeasures. The digital<br> twin is here modelled as a grey-box, which unites the high-accuracy of the white-box, numerical physical<br> model (i.e. FE model) with the computation speed of a black-box data-based Kriging surrogate model. A<br> key feature of the proposed approach is active refinement of the black-box model in parallel to performing<br> process control. The former is an adaptive calibration scheme, considering process drift as well as accuracy<br> of the surrogate.
BibTeX
@inproceedings{dodt2022active,
  author = {Miriam B. Dodt and Augustin Persoons and Matthias G.R. Faes and David Moens},
  title = {{Active learning in grey-box models for near-real-time online monitoring of dynamic processes}},
  booktitle = {Zenodo},
  year = {2022},
  doi = {10.5281/zenodo.7568274},
}

Loading related content...

Figure