A region-based machine learning approach for self-diagnosis of a 4D digital thermal twin
In this paper, we explore the applicability of machine learning as a tool for self-diagnosis of 4D digital twins with a focus on simulated surface temperatures. Generation of digital twins involves abstractions, simplification, and the closed-world assumption. Hence, performing thermal simulation in order to obtain surface temperatures involves not only mathematical modeling of the physical phenomena, but also temporal uncertainties on external conditions. To identify the types of simulation inaccuracies, our proposed method is based on thermal image comparison, i.e. the corresponding measured thermal image and the simulated thermal image resulting from the 4D digital twin. First, a statistical necessary condition is defined to obtain regions of interest in the simulated image. Second, after manual labeling of these regions into the two inaccuracy classes, we conduct a detailed feature analysis and subsequently train our Random Forest classifier. The results show a good separability of the two classes despite the limited training data, allowing to achieve values of overall accuracy around 93.5%.