Réduction et recalage de modèles d'un module de puissance : vers un jumeau numérique probabiliste pour la prédiction de durée de vie restante

Power electronic modules transform the electrical current from the grid to meet the requirements of electric motors. These components are essential to numerous electrical systems. During their operation, losses generate heat within the module, leading to thermal stress, and eventually resulting in component failure. Current physic-based lifetime models are still limited by their computational time, making them impractical for real-time use in predicting the remaining lifetime. Furthermore, existing models do not account for the numerous sources of uncertainty that influence the module lifetime.This research work focuses on developing parameterized and multiphysics reduced models of an IGBT power module. The developed reduced models are based on the Proper Generalized Decomposition method. They decrease the computational cost of numerical model and can be used for uncertainty quantification studies. In this context, a non-intrusive implementation of an electro-thermo-mechanical model is developed in Ansys. The reduced model is then employed in real time to predict the remaining lifetime of a module. First, the method uses Bayesian inference and experimental measures to obtain a probability density over the model parameters. Then, the density is sampled using Transport Maps sampling and propagated through the numerical model to obtain online a probabilistic estimation of the remaining lifetime.