The paper aims at proposing a new strategy for real-time identification or updating of structural mechanics models defined as dynamical systems. The main idea is to introduce the modified constitutive relation error concept, which is a practical tool that enables to efficiently solve identification problems with highly corrupted data, into the Kalman filtering, which is a classical framework for data assimilation. Furthermore, a PGD-based model reduction method is performed in order to optimize capabilities of the online updating strategy. Performances of the proposed approach, in terms of robustness gain and computational cost reduction, are illustrated on several unsteady thermal applications.