In Europe, the building sector accounts for about 40% of total energy consumption and 36% of greenhouse gas emissions. Hence, robust in-situ identification methods providing a credibility interval on the wall thermal resistance are necessary for diagnosing the insulation of existing building walls and checking the expected energy performance of renovated walls. Herein, we propose Bayesian inverse methods using various physical models based on PDEs and a small number of non-intrusive sensors placed only on both wall sides. A prototype was also developed, in ANR RESBIOBAT project, to acquire measurements and to apply localized controlled thermal excitation on the interior side of the wall, enabling identification results to be obtained in a reduced test time and to be less sensitive to outdoor weather conditions. The use of Markov Carlo Markoc Chain in Bayesian inference requires a large number of physical model calls. Therefore, a multi-fidelity meta-modeling approach based on Gaussian process regression (GPR) was adopted to provide uncertainty quantification and good prediction accuracy in our application [1]. Bayesian inverse methods were adapted according to wall typology. For insulated internal walls, a standard Bayesian method with simple physical models (RC model and 1D thermal equation) used as two levels of fidelity is sufficient to accurately estimate the minimum wall thermal resistance using a short excitation time (10 h) and reduced instrumentation [2]. For the single-wall typology, a more sophisticated thermal model, i.e. 2D axisymmetric heat equations, is necessary and thus integrated into a three-level fidelity metamodeling and used in a sequential Bayesian inference framework [3] in order to better predict the a posteriori distribution of parameters iteratively while limiting the computational cost. Applications will be presented on virtual walls with synthetic data and on real walls with experiments conducted at Cerema Nancy and at the Sense-City facility.