Inference on stiffness and strength of existing chestnut timber elements using Hierarchical Bayesian Probability Networks

Abstract

The assessment of the mechanical properties of existing timber elements could benefit from the use of probabilistic information gathered at different scales. In this work, Bayesian Probabilistic Networks are used to hierarchically model the results of a multi-scale experimental campaign, using different sources of information (visual and mechanical grading) and different sample size scales to infer on the strength and modulus of elasticity in bending of structural timber elements. Bayesian networks are proposed for different properties and calibrated using a large set of experimental tests carried out on old chestnut (Castanea sativa Mill.) timber elements, recovered from an early 20th century building. The obtained results show the significant impact of visual grading and stiffness evaluation at different scales on the prediction of timber members’ properties. These results are used in the reliability analysis of a simple timber structure, clearly showing the advantages of a systematic approach that involves the combination of different sources of information on the safety assessment of existing timber structures.