Propagation of visual inspection on timber members through Bayesian methods

Abstract

In this work, the variation of bending stiffness parameters of existing timber elements is assessed by analysis of an existing database of empirical results and by using Bayesian inference methods. The framework of this study initially considers the analysis of existing results from visual inspection and bending tests made to chestnut timber elements including a statistical analysis of the significance of different visual grades within the same size scale. After, Bayesian Probabilistic Networks are used to analyze the distribution of defects and to infer on the visual grading of neighboring segments for predicting the mechanical properties of the element. Finally, the results of the inference process are implemented in a finite element model of random generated elements where the information given by visual inspection on a local level is propagated to the global scale. The comparison between the experimental results and the results obtained through this methodology provided low percentage errors (lower than 3%) given that a significant benchmark sample size was available.