Wyodrębnianie struktury sieci Bayesowskiej z danych w prognozowaniu uszkodzeń żelbetowych budynków prefabrykowanych na terenach górniczych

Abstract

This article presents the results of the research on the construction of a model for assessing the risk of damage to building structures located in mining areas. The research was based on the database on the structure, technical condition and mining impacts regarding 129 prefabricated reinforced concrete buildings erected in the industrialised large-block system, located in the mining area of the Legnica-Glogow Copper District (LGCD). The methodology of the Bayesian Belief Network (BBN) was used for the analysis. Using the score-based Bayesian structure learning approach (Hill-Climbing and Tabu-Search) as well as the selected optimisation criteria, 16 Bayesian network structures were induced. All models were subjected to quantitative and qualitative evaluation by verifying their features in the context of accuracy of prediction, generalisation of acquired knowledge and cause-effect relationships. This allowed to select the best network structure together with the corresponding optimisation criterion. The analysis of the results demonstrated that the Tabu-Search method adopting the optimisation criterion in the form of Locally Averaged Bayesian Dirichlet score (BDla) led to obtaining a model with the best features among all the selected models. The results justified the adoption of the BBN methodology as effective in the context of assessing the extent of damage to building structures in mining areas.