Bayesian predictive system for assessing the damage intensity of residential masonry buildings under the impact of continuous ground deformation.

The paper introduces a method for predicting damage intensity in masonry residential buildings situated in mining areas, focusing on the impact of large-scale continuous ground deformation. The research utilizes in situ data collected in a database, encompassing structural and material features, as well as information on maintenance quality and building durability. In addition to this information, the database collected data on the intensity of continuous deformation of the mining area at the location of the building, as well as the range and intensity of damage identified in buildings. The information included in the database was the result of many years of observations of buildings during the disclosure of impacts from mining exploitation and was based on: the results of in-situ building inventory, analysis of available building documentation and information provided by mining companies. The archived data were categorized variables labeled. The transformation of the data to a labeled value was dictated directly by the assumptions of the GOBNILP algorithm. Ultimately, a predictive model, represented by an optimal Bayesian network structure, is established. The optimisation of the network structure is achieved through the adaptation of the GOBNILP Bayesian network learning algorithm from data. This optimisation process is executed through the Gurobi Optimizer. It is worth noting that this interdisciplinary approach represents one of the first applications of such a methodology in the field of civil and environmental engineering. The results obtained can therefore be of significant value given the fact that the methodology of detecting the structure of Bayesian networks from data is still developing intensively in other scientific fields. In the course of the analyses, metric scores are examined, and various network structures are assessed based on their complexity. Great values of classification accuracies over 91% were obtained. This meticulous evaluation allows for the selection of the optimal Bayesian network that best generalises the knowledge acquired during the learning process. The paper also demonstrates the potential application of the obtained model in diagnosing damage causes and predicting future occurrences, highlighting the versatility of the proposed approach for addressing issues in the field.