Bayesian Classification of Damage Modes in Existing Masonry Buildings from Descriptive Vulnerability Factors

Seismic risk mitigation strategies for urban vernacular buildings can leverage descriptive vulnerability factors to infer, through predefined functions or machine learning techniques, the expected damage. This can be described as an overall damage grade, as in EMS-98, or as modes, i.e., masonry crumbling, out-of-plane tilting and in-plane response of piers and spandrels. Predictions regarding the modes can improve the reliability of the scenario compared to overall damage. This paper presents an automated tool for the prediction of the expected damage mode based on 21 vulnerability factors. The qualitative relationships between damage and vulnerability are explored through a Naive-Bayes classification algorithm, which obtains the conditional probability of each input item (building) to belong to each output class (damage mode). This Bayesian network is trained on a dataset of 314 masonry buildings, hit by the 2016 Central Italy earthquake, whose 21 vulnerability factors, EMS-98 overall damage and main damage modes were known. Moreover, the influence of each factor on damage was assumed as an improvement coefficient of the estimates, and additional coefficients corrected the classification if it guessed wrong. The predicting capacity of the network was tested on 20 additional buildings, obtaining a 20–50% rate of true positives, increased to 40–60% after the application of the correction factors.