Crack pattern–based machine learning prediction of residual drift capacity in damaged masonry walls

Abstract

In this paper, we present a method based on an ensemble of convolutional neural networks (CNNs) for the prediction of residual drift capacity in unreinforced damaged masonry walls using as only input the crack pattern. We use an accurate block-based numerical model to generate mechanically consistent crack patterns induced by external actions (earthquake-like loads and differential settlements). For a damaged masonry wall, we extract the crack width cumulative distribution, we derive a crack width exceedance curve (CWEC), and we evaluate the drift loss (DL) with respect to the undamaged wall. Numerous pairs of CWEC and DL are thus generated and used for training (and validating) an ensemble of CNNs generated via repeated $k$-folding cross validation with shuffling. As a result, a method for damage prognosis (Level IV of SHM) is provided. Such method appears general, inexpensive, and able to adequately predict the DL using as only input the CWEC, providing real-time support for decision making in damaged masonry structures.