A neural network-based automated methodology to identify the crack causes in masonry structures

Abstract

Most masonry constructions exhibit significant crack patterns caused by differential foundation settlements. While modern numerical methods effectively address forward displacement-based problems, identifying the settlement causing a specific crack pattern remains an unsolved yet crucial challenge. For the first time, this research solves this highly non-linear back-engineering problem by proposing a robust and automated methodology synergizing artificial neural networks (ANNs) and the piecewise rigid displacement (PRD) method. The PRD's fast computational solving allows the generation of large datasets used to train specific ANNs through Levenberg–Marquardt and conjugate gradient algorithms. Using the location and widths of the main structural cracks as input, the proposed approach offers an instantaneous and accurate ANN-based identification of foundation settlements that cause the detected damage scenario. The method is first validated on semicircular arches, and after that, its potential and effectiveness are demonstrated in a real engineering scenario, represented by the Deba bridge in Spain.