Simulation-driven machine learning for real-time damage prognosis in masonry structures

Abstract

Static structural health monitoring of masonry and heritage structures typically consists of tracking crack width evolution over time. However, the health evaluation of the current structural condition is not easily relatable to the actual cracks widths. In this paper, crack patterns in masonry walls are related to a stress increase indicator based on data generated through simulations employing accurate block-based numerical models of masonry walls damaged by differential settlements- and earthquake-like scenarios. Such stress increase indicator is defined through a percentile of the static cumulative minimum principal stresses distribution in a damaged wall, so it can be straightforwardly related to the occurrence of crushing failure. Driven by this simulation-generated dataset, a machine learning predictor is trained, validated and tested to provide stress increase indicators in damaged masonry walls by using as only input the crack width distributions of the walls. This allows to originally provide a crack pattern-based real-time damage prognosis tool in static monitoring of cracked masonry walls and structures.