Machine Learning-aided Structural Health Monitoring for uncertainty reduction in seismic evaluation and robust damage detection in existing masonry buildings

A significant part of the existing building stock in Europe has been designed without modern seismic considerations and is, in the meantime, exceeding its design life-span. Even in sites of moderate seismicity, earthquakes bear large loss potential, driven by low-probability but high-impact events. Reducing seismic risk remains a major scientific frontier and reliable estimation of building vulnerability prior to and in the direct aftermath of earthquakes is a fundamental step towards quantifying and, eventually, reducing seismic risk. Structural Health Monitoring (SHM) offers tools to leverage measurement data to accelerate evaluations and reduce the uncertainty pertaining to engineering models. Despite recent advances in sensor technology and computational capabilities, SHM retains unexploited potential and is often limited to the research domain. This is mainly attributed to missing guidelines and standardized frameworks for efficient treatment and robust interpretation of monitoring data, as well as to the limited amount of recordings from real buildings. This thesis aims to provide practical techniques to support the seismic evaluation of existing buildings with monitoring data, as well as to develop robust post-earthquake damage indicators. A first step towards this goal consists in collecting data from real buildings with higher amplitudes than ambient vibrations and in different damage states. To facilitate and formalize data collection, a framework for monitoring unreinforced masonry (URM) buildings prior to and during planned demolitions is proposed. URM buildings comprise a significant part of the existing building stock in Europe, one that is particularly vulnerable to earthquakes. Vibration-based monitoring in operating buildings is hindered by accessibility issues and is limited to low-amplitude ambient vibration measurements. Demolitions of URM buildings are conducted progressively with an excavation shovel, providing a cost-efficient way to collect vibrational data. Studying the dynamic response at various amplitude levels in the commonly assumed linear-elastic regime exposed reversible stiffness reduction, which affects the predicted seismic performance and is mainly attributed to “breathing” cracks in masonry. Based on monitoring recordings from nine URM buildings undergoing demolition, a Bayesian model-updating framework is proposed and implemented, demonstrating a significant reduction in uncertainties related to predicted seismic performance. This enables the direct comparison of multiple buildings featuring either similar or different structural characteristics. Furthermore, by extrapolating the information extracted from monitored buildings that are representative of a broader typological class, provides the means to dynamically update vulnerability and risk maps at regional scale. Clustering together buildings featuring similar seismic performance is eminent, to define representative buildings at regional scale, thus paving the way to data-informed regional post-earthquake loss assessment. Therefore, damage-sensitive features (DSFs) derived from vibrations are used to quantify the damaging impact of ground motions. An extensive set of DSFs computed on low-amplitude response prior to and directly after earthquakes is evaluated in terms of damage-prediction accuracy. The predictive performance of fragility functions involving individual DSFs is compared with Engineering Demand Parameters (EDPs) and intensity Metrics (IMs). DSFs are shown to outperform peak ground acceleration, which is a typical IM and is currently used in regional risk and damage assessment. Gradientboosted decision trees and convolutional neural networks are deployed to fuse multiple DSFs into robust damage classifiers. Owing to the uncertainties arising when simulating nonlinear building responses to earthquakes, a domain adaptation framework is proposed and enables successful transfer of the knowledge obtained from physics-based simulation models to real data. While such techniques are very promising when applied to available data, long-term monitoring application are undermined by the limited life-cycle of sensors compared to the monitored structures. To alleviate this issue, timely identification of sensor faults and robust characterization of their origin are required. A semi-supervised framework for sensor fault detection is combined with an approach to interpret black-box model predictions, on the basis of the herein proposed “decision trajectories”. Additionally, a measure of correlation, namely the Decision Trajectory Assurance Criterion (DTAC), which enables the characterization of new fault types and the automatic classification of anomalies to known fault classes. The framework is designed to be independent of structure-specific characteristics and the type, amount, and locations of sensors used, providing an easy to train, project-agnostic tool, which can be seamlessly integrated into the pre-processing part of many SHM applications. This thesis contributes to seismic vulnerability assessment, by confronting the main challenges that currently limit the broad application of vibrationbased SHM for data-driven enhancement of seismic evaluation before and after earthquakes. Extensive monitoring data from real structures has been acquired during demolitions and used to validate the proposed frameworks, together with benchmark monitoring data from long-term monitored structures and a large-scale shake-table test. Thorough comparisons with the state of practice have been conducted to quantify the benefit of data-informed engineering solutions. Overall, the outcome of this work sheds light on the practical value of integrating SHM in seismic evaluation, towards SHM-based vulnerability assessment, and provides a baseline for the broad establishment of vibration-based SHM as a valid tool in the engineering community.