Fusing damage-sensitive features and domain adaptation towards robust damage classification in real buildings

Structural Health Monitoring (SHM) enables the rapid assessment of structural integrity in the immediate
aftermath of strong ground motions. Data-driven techniques, often relying on damage-sensitive features (DSFs)
derived from vibration monitoring, may be deployed to attribute a specific damage class to a structure. In
practical applications, individual features are sensitive to specific levels of damage, and therefore combining
multiple DSFs is required to formulate robust damage indicators. However, the combination of DSFs typically
involves empirical thresholds that are often structure-specific and hinder generalization to different structural
configurations. This work evaluates the predictive performance of a large ensemble of DSFs, computed on
an extensive dataset of nonlinear simulations of frame structures with varying geometrical and material
configurations. Gradient-boosted decision trees and convolutional neural networks are deployed to fuse
multiple DSFs into damage classifiers, improving the predictive accuracy compared to best-practice methods
and individual DSFs. A Domain Adversarial Neural Network (DANN) architecture enables the transfer of
knowledge obtained from numerical simulations to real data from a large-scale shake-table test. After exposure
to limited data, exclusively from the healthy state, the DANN framework yields satisfactory performance in
predicting unseen damage states in the experimental data. The results demonstrate the potential of DANN
in transferring knowledge from simulations to real-world monitoring applications, where only limited data
characterizing exclusively the current, typically healthy, structural state is available. Overall, this work
comprises the definition of multiple DSFs, their fusion through ML approaches, and the generalization of the
knowledge obtained from simulations to real data through domain adaptation.