“…Once damage-sensitive features have been extracted from the dataset, the final step of a data-driven SHM method is to analyze the features themselves for decision-making, providing outcomes in terms of early damage detection, localization, and quantification. At this stage, different techniques can be adopted, including statistical distance metrics (e.g., the Mahalanobis distance [ 23 , 24 ] or the Kullback–Leibler divergence [ 10 , 21 , 25 ]), Bayesian approaches [ 26 , 27 ], artificial neural networks [ 28 , 29 ], principal component analysis [ 30 , 31 ], and clustering [ 32 , 33 , 34 ]. In spite of their applicability, they may not perform efficiently when damage-sensitive features are of a high-dimensional nature, namely in the presence of big data to process; this leads to a time-consuming and unreliable decision-making process [ 10 , 35 ].…”