A challenging issue in data-driven methods is the limitation of applying high-dimensional damage-sensitive features (HD-DSFs) to localize and quantify structural damage. This is because the use of such data causes a time-consuming process for feature classification. Low damage detectability is another important challenge that may cause unreliable and erroneous results. The main objective of this study is to deal with these challenging issues by proposing hybrid distance methods as combinations of a distance-based indicator called residual relative error (RRE) and Kullback–Leibler divergence (KLD) with the well-known Mahalanobis distance. Autoregressive modeling is used to model vibration responses and extract the residual samples from normal and damaged conditions as HD-DSFs. The major contribution of this article is to propose two hybrid distance-based methods in unsupervised learning manners for locating and quantifying damage using low-dimensional feature samples obtained from the RRE index and KLD. The great advantage of these methods is that they not only reduce the size of HD-DSFs and behave as dimensionality reduction tools but also can increase damage detectability by the low-dimensional feature samples that are distance quantities. For this reason, both the RRE and KLD are individually capable of detecting early damage. An ability to locate and quantify damage under varying environmental and operational conditions is the other advantage of these methods. To validate the proposed methods, experimental vibration datasets of two benchmark laboratory structures are utilized. Results demonstrate that the proposed methods are practical and reliable tools to accurately localize and quantitatively quantify damage by addressing the limitation of using HD-DSFs.

Keywords