Probabilistic damage localization by empirical data analysis and symmetric information measure

Abstract

This article aims to address some challenges in data-based damage localization by proposing innovative methods including a hybrid algorithm for feature extraction, a symmetric information measure for feature classification, and a probabilistic approach to threshold estimation. The hybrid feature extraction combines an autoregressive (AR) model with a novel non-parametric estimator of probability density function (PDF) under empirical data analysis. The great novelty of this method is to propose a new probabilistic feature as an empirical PDF of the AR coefficients. The proposed information measure is a symmetric divergence aiming at addressing the main limitation of the classical Kullback-Leibler divergence regarding its non-symmetric characteristic. Finally, the proposed probabilistic approach exploits the concept of Markov Chain Monte Carlo to estimate a trustworthy threshold for locating damage. Vibration responses of two civil structures are used to verify the proposed methods with several comparisons. Results confirm that the methods are successful in identifying damage.