Abstract
Environmental effects often cause variability in dynamic features, obscuring actual damage indicators and leading to false alarms in damage detection. The Gaussian mixture model (GMM) based method is an effective solution, but challenges such as selecting initial model parameters and determining the optimal number of Gaussian components can hinder its performance. To address these challenges, we propose a two-step method that combines sequential iteration with the GMM approach. In the first step, sequential iteration is employed to determine initial model parameters and the optimal number of Gaussian components for a reliable GMM. In the second step, the expectation-maximization (EM) algorithm is used to establish the GMM, clustering the training data into local subsets. For each subset, the Mahalanobis squared distance (MSD) between each sample point and the center of its Gaussian component is calculated. This distance is used to create a novelty index based on the minimum Mahalanobis squared distance (MMSD), facilitating effective damage detection by the statistical control chart. Moreover, generalized extreme value distribution modeling method is presented to determine an accurate control limit. We validate our method using real data from two bridges, demonstrating its effectiveness through comparative analysis.