Speaker
Description
Environmental changes are inevitable parts of long-term structural health monitoring (SHM). For vibration-based damage detection of civil structures using modal frequencies, these changes significantly impact the accuracy of SHM. This paper proposes an unsupervised anomaly detection framework based on a hybrid clustering mechanism. In this first stage, Gaussian mixture model (GMM) is employed to identify the underlying probability distributions of the modal frequency data and segment them into initial clusters. Accordingly, the centroids derived from GMM serve as the main outputs for the next clustering framework. In this case, the second stage leverages k-means clustering (KMC) by employing the GMM-extracted centroids as the initialization points. This stage optimizes the initial centroids by reducing intra-cluster variance, which allows KMC to produce well-defined and statistically robust clusters without the need for additional heuristic methods to determine the number of clusters. The output centroids from KMC form the basis for a clustering-aided damage indicator (CDI), which quantifies the deviation of each modal frequency sample from its associated KMC-extracted centroid. During the anomaly detection phase, the CDI values of new samples are calculated and compared against a threshold estimated using the Generalized extreme value (GEV) model. This statistical approach effectively distinguishes between environmental variability and structural anomalies, enabling accurate and robust damage detection. Long-term modal frequencies of a full-scale concrete bridge are used to verify the proposed method. Results show that this method can significantly mitigate the impacts of environmental variability and yield reliable damage detection.
