Damage Classification and Feature Extraction in Steel Moment-Resisting Frame Using Time-Varying Autoregressive Model

Abstract

In this paper, the time-varying autoregressive (TVAR) model is integrated with the K-means—clustering technique to detect the damage in the steel moment-resisting frame. The damage is detected in the frame using nonstationary acceleration response of the structure excited using ambient white noise. The proposed technique identifies and quantifies the damage in the beam-to-column connection and column-to-column splice plate connection caused due to loosening of the connecting bolts. The algorithm models the nonstationary acceleration time history and evaluates the TVAR coefficients (TVARCs) for pristine and damage states. These coefficients are represented as a cluster in the TVARC subspace and segregated and classified using K-means—segmentation technique. The K-means—approach is adapted to simultaneously perform partition clustering and remove outliers. Eigenstructure evaluation of the segregated TVARC cluster is performed to detect the temporal damage. The topological and statistical parameters of the TVARC clusters are used to quantify the magnitude of the damage. The damage is quantified using the Mahalanobis distance (MD) and the Itakura distance (ID) serving as the statistical distance between the healthy and damage TVARC clusters. MD calculates a multidimensional statistical distance between two clusters using the covariance between the state vectors, whereas ID measures the dissimilarity of the autoregressive (AR) parameter between reference state and unknown states. These statistical distances are used as damage-sensitive feature (DSF) to detect and quantify the initiation and progression of the damage in the structure under ambient vibrations. The outcome of both the DSFs corroborate with the experimental investigation, thereby improving the robustness of the algorithm by avoiding false damage alarms.