Damage detection of a cable-stayed bridge using feature extraction and selection methods

Abstract

In this paper, a novel approach is presented to detect damage in a cable-stayed bridge based on feature extraction and selection. In the first part of the paper, several features are used in time domain, frequency domain, and time-frequency domain to detect damage. Next, several spectral parameters are introduced as effective features to reduce false alarms. Support vector machine (SVM) technique is employed as a classifier to compare the detection ability of the proposed and conventional features. In the second part of the paper, several feature selection techniques including forward selection, backward elimination, the minimum redundancy and max relevancy, ReliefF and SVM approach based on recursive feature elimination (SVM-RFE) algorithm are employed to improve feature extraction accuracy through selecting and ranking the most discriminative feature as an optimal feature subset. The capability of the proposed approach to detect damage is investigated by a data set obtained from a health monitoring system of a real benchmark problem. The data consists of vibration acceleration of the Yonghe bridge in both healthy and damaged states. Results show that the proposed procedure can effectively reduce false alarms.