Online early damage detection and localisation using multivariate data analysis: Application to a cable-stayed bridge

Abstract

An online data-based methodology for early damage detection and localisation under the effects of environmental and operational variations (EOVs) is proposed. The methodology is described in detail and implemented in a large prestressed concrete cable-stayed bridge of which 3.5 years of data are available. The effects of EOVs are suppressed by the combined application of two well-established multivariate data analysis methods: multiple linear regression and principal component analysis. Criteria for the systematic choice of the predictor variables and the number of principal components to retain are proposed. Because the bridge is new and sound, the experimental time series are corrupted with numerically simulated damage scenarios in order to evaluate the damage detection ability. It is demonstrated that the sensitivity to damage is increased when daily, 2-day, or 3-day averaged data are used instead of hourly data. The effectiveness of the proposed methodology is also demonstrated with the detection of a real, small, and temporary sensor anomaly. The implemented methodology has revealed to be robust and efficient, presenting a contribution to the transition of structural health monitoring from academia to industry.