On vibration-based damage detection by multivariate statistical techniques: Application to a long-span arch bridge

Abstract

Structural health monitoring allows the automated condition assessment of civil infrastructure, leading to a cost-effective management of maintenance activities. However, there is still a debate in the literature about the effectiveness of available signal processing strategies to timely assess the health state of a structure. This paper is a contribution to this debate, by presenting the application of different vibration-based damage detection methods using up-to-date multivariate statistical analysis techniques applied to data acquired from a permanently monitored long-span arch bridge. Techniques based on dynamic regression models, linear and local principal component analysis, as well as on their combinations, including, in particular, the newly proposed method based on the combination of dynamic multiple linear regressions and local principal component analysis, and, finally, a method based on the recently proposed approach of cointegration, are considered. A first effort is made to formulate these methods within a unique mathematical framework, highlighting, in particular, the relevant parameters affecting their results and proposing objective criteria for their appropriate tuning and for choosing the length of the training period. Then, the considered damage detection methods are implemented and applied to field data, seeking for damage-sensitive features in the presence of variable environmental and operational conditions. The considered techniques are applied to time histories of identified modal frequencies of the bridge and their capability to reveal structural damage of varying severity is assessed using control charts. The case of an artificially imposed non-linear correlation between the features is also considered. The results provide, for the first time in the literature, an estimation of the minimum level of damage that can be realistically detected in the bridge using dynamic signatures and up-to-date signal processing algorithms, thus contributing to a more aware use of monitoring data and reliance over related health state assessment information.