Summary
Structural health monitoring is a problem that can be addressed at many levels. One of the most promising approaches used in damage assessment problems is based on pattern recognition. The idea is to extract features from data that characterize only the normal condition and to use them as a template or reference. During structural monitoring, data are measured, and appropriate features are extracted as well as compared with the reference. Any significant deviations are considered as signal novelty or possible damage. Several studies present in the literature are based on the comparison of measured vibration data such as natural frequencies and vibration modes in undamaged and damaged states of the structure. This methodology has proven to be efficient; however, its application may not be the most adequate in cases where the engineer needs to know with certain imperativeness the condition of a given structure. This paper proposes a novelty detection approach where the concept of symbolic data analysis is used to manipulate raw vibration data (i.e., acceleration measurements). These quantities (transformed into symbolic data) are combined to three unsupervised classification techniques: hierarchy agglomerative, dynamic clouds and soft c‐means clustering. In order to attest the robustness of this approach, experimental tests are performed on a simply supported beam considering different damage scenarios. Moreover, this paper presents a study with tests conducted on a motorway bridge, in France, where thermal variation effects also play a major role. In summary, results obtained confirm the efficiency of the proposed methodology. Copyright © 2015 John Wiley & Sons, Ltd.