Structural health monitoring (SHM) strategies should ideally consist of continuous on-line damage detection processes, which do not need to rely on the comparison of newly acquired data with baseline references, previously defined assuming that structural systems are undamaged and unchanged during a given period of time. The present paper addresses the topic of SHM and describes an original strategy for detecting damage in an early stage without relying on the definition of data references. This strategy resorts to the combination of two statistical learning methods. Neural networks were used to estimate the structural response, and clustering methods were adopted for automatically classifying the neural networks' estimation errors. To ensure an on-line continuous process, these methods were sequentially applied in a moving windows process. The proposed original strategy was tested and validated on numerical and experimental data obtained from a cable-stayed bridge. It proved highly robust to false detections and sensitive to early damage by detecting small stiffness reductions in single stay cables as well as the detachment of neoprene pads in anchoring devices, resorting only to a small amount of inexpensive sensors. Copyright © 2015 John Wiley & Sons, Ltd.

This article addresses the subject of data-driven structural health monitoring and proposes a real-time strategy to conduct structural assessment without the need to define a baseline period, in which the monitored structure is assumed healthy and unchanged. Independence from baseline references is achieved using unsupervised discrimination machine-learning methods, widely known as clustering algorithms, which are able to find groups in data relying only on their intrinsic features and without requiring prior knowledge as input. Real-time capability is based on the definition of symbolic data, which allows describing large amounts of information without loss of generality or structural-related information. The efficiency of the proposed methodology is illustrated using an experimental case study in which structural changes were imposed to a suspended bridge during an extensive rehabilitation programme. A single-value novelty index capable of describing multi-sensor data is proposed, and its effectiveness in identifying structural changes in real time, using outlier analysis, is discussed.

Structural health monitoring (SHM) relies on data acquired from sensorial systems installed on site, and is nowadays being used more often not only for asset management, but also in critical structures when there is the need to detect damage in an early stage, before it impairs structural performance and safety. Early detection of damage in critical structures relies on the acquisition of continuous streams of information and on reliable techniques capable of analyzing it in real time, without generating false alerts. In this context, the combination of data fusion strategies, capable of converting large amounts of data into small pieces of information, with pattern recognition algorithms, which are able to analyze this information in real time, is addressed in the present paper with the objective of developing an original strategy capable of (1) removing the effects of regular actions imposed to structures without the need to measure them and of (2) compressing entire SHM data sets of arbitrary dimensions into a sensitive single-valued damage index. These capabilities are achieved by combining principal component analysis, the broke-stick rule, clustering methods, symbolic data objects, and symbolic distances. The proposed strategy was tested and validated with a numerical model of a cable-stayed bridge, using experimental data as input. From this analysis it was observed that, under the noise levels measured on site, the proposed strategy is able to automatically detect damage as small as 1% of stiffness reduction in a single stay cable. This sensitivity evidenced by the proposed strategy can be considered particularly high because it was obtained from a small amount of inexpensive sensors with a static character and because it was associated with a false detection incidence of only 0.1%.