Assignment of structural behaviours in long-term monitoring: Application to a strengthened railway bridge

Abstract

Novelty detection, the identification of data that is unusual or different, is relevant in a wide number of real-world scenarios, ranging from identifying unusual weather conditions to detecting evidence of damage in mechanical systems. Using novelty detection approaches for structural health monitoring presents significant challenges to the non-expert user. In this article, symbolic data analysis is introduced to model variability in tests. Hierarchy-divisive methods and dynamic clouds procedures are then used to discriminate structural changes used as novelty detection approaches for classifying structural behaviours. This article reports the study of experimental tests performed on a railway bridge in France. This bridge has undergone reinforcement works during the summer of 2003. Through the years of 2004-2006, new sets of dynamic tests were recorded. The main objective was to analyse the evolution of the bridge's dynamic behaviour over time. To this end, the symbolic data analysis-based clustering methods are used for assigning new tests to clusters identified before and after strengthening or to highlight a totally different structural behaviour.

Novelty detection for SHM using raw acceleration measurements

“…Once more, both CH and Γ indicate six clusters, but C *, however, indicates nine. This latter tends to indicate a higher number of clusters, as observed in references .…”

Baseline-free real-time assessment of structural changes

“…This ability allows classifying clustering methods as unsupervised, or baseline-free, in opposition to supervised or baseline strategies, which are based on a previous definition of data references, in which the structural condition is assumed to be known and to remain unchanged. Baseline-free clusterbased discrimination has been sucessfully applied in previous works addressing the structural assessment of civil structural and infrastructural systems (Cury & Crémona, 2012;Jung & Koh, 2009;Silva, Dias Júnior, Lopes Junior, & Brennan, 2008).…”

“…Examples of statistical features found in the literature comprise regression's residual errors (Loh, Chen, & Hsu, 2011;Mata, 2011), principal components (Lanata & Grosso, 2006;Posenato et al, 2010;Santos et al, 2013), autocorrelation functions (Loh et al, 2011) or symbolic data (Cury & Crémona, 2012;Santos et al, 2012b). Symbolic data as SHM feature was only recently applied to SHM and proved particularly effective in fusing multi-sensor data without loss of damage-related information.…”

Structural modification assessment using supervised learning methods applied to vibration data

“…Before presenting the experimental applications explored in this paper, it must be kept in mind that most of these classification techniques -BDT, NN and SVM -were already applied to modal parameters (natural frequencies and mode shapes) obtaining very good results [15]. Now, the authors want to further explore the potentialities of the proposed approach using uniquely raw data i.e.…”

“…This combined methodology (SDA + supervised classification methods) has already proven its efficiency when modal parameters are used [11,15]. Now, this paper attempts to answer the following questions: (i) is it possible to classify different structural states using raw data (measured accelerations) only?…”