On-line unsupervised detection of early damage

Abstract

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. detection approaches rely on signal processing and statistical learning techniques to extract sensitive information from time-series acquired on site [8][9][10][11]. Their computational simplicity makes them cost-effective and the most suitable candidates for carrying out automated on-line damage detection [2], either based on modal information [12][13][14] or based on statistical and time-series features [2,8,15].Data-driven SHM approaches rely on two mandatory steps for conducting damage detection: response modelling and statistical classification. The first aims at separating the variations imposed by 'normal' environmental/operational actions from those caused by damage [16]. It relies on training statistical learning algorithms so that they can accurately estimate the 'normal' structural response. Any 'abnormal' variations can afterwards be highlighted by comparing the estimates with the actual responses. The most reported statistical modelling algorithms found in SHM literature consist of multi-layer perceptron (MLP) neural networks [17][18][19], support vector regressions [20], linear regressions [2], principal component analysis [21] and auto associative neural networks [22]. Regardless of the chosen algorithms, response modelling has been reported in the literature as a supervised problem, where the statistical learning algorithms are trained a priori with reference data, in which the structural systems must be assumed undamaged and unchanged [9,11,23].Statistical classification consists of discriminating SHM data as related to identical or distinct structural conditions [24,25]. This step has also been addressed under supervised approaches, where classification algorithms are trained with reference data sets (in general, the same ones used for response modelling) to define boundaries that s...

Analysis of measurement and simulation errors in structural system identification by observability techniques

“…In fact, the state of the art of these methods has been reviewed in a number of works . According to most of these works, system identification methods can be classified as parametric and non‐parametric (genetic algorithms , evolutionary strategy , neural networks or least‐squares estimation ).…”

Bridge Damage Detection Based on Vibration Data: Past and New Developments

“…It relies on training statistical learning algorithms so that they can accurately estimate the "normal" structural response. The most reported statistical modeling algorithms found in SHM literature consist of multilayer perceptron neural networks, support vector regressions, linear regressions, principal component analysis (PCA), and autoassociative neural networks (Santos et al, 2016a). Dervilis et al (2015) conducted a study on the Tamar and Z24 Bridge to test a novel regression-based methodology for damage detection in changing environment and operational conditions.…”

“…The frequencies of vibration from the Z-24 bridge were used as modal parameters for the damage detection. In the case of the International Guadiana Bridge, a cable-stayed bridge with a main span of 324 m, the neural network technique was used to avoid the environmental/operational effects (Santos et al, 2016a). On the other hand, they used PCA in the same bridge for data normalization because it implicitly allows considering the effects of different actions without the need to measure them (Santos et al, 2015).…”

Constrained observability method in static structural system identification

“…SSI methods can be also classified as physics‐based (e.g., finite element models [FEMs]) or nonphysics‐based models (neural networks models, autoregressive models, or rational polynomial models). The parameters of physics‐based models represent the structural characteristics, for example, elastic modulus, inertias, and mass, whereas the parameters of the nonparametric models are weight factors of the adopted basis functions, which have no physical meaning and are determined by minimizing the discrepancy between the predicted and the measured response.…”