Three-way analysis of structural health monitoring data

Structural health monitoring aims to detect damages in man-made engineering structures by monitoring changes in their vibration response. Unsupervised learning algorithms can be used to obtain a model of the undamaged condition and detect which new samples of the structure are not in agreement with it. However, in real structures with a sensor network configuration, the number of candidate features usually becomes large. Therefore, complexity increases and it is necessary to perform feature selection and/or dimensionality reduction to achieve good detection accuracy. In this paper, we propose to exploit the three-way structure of data and apply a true multi-way data analysis algorithm: Parallel Factor Analysis. A simple model is obtained and used to train novelty detectors. The methods are tested both with real and simulated structural data to assess that the three-way analysis can be successfully used in structural health monitoring. Furthermore, the usefulness of the approach for feature selection is also analyzed.