Novelty Detection in Projected Spaces for Structural Health Monitoring

The aim of Structural Health Monitoring (SHM) is to detect and identify damages in man-made structures such as bridges by monitoring features derived from vibration data. A usual approach is to deal with vibration measurements, obtained by acceleration sensors during the service life of the structure. In this case, only normal data from healthy operation are available, so damage detection becomes a novelty detection problem. However, when prior knowledge about the structure is limited, the set of candidate features that can be extracted from the set of sensors is large and dimensionality reduction of the input space can result in more precise and efficient novelty detectors. We assess the effect of linear, nonlinear, and random projection to low-dimensional spaces in novelty detection by means of probabilistic and nearest-neighbor methods. The methods are assessed with real-life data from a wooden bridge model, where structural damages are simulated with small added weights.