Baseline-free real-time novelty detection using vibration-based symbolic features

The present works aims at describing a new strategy for vibration-based detection of structural changes, or novelties. It is based on the definition of symbolic modal quantities and in analyzing them with clustering methods under a time-windows process which allows for baseline-free analysis capable of detecting changes in real-time. At each time-window, a new single-valued feature is defined as the average distance between clusters of symbolic modal features. This feature is then statistically characterized and tested for novelty detection. The influence of the temperature action on frequencies, mode shapes and damping ratios is removed resorting to neural networks, applied independently at each time-window. The effectiveness of the proposed strategy is tested on a case study consisting of vibration monitoring on a railway bridge. The analyses conducted on this data revealed greater sensitiveness of damping ratios to structural changes but earlier detections obtained from modal frequencies. Modal Features, Symbolic Data, Clustering Methods, Unsupervised Outlier Detection, Moving Time-Windows