Enhanced features in principal component analysis with spatial and temporal windows for damage identification

Principal component analysis (PCA) methods have been widely applied to damage identification in the long-term structural health monitoring (SHM) of infrastructure. Usually, the first few eigenvector components derived by PCA methods are treated as damage-sensitive features. In this paper, the effective method of double-window PCA (DWPCA) and novel features are proposed for better damage identification performance. In the proposed method, spatial and temporal windows are introduced to the traditional PCA method. The spatial windows are applied to group damage-sensitive sensors and exclude those sensors insensitive to damage, while the temporal window is applied to better discriminate eigenvectors between the damaged and healthy states. In addition, the length and directional angle of the eigenvector variation between the healthy and damaged states are used as the damage-sensitive features, instead of the components of the eigenvector variation used in previous studies. Numerical simulations based on a large-scale bridge reveal that the proposed features are successful in identifying the damage located far from sensors due to the use of both spatial and temporal windows as well as the length of the eigenvector variation. In addition, compared to the previous PCA and moving PCA methods, the novel features have higher sensitivity and resolution in damage identification.