Fractal-based manifold learning for structure health monitoring

Structure Health Monitoring (SHM) has been widely used in various engineering fields to ensure the safety of structures. Many of SHM methods are available based on machine learning to recognize the damage pattern, which are very time-consuming. A great challenge for most existing machine learning techniques is that their performances decrease as number of sensors increased for structure under analysis. In this paper, a new SHM technique, integrating manifold learning and fractal analysis, is proposed to detect structural damage. Both temporal and spatial features will be represented in a low dimensional embedding through dimensionality reduction. There are two procedures of the proposed method: temporal dimension reduction by fractal analysis, and spatial dimension reduction by manifold learning (Uniform Manifold Approximation and Projection-UMAP). The proposed methodology is applied to classify seven damage scenarios of benchmark study. The results showed high accuracy to classify different benchmark scenarios and can be potentially used for structure analysis which requires large number of sensors.