Manifold learning-based unsupervised feature selection for structural health monitoring

Abstract. Feature selection plays an important role in enhancing the performance of a data-driven structural health monitoring (SHM) system by selecting more informative and lower-dimensional features from the original ones. In this paper, an unsupervised feature selection method based on manifold learning (MLUFS) is proposed to identify feature sets that are beneficial to preserve the structure of the data. This method leverages nonlinear dimensionality reduction methods and canonical correlation analysis to address the challenges associated with acquiring costly labels in the SHM field. It is applicable to high-dimensional data that lie on a low-dimensional manifold within the ambient space. A case study on the handwritten digit datasets is presented to compare the classification performances corresponding to the features selected by the MLUFS method and a supervised feature selection method. The results show that the proposed method can provide feature sets with higher generalization ability compared to the selection method using labels. Another simulated acceleration dataset is used to demonstrate the effectiveness of the proposed method in anomaly detection for an SHM system.