Selecting meaningful predictor variables: A case study with bridge monitoring data

Abstract

Bridge Health Monitoring (BHM) is an important problem in many countries, including Viet Nam. Therefore, there have been many proposed mechanical, mathematical, statistical, etc. methods created to solve this problem. In BHM process, one important step is to reduce and extract important information from realistic datasets obtained from bridge monitoring. Our contribution in this study is on reduction of variables measured on the bridge using the Principal components analysis (PCA), in coupling with some additional methods. Specifically, after achieving a new dataset having new uncorrelated variables using PCA, this study uses the idea of cross-validation to point out some first few components enough to be able to reconstruct the original data with appropriate information (variance). Finally, for the purpose of variable reduction, the Canonical correlations analysis is used to decide which subset of the original dataset keeps the most information. To validate the new method in practical usage, this study uses Sai Gon Bridge's vibration measure data created by Laboratory of Applied Mechanics, Ho Chi Minh city University of Technology, implementing it in statistical programming environment R and a small Java application to allow users to interact and see the results graphically.