Optimal feature selection for supervised bridge damage classification using responses collected from a passing vehicle

This paper proposes an innovative feature engineering approach for supervised bridge health monitoring (BHM) using signals measured from a passing vehicle. The proposed feature engineering approach consists of three main components:(a) signal-pre-processing, (b) feature extraction, and (c) feature selection (FS). The FS component adopts filter-based and wrapper-based approaches to find an optimal feature selection method for the damage classification using drive-by measurements. This component selects a subset of features characterized by minimal internal similarity and maximal relevance to the target class. In order to provide a deep insight to the FS component, 24 hybrid models for FS-based approaches and 36 hybrid models for wrapper-based FS are proposed in the drive-by BHM system and the strengths and weaknesses of the different methods are evaluated. Numerical simulations along with extensive laboratory experiments are used to test the performance of the proposed approach. Results show that the wrapper-based FS approach demonstrating an improvement of 10–20 % compared to filter-based FS approaches in damage classification. Furthermore, damage classification performance indexes of over 90 % are obtained for both numerical and experimental data sets, when adopting various machine learning (ML) algorithms to carry out the classification process. Also, an extensive analysis of the impact of bandpass filter (BPF) on framework accuracy is carried out. It is shown that applying BPF can significantly increase the classification accuracy. This study demonstrates that ML techniques when coupled with advanced feature engineering can provide detailed insights into the damage condition of bridges, purely based on the signals measured on a passing vehicle.