Decision Tree-based Information Gain Approach for Feature Selection in Structural Health Monitoring of Bridges

Abstract Recent advancements in structural health monitoring (SHM) of bridges have increasingly relied on the statistical pattern recognition paradigm, where extracting damage-sensitive features from raw data is critical. The large volume of features generated from multiple sensors and signals often includes irrelevant data and noise, which hinders the accurate detection of damage. Effective feature selection is therefore essential to advance bridge SHM from research to practical implementation. Thus, this paper proposes a decision tree-based information gain approach for automated feature selection, designed to enhance sensitivity of features to structural damage. The method is evaluated using data from the Z24 Bridge, covering its undamaged condition and eight damaged state conditions. Features are extracted using frequency- and time-domain techniques. The proposed approach is compared against two other alternatives based on permutation importance and F-statistics, showing a superior trade-off in classification performance, identifying with a smaller and more effective subset of features. An extreme gradient boosting classifier is employed to address missing data, a common challenge in SHM applications. Time-domain features are further explored for real-world bridge applications as an alternative to traditional frequency-domain features. Finally, a sensor-level sensitivity analysis revealed new possibilities for applying tree-based algorithms in SHM, underscoring their practical impact even for damage localization.