Structure damage detection based on random forest recursive feature elimination

•A two-stage feature selection method in structural damage detection is proposed.•RF–RFE can select most important damage features.•Not all the WPD features are beneficial for the damage detection.•Using RF–RFE can get better performance than RF–VIA.•Using the fewer features selected from higher-level WPD can achieve better performance than using all lower-level WPD features.

Feature extraction is a key former step in structural damage detection. In this paper, a structural damage detection method based on wavelet packet decomposition (WPD) and random forest recursive feature elimination (RF–RFE) is proposed. In order to gain the most effective feature subset and to improve the identification accuracy a two-stage feature selection method is adopted after WPD. First, the damage features are sorted according to original random forest variable importance analysis. Second, using RF–RFE to eliminate the least important feature and reorder the feature list each time, then get the new feature importance sequence. Finally, k-nearest neighbor (KNN) algorithm, as a benchmark classifier, is used to evaluate the extracted feature subset. A four-storey steel shear building model is chosen as an example in method verification. The experimental results show that using the fewer features got from proposed method can achieve higher identification accuracy and reduce the detection time cost.