Abstract

This study aims to propose an Extreme gradient boosting (XGBoost)-driven metamodel for damage detection of beams using free vibration signals optimally measured at a limited number of sensors based on reduced-order model. Instead of utilizing an exceedingly huge and complicated learning model constructed by big data, this methodology only builds a learning model via a series of smaller and simpler XGBoost models with only small and moderate data. The sensor placement locations are found by adaptive hybrid evolutionary firefly algorithm (AHEFA) via the optimization problem based on a reduced-order model. Then, to obtain data for feeding into the above learning models, the isogemetric analysis (IGA) combined with the third-order shear deformation (TSDT) is employed. In which, the eigenvector values at a number of degrees of freedom (DOFs) corresponding to the optimized sensor positions are treated as the input data, whilst the randomly assumed damage ratios of beam elements are considered as the outputs. A modal strain energy-based index (MSEI) is applied to remove low-risk elements before employing the suggested metamodel built by a series of XGBoost models. A simply supported beam with two damage scenarios is investigated. The results obtained by the present methodology have shown the reliability and efficiency in identifying the locations and ratios of damaged elements in beam structures.