Structural damage identification through machine learning approaches using FRFs

Resumo

In this paper, experimental tests and numerical simulations are conducted to evaluate the performance of different models for structural damage identification and quantification. For this purpose, an aluminum beam in Laboratory conditions is utilized as a test structure. Firstly, impact tests are performed to identify the modal parameters and frequency response functions (FRFs) of the healthy structure. Then,
different damages are induced in the beam by means of rectangular notches, and FRFs from each damage scenario are measured. Meanwhile, a simplified numerical model of finite elements of the beam is developed and calibrated with respect to experimental data. The calibrated model is used to generate a set of simulations representing the different damage scenarios induced experimentally. The damage is introduced in the numerical model by reducing the cross-sectional area. Normalized FRF amplitudes are used as the damage indexes. To increase the predictive capability of the models, uncertainties are introduced considering the FRF amplitudes as random variables. Afterward, different datasets are constructed and several well-established machine learning classifiers such as Decision Tree, SVM and KNN are trained to perform damage identification and quantification. Finally, experimental data measured on the damaged beam are used as input variables to evaluate the prediction capacity of the trained classifiers. Undamaged and damaged data are correctly classified by most of the classifiers. However, to quantify the degree of damage some shortcomings are found.