Structural damage identification utilising PCA-compressed frequency response functions and neural network ensembles

This paper presents a damage detection method that utilises FRF data to identify damage in beam structures. The proposed method uses artificial neural networks (ANNs) to map changes in FRFs to damage characteristics. To obtain suitable patterns for ANN inputs, the size of the FRFs is reduced adopting Principal Component Analysis (PCA) techniques. A hierarchy of neural network ensembles is created to take advantage of individual differences from sensor signals. To simulate field applications, the time history data are polluted with white Gaussian noise. The method involves finite element modelling of undamaged and damaged steel beams. By performing transient analysis with the numerical beams, the time histories are obtained and subsequently polluted with different levels of white Gaussian noise. FRFs are determined and compressed utilising PCA techniques. The PCA-reduced FRFs are then used as input patterns for training and testing of neural network ensembles giving the characteristics of the damage.