Damage Detection Using Neural Networks: An Initial Experimental Study on Debonded Beams

Abstract

Frequency response data obtained from a pieoelec tric actuator/sensor pair bonded to a composite/aluminum beam structure with a debond between the interface is used to train an ar tificial neural network by backpropagation to identify the severity and presence of a delamination. The PZT actuator/sensor pair is so arranged that the damage site lies between the actuator and sensor. The damage consists of an artificially created de-bonding between an aluminum beam and a bonded composite patch. The experi mentally obtained transfer function data in the form of a magnitude and phase, over a specified frequency range, is obtained from a sig nal analyzer. The training process consists of training the network with several fully bonded specimens and several debonded speci mens with various sized damage. The effectiveness of several different configurations of the network applied to this problem is investigated. The neural network after training on a limited number of training data is able to identify the damaged specimens with substantial accuracy.