A Neural Network Approach for Detection of Damage in a Vibrating Beam

This study investigates the potential of using measured data of modal frequency for detecting the location and the size of defects in a vibrating beam. The experimental layout included a beam on which defects were emulated via masses attached to the beam in user-defined locations. The beam was subjected to forced vibration using a wide bandwidth white noise input. The measured natural frequencies of the beam's first five modes of vibration, the location and the size of damage were employed in training a neural network (NN). Neural networks present a viable computational method, with both pattern recognition and prediction capabilities for dynamic system response. A NN for the direct problem was designed, when the damage characteristics were known and the modal frequencies were predicted. A NN for the inverse problem when location and size of damage were predicted based on the measured modal frequencies, was also built. The performance and prediction capabilities of both NNs are assessed.