Non-destructive evaluation tool for monitoring and detection of structural damage by using neural network.

Abstract

This work studies the capability of generalization of Neural Network using vibration based measurement data aiming at operating condition and health monitoring of mechanical systems. The procedure uses the backpropagation algorithm to classify the input patters of a system with different stiffness ratios. It has been investigated a large set of input data, containing various stiffness ratios as well as a reduced set containing only the extreme ones in order to study generalizing capability of the network. This allows to definition of Neural Networks capable to use a reduced set of data during the training phase. Once it is successfully trained, it could identify intermediate failure condition. Several conditions and intensities of damages have been studied by using numerical data. The Neural Network demonstrated a good capacity of generalization for all case. Finally, the proposal was tested with experimental data.