Neural Networks Trained by Analytically Simulated Damage States

Abstract
Identifying changes in the vibrational signatures of a structure is a promising tool in structural monitoring. Neural networks can be used for this purpose. For a neural network to diagnose damage correctly, it must be trained with successfully diagnosed damage states (learning or training samples). Training samples can be developed over time as actual damage states are experienced by the structure. They can also be obtained from a destructive test program in which the variations in vibrational signatures are recorded. Both of these methods of obtaining learning samples are difficult to implement and make the approach impractical. This paper investigates the feasibility of using analytically generated training samples to train neural networks. These networks, trained with analytically generated states of damage, were used to diagnose damage states obtained experimentally from a series of shaking‐table tests of a five‐story steel frame. The results show that neural networks, trained with analytically obtained sample cases, have a strong potential for making on‐line structural monitoring a practical reality.

This publication has 6 references indexed in Scilit: