Influence of input parameters on the performance of an artificial neural network used to detect structural damage

Structural damage detection is a very important research topic and, currently, there are not specific tools to solve it. A promising tool that can be used is the artificial neural network, ANN, which can deal with hard problems. This paper uses a back propagation ANN with Bayesian regularization training to locate and quantify damage in truss structures. The input parameters corresponded to natural frequencies combined with shape modes, modal flexibilities or modal strain energies. The ANN was trained by considering only simple damage scenarios, random multiple damage scenarios or a combination of them. The results are shown in terms of the percentage of cases in which the ANN trained achieves a determined performance in assessing both the damage extension and the presence of damaged elements. The best performance for the ANN is obtained by using modal strain energies and multiple damage scenarios.