Rotor crack identification based on neural networks and modal data

A model-based procedure for rotor crack localization and assessment is presented in this paper. The procedure is applied to a small-size test rig provided with a notch. Both the position and depth of the notch are estimated through a neural network on the basis of the first four natural frequencies of the rotor. A 3-D finite element model is used to generate the data for training the net. One of the contributions of this paper consists of a meshing procedure that reduces the systematic errors of the model, which have a significant influence in identification accuracy. A sensitivity analysis has been carried out for any size and position of the notch, which constitutes another original contribution in this field. In the studied case, the proposed procedure is able to predict both the position and depth of the notch when the notch depth is greater than 20 % of the rotor diameter. The sensitivity analysis reveals that there are blind spots in the rotor as regards notch identification.