Crack-Depth Determination by a Neural Network with a Synthetic Training Data Set

A neural network with an analog output is presented for crack-depth estimation from ultrasonic signals backscattered from a surface-breaking crack in a steel plate. The network has only one response unit and this unit directly reports the crack depth from the measured signals. A completely synthetic data set, spot-checked by comparison with experimental results, is utilized for the training of the network. The synthetic data set has been obtained by solving governing boundary integral equations by the boundary element method. A Gaussian modulated sinusoid has been utilized as incident signal. The architecture of the present network, which is a feedforward three-layered network together with an error back- propagation algorithm, has been discussed in Refs. [1,2].