A neural network for crack sizing trained by finite element calculations

Abstract A neural network is combined with a self-compensating ultrasonic technique to size cracks emanating from rivet holes. The network is trained with a combination of experimental and synthetic data. A 2-D finite element (FE) method is used to generate synthetic data for scattering of ultrasound by cracks emanating from rivet holes in a thin aluminium sheet. Both the back and forward scattered fields are calculated at several positions parallel to the crack line. An experimental pulse from an ultrasonic transducer is applied as input excitation to the FE grid. The FE program generates a set of back- and forward-scattered signals for cracks of various lengths in the range; 0.25–3.00 mm with length increments of 0.25 mm. Data obtained in this way are used to train the neural network (NN) classifier that categorizes the data according to crack length. Once the network is trained, its performance is tested on self-compensated experimental data obtained by ultrasonic testing of specimens containing cracks of lengths in the range 0.50–3.00 mm with length increments of 0.50mm. The use of FE modelling to train the NN eliminates the need for a large number of experiments.