Neural network assisted crack and flow identification in transient dynamics

Crack and flaw identification problems in two-dimensional elastomechanics are numerically studied in this paper. The mechanical modelling is based on boundary element techniques, with special care of hypersingular issues for the cracks. The possibility of partially or totally closed cracks (unilateral contact effects) is taken into account by linear complementarity techniques. Backpropagation neural networks are used for the solution of the inverse problems. For dynamical problems, a suitable preprocessing of the input data enhances the effectiveness of the procedure. For the two-dimensional examples presented here, the proposed method has similar performance for classical crack and flaw identification problems. The identification of unilateral cracks is a considerably more difficult task, which nevertheless, can also be solved by the same method, provided that a suitable dynamical test loading is applied.