Neural network-based inverse analysis for defect identification with laser ultrasonics

This paper describes an application of the neural network-based inverse analysis method to the identification of a surface defect hidden in a solid, using laser ultrasonics. The inverse analysis method consists of three subprocesses. First, sample data of identification parameters versus dynamic responses of displacements at several monitoring points on the surface are calculated using the dynamic finite-element method. Second, the back-propagation neural network is trained using the sample data. Finally, the well-trained network is utilized for defect identification. Fundamental performance of the method is examined quantitatively and in detail, through both numerical simulations and laser ultrasonics experiments. Locations and depths of vertical defects are successfully estimated within 12.5% and 4.1% errors relative to the specimen thickness, respectively.