Neural networks for delamination flaw detection in FRP laminated composite plates

In this study, the detection of delamination flaws in laminated composite plates is carried out using artificial neural networks (ANN) in a two-level cascading manner. The three damage parameters detected using ANN are the size of the delamination, its vertical location (across the plate thickness) and horizontal location (along the plate surface). The numerical data in the form of frequency domain Green's function for the displacement response on the surface of the plate containing the delamination flaw is generated first using an available numerical method. Pseudo-experimental data is generated adding artificial random noise into the numerical data. At the first level, a counterpropagation neural network (CPN) is trained for qualitatively classifying the damage parameters using the numerical data generated above. Next, a second level back-propagation network (BPN) is used for each subclass to quantify the damage parameters. An overlapping data set is used for the training of each class of the second level network. As a result, any pattern misclassified by the CPN due to its closeness to the boundary of any two classes is still quantified correctly. By feeding pseudo-experimental data to the trained networks, it is seen that the classification success rate and noise tolerance level of CPN is excellent. The quantification of damage by the second level BPN is also good. It is possible to stop after the first level if only a qualitative assessment of the damage and its approximate location is required. These cascaded networks show promise in providing a successful delamination damage detection tool.