Application of evolutionary neural network in impact acoustics based nondestructive inspection of tile-wall

The degradation of the bonding of the tile-wall structure has emerged as an urgent issue in metropolitan cities, with more and more tile-dropping accidents caused. In order to ascertain the bonding integrity of tile-walls, impact acoustics features obtained from the sound signals generated by controlled impact are employed to develop a rapid and effective nondestructive inspection technique. To facilitate the automatic interpretation, the multilayer artificial neural network (ANN) is used as a cost-effective classifier superior to traditional statistics methods. Nonetheless, the classical gradient descent based backpropagation (BP) training strategy of the multilayer neural network faces certain drawbacks, e.g., very slow convergence, easily getting stuck in a local minimum. In this paper, an evolutionary algorithm based training method is developed to train the ANN and perform the automatic classification of bonding integrity. The design, feature extraction approach, training and application of the proposed evolutionary neural network are presented. The classification results obtained experimentally from prepared sample slabs are presented and compared with that with BP algorithm, demonstrating the validity of the proposed methodology.