Diagnosis of structural cracks using wavelet transform and neural networks

Abstract

This paper proposes an approach to detecting and characterizing structural cracks emanating from rows of rivet holes in thin metallic plates using lamb wave, wavelet transform and neural networks. When lamb waves propagate through rows of holes, there exist strong wave interferences among different transmitted waves which make health diagnosis complicated. An active sensing network is mounted on the plate and wavelet transform is used to extract a robust and effective feature called energy ratio change from time domain signals. The relationship between energy ratio change and the crack characteristics is analyzed and the effective detection paths are found out. Neural networks are then developed using the feature to diagnose health condition in stages. One neural network is first used to diagnose plate integrity. If cracks are detected, then the second neural network is called to determine their locations. The method is first examined in simulation data, then the NNs trained by simulation data are used to diagnose real plates. The results shows that the method can effectively detect cracks and identify their locations in both simulation and experimental data, hence demonstrating the feasibility of using the method to develop built-in real time intelligent diagnosis systems.