Numerical investigations of deep learning-assisted delamination characterization using ultrasonic guided waves

Abstract

Ultrasonic guided waves have been applied extensively for nondestructively detecting delamination in multilayered structures. However, traditional guided-wave-based nondestructive evaluation (NDE) techniques require highly skilled users to interpret the complex wave field scattered by the delamination. To overcome this challenge, this study proposes an approach that combines a deep-learning (DL) neural network with traditional ultrasonic NDE techniques. The NDE technique used here is based on guided waves generated and received in a transmitter-receiver configuration. A 2D finite element analysis (FEA) model is constructed to simulate the guided wave interactions with a delamination between two metallic layers, which yields both the training and testing data. The tailored DL neural network is a convolutional neural network (CNN) combined with bi-directional long short-term memory (BiLSTM). This hybrid neural network is trained by a set of pulse-echo or pitch-catch time-domain data. Once trained, the DL neural network predicts the location of delamination using the recorded pulse-echo time-domain signals as input, and the length of delamination using the recoded pitch-catch time-domain as input. This process of nondestructively characterizing the location and size of delamination can be carried out automatically without the need to analyze the complex wave fields in the ultrasonic tests. The predicted results on both within-range and out-of-range unseen data demonstrate that the proposed technique has tremendous potential for characterizing delamination in practical NDE and structural health monitoring (SHM) applications.