Automated characterization of debonding based on ultrasonic guided waves and a simulation-trained deep neural network

This article proposes an automated nondestructive evaluation (NDE) technique for debonding characterization using ultrasonic guided waves and a simulation-trained deep neural network. This technique utilizes guided waves in a transmitter–receiver configuration and leverages their interaction with a debonding between two metallic layers. First, a multitude of two-dimensional finite element simulations are conducted to obtain time-series pulse-echo and pitch-catch debonding responses. These signals serve as training data for a hybrid neural network that combines a convolutional neural network (CNN) with a bi-directional long short-term memory (BiLSTM) layer. Once trained, this deep-learning model is able to automatically characterize the location and size of debonding damage by inputting either simulated or experimentally measured guided wave signals. The developed deep-learning model is validated by conducting guided wave active sensing experiments on a pristine plate and four debonding specimens with various debonding locations and sizes. These experimental results demonstrate that the developed neural network, once trained by simulated data, is capable of accurately characterizing debonding sizes. These findings indicate that the proposed technique has tremendous potential for characterizing interfacial debonding in practical NDE and structural health monitoring (SHM) applications.