Session: 06-01: Machine Learning and Statistical Methods in NDE I

Paper Number: 134721

134721 - Guided Waves-Based Disbond Detection of Double-Layer Plates Using LSTM Networks 

Abstract: 

Elastic adhesively bonded material structures are of great interest in a wide range of industries, such as aerospace, nuclear, and automobile. Interfacial defects like disbond and deamination are prone to appear in these bonded multilayered structures during both the fabrication process and service life. Guided waves have been verified as a powerful tool for the effective identification of interfacial defects. For example, the location and size of the disbond can be inferred by analyzing the characteristic wave components. However, these diagnostic waveforms may overlap with other futile waves, making it difficult to render accurate damage detection results. Besides, this task highly depends on the experience and judgment of inspection personnel. To reduce the reliance on human expertise, deep learning algorithms provide a promising solution to solve this kind of inverse problem. This paper utilizes the long short-term memory (LSTM) neural network to detect disbond in a double-layer plate.

We consider the double-layer plate consisting of an aluminum substrate with a stainless-steel coating. The sensitive wave mode is initially selected to ensure effective damage interrogation. Subsequently, finite element simulations are conducted for an extensive variety of locations and sizes of disbond to generate the necessary ultrasonic data. The temporal signals of the out-of-plane displacement are extracted from the sensing point. To account for uncertainty and noise in the experimental measurements, Gaussian random noise will be introduced in the sequence data. Then, an LSTM regression model is developed and trained using the time trace signals obtained from the forward computations. Once trained, the neural network is ready to solve the inverse problem by testing the temporal signals to predict the location and size of disbond. Finally, the performance of the network is assessed based on loss and accuracy. This method offers a feasible approach for nondestructive evaluation (NDE) practical application on disbond localization and sizing.

Presenting Author: Junzhen Wang Stevens Institute of Technology

Presenting Author Biography: Dr. Junzhen Wang received his bachelor's degree in Aircraft Design and Engineering from Nanjing University of Aeronautics and Astronautics in 2017, and PhD degree in Mechanical Engineering from Shanghai Jiao Tong University in 2022. He is currently working with his advisor Professor Jianmin Qu as a postdoctoral researcher in Stevens Institute of Technology. His research interests cover guided waves, nonlinear ultrasonics, structural health monitoring, and non-destructive evaluation.

Authors:

Junzhen Wang Stevens Institute of Technology
Jianmin Qu Stevens Institute of Technology