1D-convolutional neural network for detection of disbonding in a stiffened aluminum panel using Lamb wave responses

Abstract

Stiffened panels enhance stability and load-bearing capacity, but issues like stiffener disbonding stemming from factors such as insufficient bonding or environmental exposure, pose significant risks to structural performance and safety. This study adopts a dual-level investigation approach involving both the detection of disbond and predicting the extent of disbonding in a stiffened aluminum panel using two distinct one-dimentional convolutional neural networks (1D-CNNs) models. In the initial phase, a CNN-based classifier model (CNN-C) is employed to identify the presence of disbonding, treating it as a binary classification task. Subsequently, the focus shifts to the second phase, where the width of disbonding is predicted through a CNN-based regression model (CNN-R). Lamb waves from an experimental setup with four different types of stiffened aluminum panels constitute the dataset. These panels exhibit varying disbonding conditions, encompassing instances without disbonding to various widths of disbonding induced by Teflon strips. While conventional CNNs are traditionally designed for image processing, here, both the proposed CNN models are tailored to process 1D Lamb wave responses as input. CNN-C generates a probability distribution to classify the classes representing the presence and absence of disbonding, while CNN-R produces a continuous value representing the width of disbonding. The developed models are evaluated using an unseen test dataset, which remains undisclosed during their training. The CNN-C model achieves a perfect 100% accuracy in classifying the state of stiffened panels, while the CNN-R model effectively predicts disbonding width with a mean absolute error of 0.096 mm and an average relative error of 0.56% on the unseen test dataset.