Graph-in-Graph Convolutional Network for Ultrasonic Guided Wave-Based Damage Detection and Localization

Abstract

Interpretation of guided wave signals is a central challenge for ultrasonic guided wave-based damage detection and localization technology. Because of the complexity of the guided waves that are scattered from structural damage, existing guided wave-based damage detection methods cannot be used to extract the relationship information hidden in the guided waves for use in damage detection and localization. A graph-in-graph convolutional network is thus proposed for guided wave-based damage detection and localization that constructs spatial–temporal feature representations of the guided wave signals and interconnects them into a global graph to indicate the inherent differences among these signals. By converting the guided wave characteristics into structural and topological information in non-Euclidean space, the proposed method correlates the global graph with the damage location directly and achieves greater damage detection accuracy with fewer training data. Validations are performed using two different experimental datasets, which were collected from aluminum plates and a composite laminate. The results indicate that the proposed method achieves superior performance with high accuracy and stability for even limited and imbalanced datasets acquired with only three transducers.