Terahertz Characterization for Debonding Defects of Composites based on the Group-CNN Network

Damage detection in composite materials is crucial for monitoring the condition of components throughout their lifecycle and enabling the maintenance and potential replacement. In practical manufacturing and in-service processes of Glass Fiber Reinforced Polymer (GFRP) composites, unpredictable damage such as delamination often occurs on the surface or inside due to the varying uncertainties in the manufacturing process and service environment. Terahertz (THz) non-destructive testing technology, as a new detection method, provides a new approach for characterizing delamination defects in GFRP composites. Nevertheless, conventional techniques for complex signal processing in THz testing prove to be arduous, time-intensive, and necessitate a substantial degree of prior knowledge to mitigate the adverse impact of interference phenomena, such as noise and dispersion, on THz signals. Therefore, we propose an effective end-to-end 3D THz characterization system based on deep learning methods to achieve automatic localization and imaging of delamination defects in GFRP composites without additional signal processing or prior knowledge. In the localization process, a Deep Separable Convolutional Neural Network (DW-CNN) using group convolutions is designed to accurately classify THz signals. In the imaging process, a classification encoding strategy is employed to obtain two-dimensional and threedimensional information about layered defects based on the classification results. Finally, the system effectiveness is validated through a series of experiments, which provides a new solution for intelligent and automated THz characterization of delamination defects within composite materials.