Terahertz Automatic Localization and Imaging for Multiscene Delamination Damages of Composites Using Multilabel Attention Fusion Network

Multiscene delamination damages caused by the multiinterface characteristics of composites, including single delamination damage and overlapping compound delamination damage types, have a huge impact on the structural integrity and dynamic service performance of composite structure. Terahertz nondestructive testing (THz-NDT) technique, as a novel and high-resolution measurement method, has emerged great potentials in the integrity assessment of composite. However, current damage extraction methods in THz-NDT, including manual feature-based and data-driven approaches, are devoted to single delamination damage type, which brings a big challenge for overlapping compound delamination damages due to their complexity and unpredictability. In order to address the general characterization problem for multiscene delamination damages, in this work, a general multilabel THz intelligent characterization framework is proposed. The key is to simplify the general characterization problem of multiscene delamination damages into a multilabel classification task. This framework contains both specially designed core components: the multilabel attention fusion network (MLAF-Net) and the multilabel component encoding (MLCE) THz imaging strategy. Both components are used to perform the accurate localization and high-resolution imaging of multiscene delamination damages, respectively. Finally, a series of comparison and ablation experiments are carried out to demonstrate the advantages of the framework, which will provide a novel and general insight and paradigm for automatic localization and identification of complex multiscene damages in THz-NDT or other NDT fields.