A Three-Branch Time-Frequency Feature Fusion Method Based on Terahertz Signals for Identifying Delamination Defects...

Composite materials are critical components in advanced equipment such as aerospace, however, delamination defects that readily arise during manufacturing and in service present serious risks to equipment safety. Terahertz non-destructive testing is highly effective for analyzing the internal structure of composite materials, making it an effective approach for precise identification of delamination defects. Current terahertz detection approaches mainly depend on single domain features, making it difficult to capture complementary information from both the time and frequency domains. To address this, a Time-Frequency Feature-fusion Network (TFFN) is proposed. In this network, a three-branch architecture is employed: local transient patterns and pulse-related structural features are extracted by the local time-frequency branch; damage-sensitive frequency bands are focused on by the frequency-domain branch through a channel-space-frequency band attention mechanism; and deep integration of time-frequency features is achieved by the time-frequency fusion branch using Manifold Mixup. Finally, the features extracted from the three branches are adaptively fused via a cross-branch attention mechanism, and defect identification is accomplished by the classifier. Experimental results show that this method achieves accuracies of 98.40% on the glass fiber reinforced polymer (GFRP) dataset and 98.63% on the quartz fiber reinforced polymer (QFRP) dataset, surpassing the best existing method by 2% and 1.25%, respectively. A substantial improvement in both defect identification accuracy and the model's generalization ability for layered structures is thereby achieved.