Attention-fused dual-stream learning for defect classification in thick aerospace CFRPs with complex microstructures...

Reliable detection of critical defects in thick Carbon Fiber Reinforced Polymers, particularly delamination, is a significant challenge. This task becomes severely complicated when complex microstructures such as fiber waviness limit the effectiveness of conventional ultrasonic testing. To address this, a dual-stream deep learning framework with an efficient and interpretable AttentionFusion module is proposed, which synergistically integrates spatial-morphological information from B-scan images with physics-rich, multi-angle scattering signatures from raw Full Matrix Capture data. Through the adaptive weighing of both static B-scan and dynamic multi-angle inspection streams, the most salient features are leveraged by a YOLOv8-based detector for defect identification. When validated on a dataset consisting of 2776 samples, a 25.8% relative mAP50 improvement over a single-stream baseline was achieved, with this margin increasing to 29.9% on challenging wavy-fiber samples. The critical contribution of the AttentionFusion mechanism was confirmed via ablation studies. Furthermore, the framework's decision-making process was elucidated through visualization of attention maps, enhancing its transparency. By leveraging raw Full Matrix Capture data often discarded in traditional pipelines, a more accurate and trustworthy solution for automated nondestructive testing in complex aerospace composites is provided.