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

Area of Science:

• Materials Science
• Artificial Intelligence
• Non-Destructive Testing

Background:

• Detecting critical defects, such as delamination, in thick Carbon Fiber Reinforced Polymers (CFRPs) is challenging.
• Complex microstructures like fiber waviness can significantly reduce the effectiveness of conventional ultrasonic testing (UT).

Purpose of the Study:

• To develop an advanced deep learning framework for reliable defect detection in CFRPs, overcoming limitations of traditional UT.
• To synergistically integrate diverse data streams for enhanced defect identification in complex composite materials.

Main Methods:

• A dual-stream deep learning framework incorporating an AttentionFusion module was proposed.
• The framework integrates spatial-morphological data from B-scan images with multi-angle scattering signatures from raw Full Matrix Capture (FMC) data.
• A YOLOv8-based detector was employed for defect identification, leveraging adaptive weighing of inspection streams.

Main Results:

• A 25.8% relative mean Average Precision at 50% IoU (mAP50) improvement over a single-stream baseline was achieved.
• The framework demonstrated a 29.9% performance gain on challenging wavy-fiber composite samples.
• Ablation studies confirmed the significant contribution of the AttentionFusion mechanism, and attention maps visualized the decision-making process.

Conclusions:

• The proposed dual-stream deep learning framework offers a more accurate and transparent solution for automated non-destructive testing (NDT) in complex aerospace composites.
• Leveraging raw FMC data, often discarded, enhances defect detection reliability.
• The AttentionFusion module is crucial for synergistically integrating multi-modal data for improved NDT performance.