Classification, Localization and Quantization of Eddy Current Detection Defects in CFRP Based on EDC-YOLO

Area of Science:

• Materials Science
• Nondestructive Testing
• Artificial Intelligence

Background:

• Accurate defect detection is crucial for the reliability of carbon fiber-reinforced plastic (CFRP) components.
• Eddy current nondestructive testing (ECNDT) is a key technique for inspecting these materials.
• Existing methods may face challenges in identifying and quantifying diverse defect types.

Purpose of the Study:

• To investigate the identification and measurement of common CFRP defects (cracks, delamination, impact damage).
• To enhance the You Only Look Once (YOLO) model for improved ECNDT performance.
• To develop an improved Eddy Current YOLO (EDC-YOLO) model for defect quantification.

Main Methods:

• Utilized the You Only Look Once (YOLO) model as a base for defect detection.
• Integrated Transformer-based self-attention mechanisms and deformable convolutional sub-modules to address multi-scale feature limitations.
• Incorporated CBAM for global feature extraction and Wise-IoU loss function for performance enhancement.
• Developed the improved Eddy Current YOLO (EDC-YOLO) model.

Main Results:

• The EDC-YOLO model demonstrated effectiveness in identifying and quantifying cracks, delamination, and impact damage in CFRP.
• Model enhancements led to a 4.4% increase in mAP50 for defect detection.
• The study provided insights into the correlation between impact damage characteristics and energy levels.

Conclusions:

• The EDC-YOLO model offers a robust solution for defect identification and quantification in industrial ECNDT of CFRP.
• The integrated attention mechanisms and loss function significantly boost detection accuracy.
• This AI-driven approach enhances the reliability and efficiency of CFRP inspection.