Collaborative Optimization of High-Resolution Representation and Miss-Sensitive Supervision for Aero-Engine...

Area of Science:

• Aerospace Engineering
• Materials Science
• Computer Vision

Background:

• Aero-engine blades face extreme operational conditions, leading to micro-cracks.
• Detecting these micro-cracks in borescope images is challenging due to their small size, low contrast, and complex image backgrounds.
• Conventional methods suffer from high missed-detection rates.

Purpose of the Study:

• To develop an advanced deep learning framework for accurate aero-engine blade micro-crack detection.
• To improve the robustness and reliability of automated inspection systems for critical aerospace components.

Main Methods:

• An improved YOLOv11 framework was proposed, incorporating P1/P2 shallow high-resolution detection branches.
• Focal Loss was used to address foreground-background class imbalance.
• Object-level Tversky Loss and a hard mining strategy were employed to enhance learning from difficult samples and reduce false negatives.

Main Results:

• The proposed model achieved high performance metrics: Precision (0.9981), Recall (0.9606), F1-score (0.9790), mAP50 (0.9781), and mAP50-95 (0.6938) on a real dataset.
• Significant improvements in detection accuracy, localization precision, and robustness were observed compared to the baseline YOLOv11.
• The enhanced model demonstrated superior performance in complex borescope inspection scenarios.

Conclusions:

• The improved YOLOv11 framework effectively addresses the challenges of aero-engine blade micro-crack detection in borescope images.
• The proposed enhancements lead to more accurate and reliable automated defect detection, crucial for aviation safety and maintenance.
• This work offers a robust solution for real-world industrial inspection tasks involving fine crack detection.