Automated crack detection of train rivets using fluorescent magnetic particle inspection and instance segmentation

Area of Science:

• Railway engineering
• Non-destructive testing
• Artificial intelligence

Background:

• Railway rivets are critical components prone to damage, with cracks reducing load-bearing capacity and increasing failure risk.
• Fluorescent magnetic particle flaw detection (FMPFD) is standard for train fasteners, but manual inspection is inefficient and error-prone.
• Detecting cracks in fluorescent rivet images is challenging due to their dense, multi-scaled, and uninstantiated nature.

Purpose of the Study:

• To develop an intelligent system for automatic crack detection in railway rivets.
• To address the challenges of dense, multi-scaled, and uninstantiated cracks in fluorescent rivet images.
• To improve the reliability and efficiency of rivet crack inspection.

Main Methods:

• Instance segmentation approach for automatic crack detection.
• Novel decentralized target center and low overlap rate labeling method.
• Gaussian-weighted correction for post-processing dense crack areas.
• Efficient channel spatial attention mechanism for multi-scale crack feature extraction.
• Multi-task feature learning for improved detection in uninstantiated crack regions.

Main Results:

• The proposed methods significantly outperform baseline and state-of-the-art algorithms.
• Achieved a recall rate of 86.4% and a mean Average Precision (mAP0.5) of 90.3%.
• Developed a versatile single-coil non-contact composite magnetization device for various rivet shapes.

Conclusions:

• The developed intelligent system effectively detects cracks in railway rivets, enhancing safety.
• The novel methods successfully address challenges posed by complex crack characteristics.
• The system offers a valuable advancement for automated inspection of train fasteners.