A high-precision segmentation method based on UNet for disc cutter holder of shield machine

Area of Science:

• Robotics and Automation
• Computer Vision
• Geotechnical Engineering

Background:

• Robotic disc cutter replacement in shield machines requires accurate visual positioning.
• Underground conditions like low light, dust, and debris hinder recognition accuracy.

Purpose of the Study:

• To develop a robust segmentation model for identifying disc cutter holders under challenging underground conditions.
• To improve the accuracy and efficiency of visual positioning for robotic operations.

Main Methods:

• Proposed a multi-mechanism enhanced UNet model (Res-UNet-CA) incorporating attention mechanisms.
• Utilized a hybrid loss function (dice loss + cross-entropy loss) with the Adam optimizer.
• Conducted experimental comparisons with mainstream semantic segmentation models.

Main Results:

• The Res-UNet model demonstrated superior training efficiency and segmentation accuracy.
• The Res-UNet-CA architecture achieved state-of-the-art metrics: 99.45% accuracy, 98.9% precision, 99.11% recall, 99% F1-score, and 98.63% mIoU.
• Significantly outperformed other semantic segmentation models in prediction quality.

Conclusions:

• The Res-UNet-CA model provides an innovative and effective solution for shield machine disc cutter holder detection.
• This approach enhances the reliability of visual positioning in complex underground environments.
• Offers potential for improved automation and safety in tunneling operations.