A Deep Learning-Based Machine Vision System for Online Monitoring and Quality Evaluation During Multi-Layer...

Area of Science:

• Manufacturing Engineering
• Robotics and Automation
• Materials Science

Background:

• Multi-layer multi-pass welding is critical in industries like nuclear power and shipbuilding.
• Welding distortion and defects remain significant challenges, necessitating robust monitoring and quality control.
• Dynamic adjustment during welding is essential for maintaining high-quality output.

Purpose of the Study:

• To develop a machine vision system for real-time monitoring and surface quality evaluation in multi-layer multi-pass welding.
• To enable dynamic control of the welding plan through cross-section modeling and distortion measurement.
• To create a defect inspection dataset for training and validating the quality control algorithms.

Main Methods:

• Utilized line scanner and infrared camera sensors for data acquisition.
• Developed cross-section modeling from line scanner data to measure distortion.
• Applied a normal map approach combined with deep learning for surface defect inspection, addressing material color variations.
• Integrated a burn-through defect detection algorithm and weld pool temperature monitoring.

Main Results:

• Achieved a mean average precision of 0.88 for surface defect inspection using the deep learning approach.
• Successfully measured distortion and enabled dynamic control of the welding plan.
• Demonstrated effective monitoring of weld pool temperature and detection of burn-through defects.
• Integrated the system into a graphical user interface for visualizing welding progress.

Conclusions:

• The proposed machine vision system provides a solid foundation for monitoring multi-layer multi-pass welding.
• The system demonstrates potential for the development of automatic adaptive welding systems.
• Enhanced quality control through real-time monitoring and defect detection is achievable.