Deep Learning-Based Defects Detection in Keyhole TIG Welding with Enhanced Vision.

Area of Science:

• Materials Science and Engineering
• Manufacturing Technology
• Artificial Intelligence in Industrial Applications

Background:

• Keyhole tungsten inert gas (TIG) welding offers high efficiency but requires real-time defect monitoring.
• Existing methods struggle with complex defect recognition, particularly distinguishing subtle issues in molten pool images.
• Support Vector Machines (SVMs) have limitations in handling deep feature maps and complex nonlinear relationships inherent in welding defects.

Purpose of the Study:

• To develop and validate a deep learning-based method for real-time defect detection in keyhole TIG welding.
• To enhance the accuracy and efficiency of identifying various weld states, including good welds and critical defects.
• To provide a robust solution for automated quality control in advanced welding processes.

Main Methods:

• Utilized a multi-layer deep neural network trained on a comprehensive dataset of welding images.
• Employed meticulous data preprocessing and augmentation techniques to ensure dataset reliability.
• Implemented a four-class classification task to differentiate between good welds, burn-through, partial penetration, and undercut.

Main Results:

• The deep neural network demonstrated high accuracy in classifying different weld pool images.
• The system achieved real-time performance, crucial for immediate feedback in welding operations.
• Effectively distinguished subtle defects often confused in traditional analysis, including those within the molten pool's deep feature maps.

Conclusions:

• Deep neural networks are superior to SVMs for recognizing complex welding defects due to their ability to process deep feature maps.
• The proposed method offers an effective solution for real-time quality control and defect prevention in keyhole TIG welding.
• This approach significantly advances automated inspection capabilities in modern manufacturing.