Power Control during Remote Laser Welding Using a Convolutional Neural Network.

Area of Science:

• Manufacturing Engineering
• Robotics and Automation
• Materials Science

Background:

• Complex geometries in modern manufacturing necessitate advanced control for stable welding.
• Traditional methods require extensive experimentation to determine optimal welding parameters.

Purpose of the Study:

• To develop an advanced control algorithm for remote laser welding using a convolutional neural network (CNN).
• To improve welding stability and reduce parameter optimization time for complex workpieces.

Main Methods:

• A CNN integrated with a PID controller was employed for laser power control, utilizing optical triangulation feedback.
• Experiments were conducted on AISI 304 metal sheets (1.5 mm thickness).
• Gradient-weighted Class Activation Mapping (Grad-CAM) was used for model interpretability.

Main Results:

• The CNN-PID controller achieved 94% accuracy on test datasets.
• A rapid rise time of less than 1.0 second was observed for achieving full penetration.
• Grad-CAM confirmed the CNN's focus on the weld interaction zone, enabling adaptive control.

Conclusions:

• The developed CNN-based control method offers a highly accurate and efficient solution for remote laser welding.
• Proposed feed-forward improvements can further enhance controller performance and response time.
• This approach addresses the need for advanced control in welding complex geometries.