Transfer Learning-Based Multi-Sensor Approach for Predicting Keyhole Depth in Laser Welding of 780DP Steel

Area of Science:

• Materials Science and Engineering
• Manufacturing Processes
• Artificial Intelligence in Engineering

Background:

• Penetration depth is crucial for joint strength in laser welding, but challenging to monitor in real-time.
• Optical coherence tomography (OCT) offers real-time measurement but requires precise calibration.
• Keyhole dynamics in laser welding complicate in-situ monitoring of weld quality.

Purpose of the Study:

• To develop and evaluate deep learning models for estimating penetration depth in 780 dual-phase (DP) steel laser welding.
• To compare the performance of uni-sensor and multi-sensor deep learning approaches.
• To investigate the effectiveness of transfer learning with pre-trained Convolutional Neural Network (CNN) architectures.

Main Methods:

• Developed uni-sensor models using coaxial weld pool images and multi-sensor models incorporating spectrometer signals.
• Employed transfer learning with pre-trained CNNs (MobileNetV2, ResNet50V2, EfficientNetB3, Xception).
• Used OCT signals as the ground truth for penetration depth during model training and validation.

Main Results:

• Uni-sensor models (without fine-tuning) achieved R² values from 0.502 to 0.681 and MAE from 0.152 mm to 0.196 mm.
• Fine-tuning uni-sensor models led to a decrease in R² (>17%) and an increase in MAE (>11%).
• Multi-sensor models demonstrated superior performance with R² ranging from 0.900 to 0.956 and MAE from 0.058 mm to 0.086 mm.

Conclusions:

• CNN transfer learning models show significant potential for predicting laser welding penetration depth in 780DP steel.
• Multi-sensor fusion significantly enhances prediction accuracy compared to uni-sensor models.
• The developed models offer a promising non-contact method for real-time weld quality assessment.