Weld Defect Detection in Laser Beam Welding Using Multispectral Emission Sensor Features and Machine Learning

Area of Science:

• Materials Science and Engineering
• Manufacturing Processes
• Data Science and Machine Learning

Background:

• Laser beam welding (LBW) is prone to defects like pore formation due to complex material-laser interactions.
• Interpreting electromagnetic emissions during LBW is challenging for direct defect detection.
• Existing methods lack efficient real-time weld quality assessment.

Purpose of the Study:

• To develop a data-driven framework for interpreting electromagnetic emissions in LBW for defect detection.
• To implement both supervised and unsupervised learning approaches for quantitative weld monitoring.
• To enhance the interpretability of photonic data for industrial weld quality assessment.

Main Methods:

• Conducted 81 welding experiments, recording real-time emission data across 42 spectral channels.
• Extracted statistical, temporal, and shape-based features, reducing dimensionality with Principal Component Analysis (PCA).
• Applied Long Short-Term Memory (LSTM) for supervised learning and Isolation Forest for unsupervised anomaly detection.

Main Results:

• The LSTM model achieved low error rates (MSE: 0.0029, MAE: 0.0288) on the testing set.
• Isolation Forest demonstrated high performance in detecting anomalous welds (80% accuracy, 85.7% precision).
• The framework successfully enhances the interpretability of 4D photonic data from LBW processes.

Conclusions:

• The proposed data-driven framework effectively interprets electromagnetic emissions for weld defect detection in LBW.
• The framework enables both post-process analysis and potential real-time monitoring of weld quality.
• This approach offers a scalable solution for industrial applications requiring robust weld quality assessment.