A weld point cloud recognition method based on an improved Light Gradient Boosting Machine.

Area of Science:

• Robotics and Automation
• Computer Vision
• Machine Learning

Background:

• Accurate weld identification is crucial for quality control and automated processes like grinding.
• Weld point clouds present challenges due to their irregular nature and lack of explicit topological structure.

Purpose of the Study:

• To develop a robust framework for classifying individual points in weld point clouds as either weld bead or base metal.
• To enhance the accuracy and interpretability of weld recognition models for industrial applications.

Main Methods:

• Formulated weld recognition as a binary point-wise classification task.
• Employed neighborhood-based geometric feature extraction and compared baseline machine learning models (RF, XGBoost, LightGBM).
• Optimized LightGBM hyperparameters using metaheuristic algorithms (ALA, AE, SFOA) and analyzed feature importance with SHAP.

Main Results:

• LightGBM demonstrated superior baseline performance at a 1.5 mm neighborhood radius.
• The AE-LightGBM model achieved the best overall performance, validated by statistical significance and convergence analysis.
• SHAP analysis provided insights into feature contributions, enhancing model interpretability.

Conclusions:

• The proposed AE-LightGBM framework offers an effective solution for robot-based weld recognition using 3D vision and supervised learning.
• This approach enhances the precision required for automated weld quality inspection and grinding operations.