Random forest-based real-time defect detection of Al alloy in robotic arc welding using optical spectrum

Aluminum alloy of arc welding is the main technology for the key components manufacturing in aerospace, nuclear power, ship and so on. Real-time weld defects detection is still challenging due to the complexity and diversity of weld defects. Arc optical Spectroscopy emission is the key information generated during arc welding process. However, how to select the effective spectrum feature from high dimension of arc spectrum is crucial for improving the accuracy of defects recognition. This paper proposed an on-line defects detection method for aluminum alloy in robotic arc welding based on random forest and arc spectrum. Firstly, preprocessing of arc spectrum was carried out before 50 features were extracted. Then, a quantitative index of feature importance is proposed based on mean decrease accuracy and mean decrease Gini to reduce the feature redundancy. Six spectral features were selected and analyzed in terms of the construction pattern. Furthermore, the defect identification model was established based on random forest and the optimal feature subset. Comparing with RBF and BP models, it can achieve better performance in identifying three typical defects, including incomplete penetration, burn-through and porosity. This paper can provide some guidance for data mining of optical information and intelligent manufacturing.