Selection of Arc Spectrum Features and Defect Recognition in GTAW Based on Random Forest

Because of the high redundancy of welding spectrum, the efficiency of welding recognition model cannot meet the requirements. Aiming at these problems, an on-line detection method for multiple welding defects is proposed in this paper. The background spectrum of the original spectrum is removed in order to remove trend items, and the multi-dimensional line ratio characteristics are extracted. Then, the feature importance index is constructed based on the MDA (Mean Decrease Accuracy) and MDG (Mean Decrease Gini). The feature selection is carried out by quantitatively evaluating the multi-dimensional spectral features, and the important spectral features are selected and analyzed. Normal and three types of defects, including incomplete penetration, burn through, and porosity, are distinguished by a random forest model effectively. By comparing the results of different feature recognition, the feature selection effectively removes the useless features, redundant features, and improves the computational efficiency of the subsequent models. The results of the important spectral feature analysis can provide guiding significance for the subsequent feature line selection. Comparing with RBF and BP neural network, the random forest model achieves higher identification rate, more stable results, and can be applied for on-line detection of welding defects.