Welding penetration prediction with passive vision system

Weld pool backside width is an important indicator of welding penetration. However, it is challenging to measure it directly during welding process. In this work, data driven method was established to estimate backside width during GTAW. First, key features related with 3D weld pool surface characters were extracted using computer vision approach. Second, the database which covers wide range of welding conditions was established. Two supervised machine learning method: linear regression and bagging trees were tested on the database. Through feature importance analysis, we found out that the weld pool width, trailing length and surface height (SH) played the major role to predict backside width. The performance of prediction was further improved through feature selection. Finally, the trained model was validated with butt joint welding experiments with satisfactory accuracy. The proposed method can be further applied to build real time close-loop penetration control system.