Abstract: Weld penetration state is one of the key factors affecting the quality of the weld seam. Reliable sensing and real-time prediction of the weld penetration state is important to improve the quality of the weld and the level of welding intelligence. Laser vision was used to detect the reflected laser streaks in different penetration states of low-frequency pulsed gas tungsten arc welding (P-GTAW). The correlation between the dynamic behavior of laser streak images and the free surface of the melt pool in three penetration states was analyzed by establishing a standard model of the melt pool surface. Developed a GTAW fusion penetration prediction model based on deep learning convolutional neural networks. Experimental results show that：The dynamic behavior of the P-GTAW laser streak has a clear optical correspondence with the melt penetration state on the back side of the melt pool and the surface oscillation mode of the melt pool. The developed end-to-end convolutional neural network model from laser streak image to GTAW backside melt-through state can accurately classify three states with an accuracy of 98.1%, enabling real-time sensing and prediction of weld melt-through state.

Key words: penetration inspection, laser vision method, deep learning, melt pool oscillation, end-to-end neural network, GTAW