Real-time identification of molten pool and keyhole using a deep learning-based semantic segmentation approach in penetration status monitoring

Due to the complexity and diversity of disturbances in the monitoring signals, it is still a challenge to accurately real-time monitor penetration statues in laser welding manufacturing. In this study, an innovative penetration status diagnosis strategy based on two deep learning methods is proposed. A U-net based image processing method is proposed to semantically segment the welding process monitoring images to eliminate the multiple interferences. Analysis results show that this method can extract the most accurate molten pool and keyhole contours. Then, a lightweight convolutional neural network (CNN) model with optimized network structure and input image size is established for predicting the penetration status. The comparison with other deep learning models shows that the established CNN model has the optimal model size and monitoring accuracy. The validation results on laser welding show that the proposed monitoring method has strong robustness and generalization capability, and the prediction accuracy can reach 98.68%.