Recurrent Convolutional Neural Network based defect detection in Submerged Arc Welding processes

Detecting defects predictively during the welding process, such as porosity, is of vital importance as it allows for the avoidance of degradation in the quality, durability, and productivity of the weld. Research into predictively identifying these types of defects in Submerged Arc Welding (SAW) is quite limited due to the difficulty of gathering data along the process. This remains a challenge to drive the optimization of the manufacturing of pieces that include such welds as the case of large components like pipes in the oil and gas industry. Therefore, this work addresses this challenge and proposes a methodology based on a deep hybrid neural network called recurrent convolutional neural network (RCNN). This deep learning model is capable of detecting and predicting surface porosity defects in real-time using the continuous voltage electrical signal from the SAW process. The training of the RCNN model involved using various weld beads, some with surface porosity and others without. On the one hand, defects were labeled based on the location of the pores along the weld, while on the other hand, the voltage electrical signals were processed and organized. The proposed framework based on RCNN was tested in other weld beads, where the results were satisfactory with the model achieving a high accuracy rate of around 80% in predictive pore detection. Moreover, the model's processing time was < 10 ms, meeting the requirements for real-time applications.