Abstract
A deep learning-based approach has been developed for real-time weld pool semantic segmentation and predicting the depth of penetration (DOP) during activated tungsten inert gas (A-TIG) welding of 10 mm thick 316LN stainless steel plate. A custom U-Net framework, with VGG19 as the encoder, was used for semantic segmentation of weld pool images in real-time. Model development involved three main steps: semantic segmentation of the weld pool using a convolutional neural network (CNN), surface profile measurement and DOP prediction using a back-propagation neural network (BPNN). The CNN model exhibited excellent accuracy on the segmented images. Key features were extracted from the segmented images and along with the welding current were used as inputs to the BPNN while the measured DOP values were used as output to train the BPNN. All the three steps were combined to predict the depth of penetration in real-time. The model’s execution time was found as 110 ms. Replacing the encoder with EfficientNet-B0 reduced the execution time to 70 ms. A Python-based graphical user interface (GUI) was developed to monitor the depth of penetration as a function of weld distance in real-time. Excellent agreement was achieved between the predicted and measured values.