Do We Need a New Foundation to Use Deep Learning to Monitor Weld Penetration?

Abstract

Deep learning has been successfully used to automate the modeling process that trains a network/model from a given experimental dataset to calculate the output directly using high-dimensional complex raw data. However, the trained network is an inverse of the welding process (forward process) that produces the welding phenomena/measured raw data as the output with the penetration as the input of the forward process. Now the question is in addition to the current state of the weld penetration to be estimated if the forward process also has other inputs to determine its output. If it has, then the inverse model has to be constructed accordingly. This will call for a new foundation for deep learning-based monitoring of penetration. This letter proposed a novel innovative generative adversarial network (GAN) with GRU (Gated Recurrent Unit) in the generator, i.e., GRU-GAN, to model the extremely complex forward process to generate the observed topside welding image (output of the forward process) from the backside images (as comprehensive quantification of weld penetration). It is found that the produced topside welding image is not only determined by the current backside image but also by its history. A new foundation thus must be established to guide deep learning-based monitoring of weld penetration. The prediction model/network as an inverse model must be in compliance with the forward process that includes the history of the state of the weld penetration as its input.