Welding process analysis through sequential classification of process images

Automated welding process inspection is crucial to ensure good execution of the process. With the rise of Artificial Intelligence based solutions several proposals were made in this regard. In this research we study the possibilities of detecting welding process instabilities in sequential data and using sequential deep learning models. We propose a new sequential dataset that consists of four process states to validate our approach. These process states are as follows: no welding, welding and stable, welding and deviated (or shifted), welding and unstable. What characterizes this dataset is the introduction of temporal characteristic through sequential images, and the use of active vision. We test the proposed dataset on a state-of-the-art sequential classification model: ConvLSTM. The tested model is capable of successfully retrieving the patterns in the proposed dataset and classifying the different states of the process represented by a sequence of ten images, at an accuracy of 92%.