A comparative study of deep learning models for welding process analysis using sequential models

Abstract Automated inspection of the welding process is essential for ensuring consistent quality and high operational reliability. While the opportunities offered by the recent advances in artificial intelligence have allowed for good progress in automated process monitoring, most proposals rely on static image analysis and overlook the temporal dynamics inherent to welding operations. In this work, we introduce a sequential dataset that we designed to capture four distinct welding process states: no welding, welding stable, welding deviated, welding unstable. This dataset differs from the state-of-the-art datasets by its temporal structure and the active vision setup, which enables the study of the dynamic behavior of the process. The introduced temporal dimension is essential for capturing the complex patterns associated with welding defects, particularly in scenarios involving weaving motions. Additionally, we evaluate and compare different state-of-the-art sequential data analysis deep learning models. The study carried out aims to assess the capabilities of these models to detect process instabilities from video sequences in real time and to establish a baseline for future research in intelligent welding process monitoring. Upon this benchmark, we provide recommendations on the most suitable deep learning models for similar applications involving sequential image analysis of the welding process. The obtained results demonstrated that the proposed dataset provides informative clues about the spatial and temporal state of the process. Notably, the MC3 model was capable of achieving the best compromise between performance and computational efficiency, classifying the state of the process at an accuracy of 99.12% with a relatively small latency.