Identifying anomalous welding in the bud: A forecasting approach

Abstract

To forecast anomalous welding processes, we propose a novel two-stage framework that integrates generative models and adversarial learning techniques for predicting anomalies in molten pool behavior. In the first stage, the goal is to generate molten pool videos (MPVs) for future welding operations by sequentially predicting molten pool frames under consistent welding parameters. The second stage uses one-class classification on the generated molten pool images to detect anomalies. This is done by maximizing the discrepancy between outliers (anomalies) and inliers (normal behavior) while minimizing the variation within the inliers. By leveraging the generative error introduced by spatiotemporal prediction, the framework enhances the separability between normal inliers and anomalous outliers. The proposed framework was evaluated by identifying anomalies in a variety of weld seams. Our results demonstrate that the framework successfully forecasts welding anomalies on real-world MPV datasets, highlighting its potential for practical applications in defect detection and process control.