Dynamic prediction of backside weld width: factor analysis and progressive deep learning enhancements

In scenarios where direct observation of backside weld width is impractical, Variable Polarity Plasma Arc Welding (VPPAW) systems increasingly rely on topside welding images for predictive analysis. Unlike early methods that used weld pool features (e.g., length, width, area) and welding parameters (e.g., current, speed) for regression or machine learning, deep learning provides superior accuracy. However, most research has focused on model optimization, with limited attention to the factors contributing to prediction errors. Therefore, this study systematically investigates these factors, including feature fusion, temporal sequence integration, model efficiency, and architecture. Based on this analysis, a novel framework is proposed that integrates ShuffleNetV2 for feature extraction, GRU networks to capture temporal heat input dynamics, and an Attention mechanism to adaptively weight temporal frames. This approach improves prediction accuracy by addressing the dynamic and temporal characteristics of welding. Experimental results reveal significant improvements over traditional CNN models, offering valuable insights into dynamic weld prediction. The proposed VPPAW deep learning model provides a robust foundation for further exploration of factors influencing backside weld width prediction.