Towards self explaining WAAM systems : from causal graph discovery to vision interpretation in weld dynamics

Abstract Reliable real-time monitoring of weld quality is essential for the industrial adoption of wire arc additive manufacturing (WAAM), because unstable arc and melt-pool behaviour produces defects that are costly to find after deposition. This work proposes the Causal Fusion Transformer with a vision–language module (CFT+VLM) to address this problem. Here, causal fusion means that the relationship between the parameters (modalities) is constrained to follow the physical order of the process: process parameters (current, voltage, wire feed speed and travel speed) combination leading to melt-pool video features and corresponding variations, then weld events detected by YOLOv8 (bead formation, spatter and fire spark). The framework was evaluated on videos (3,344 clips) corresponding to 11 WAAM deposition conditions. The proposed CFT + VLM architecture was then implemented on these and compared to conventional image based It produced well-separated latent clusters (silhouette score 0.70) and classified three pseudo-quality classes with an accuracy of 0.76, an F1 score of 0.7652 and an AUC of 0.8681. Grad-CAM maps and a causal graph link the predictions to the arc and melt-pool regions and to the process parameters, which provides an interpretable basis for automated WAAM quality monitoring.