Early In-Process Prediction of GMA Weld Quality from Laser Doppler Vibrometer Signals

Post-process weld quality assurance detects a defect only after the joint has been completed and therefore cannot intervene while an unstable weld is still being produced. The question addressed here is how much of a seam must be observed before its final quality state can be recognized. This study investigates whether the final quality state of a gas metal arc (GMA) weld can be recognized from the initial segment of the seam using only a laser Doppler vibrometer as the diagnostic sensor. Thirty-two bead-on-plate GMA welds were produced under stable and deliberately destabilized conditions. The welding current channel was used only for offline arc segmentation and independent reference, whereas classification used vibration-derived time, frequency, envelope, and wavelet features extracted in 0.10 s windows. Causal prefix features were evaluated with leave-one-weld-out validation and weld-level bootstrap confidence intervals, and a streaming variant updated the probability of a defective weld window by window. The final process-stability state was distinguishable after approximately 5 mm of seam (AUC = 0.94), with high discrimination through the early-to-mid seam. As an online alarm, the aggregate prefix rule produced no false alarms among six acceptable welds, detected 95% of the defective welds, and achieved a median recognition lead of ~3.6 s. However, these specificity and lead-time estimates are coarse because only six acceptable welds were available, and with a single exception, the defective welds were unstable throughout; thus, the results mainly reflect early recognition of an established state. Decision latency was nearly constant in time (0.15–0.25 s), so the spatial decision position and remaining lead time depended mainly on welding speed. The results support non-contact early recognition of unstable GMA welding, while larger balanced datasets with mid-weld transitions are needed to validate true defect-onset anticipation.