Temporal Evidence Accumulation Trades Alarm Purity Against Detection Delay in Synthetic Melt Pool Monitoring

Abstract

This fully synthetic study separates spatial frame classification from temporal evidence accumulation in melt pool monitoring. A seeded generator creates 9,600 radiance images across 150 independent sixty four frame tracks, including sustained morphology changes, track offsets and isolated spatter flashes. Penalized logistic regression learns a score from five explicit image features using sixty training tracks; twenty validation tracks determine score thresholds, and seventy test tracks provide evaluation. A two state Bayesian filter converts classifier probabilities to likelihood ratios and accumulates causal evidence. On 4,480 test frames, frame logistic F1 is 0.339, nearly identical to thresholded brightness at 0.341. Temporal F1 increases to 0.604 while false positive frames fall from 126 to 96. However, detected events decrease from sixty to fifty seven of sixty three, and mean delay rises from 2.42 to 3.65 frames. Permuting frame order reduces temporal F1 to 0.211. The results demonstrate an explicit purity versus delay tradeoff for generated morphology events, not defect detection in manufactured parts.