PI-TSAD: Detecting Keyhole Collapse

Photodiode-based detection validated against synchrotron X-ray labels and transferred to eight commercial builds without retraining.

I developed PI-TSAD, a physically informed time-series anomaly detector, to identify keyhole collapse from photodiode signals during laser powder bed fusion.

  • Analyzed 200 kHz photodiode data and selected a 0.25-30 kHz filter to isolate melt-pool dynamics.
  • Extracted statistical and frequency-domain features and trained a Random Forest classifier to identify collapse events.
  • Validated event timing on four held-out melt tracks against synchrotron X-ray labels from the Advanced Photon Source (APS).
  • Transferred the detector to eight AlSi10Mg builds on a commercial DMG MORI LASERTEC 12 SLM without retraining.
PI-TSAD collapse-event predictions compared with X-ray ground-truth labels for four held-out melt tracks
Collapse predictions for four held-out tracks, compared with synchrotron X-ray labels. Figure from the PI-TSAD paper.

Across the commercial builds, anomaly counts followed the CT pore-count trend. The lack-of-fusion sample exposed a limitation: a detector trained on keyhole collapse does not capture every mechanism that produces porosity.

PI-TSAD anomaly counts and baseline peak counts compared with CT pore counts for eight AlSi10Mg builds
Comparison with CT measurements across eight builds. Figure from the PI-TSAD paper.

This first-author work was published in the Journal of Manufacturing Processes (2026) and received the NAMRC Outstanding Paper Award in Manufacturing Processes and Student Research Presentation Award.

Read the paper · Code

Back to projects