Deterministic Keyhole-Collapse Detection

I built deterministic keyhole-collapse detection from thermal-emission signals, with no training dataset or pretrained ML model.

AI for Manufacturing · Course project

I developed a deterministic method for keyhole-collapse detection in laser powder bed fusion to remove the need for labeled training data and a trained machine-learning model.

What I built:

  • Signal filtering and sliding-window extraction of physical features from thermal-emission measurements.
  • PCA and Mahalanobis-distance scoring calculated directly from the input signal: no separate training dataset or pretrained ML model.
  • Time-aligned scores and candidate-event visualizations, compared with my PI-TSAD method as a reference.

Tools: Python · NumPy · SciPy · statistical signal processing · PCA · Matplotlib

Thermal-emission signal with PCA and random-forest event predictions and their corresponding scores
Deterministic statistical scoring compared with my PI-TSAD baseline. The random forest is shown only for comparison; it is not used by this detector.
UMAP embedding of window features with supplied collapse labels shown in red
Exploratory feature visualization from the study. This embedding is not part of the deterministic detector.

The code calculates PCA statistics from the signal being analyzed. The offline demo uses synthetic input; experimental datasets and trained models are not distributed.

GitHub repository · Back to projects