Georgia Tech · time-series ML · failure detection · evaluation

Failure detection that is tested on geometries the model has not seen before.

This Georgia Tech research develops reproducible ML workflows for identifying simulator failures in time-series thrust data, with emphasis on data auditing, leakage prevention, and generalization beyond duplicated geometries.

Rocket-motor thrust time-series and failure detection workflow — Campaign auditing and geometry-aware splitting come before classification so the reported result measures generalization.

Why the data audit comes first

A simulation campaign can fail silently: a row may exist even when the underlying run is invalid. The workflow therefore starts with campaign auditing and candidate-failure identification before any classifier is trained.

Method

I built the pipeline in Python with preprocessing, feature engineering, failure classification, clustering/statistical analysis where useful, and evaluation controls designed to prevent duplicate geometry leakage.

Generalization result

After removing duplicate-geometry leakage, the classifier reaches 96.8% accuracy and 0.986 ROC AUC on unseen geometries under the current study definition.

Why it belongs in an ML portfolio

The value of the project is the full chain: audit the data, identify the failure mechanism, choose a split that reflects deployment, quantify performance, and verify the result independently. That same evaluation discipline carries into recommendation, retrieval, and model-deployment work.