Acoustic Data-based Characterization of Incipient Boiling for Space Applications
Cal Poly SLO Data Science Senior Capstone, NASA Ames Research Center
Overview
Localized heat leaks in cryogenic fuel systems can trigger boil-off, tank over-pressurization, and long-duration mission failure. This project uses acoustic accelerometer data and unsupervised machine learning to identify distinct boiling regimes, translating raw vibration signals into physically interpretable cluster groupings that can inform future space fuel management systems.
Team: Hailey Ernest, Colin Hassett, Elliot Kunz, Cameron Hafer · Advisors: Dr. Kelly Bodwin, Dr. Alex Dekhtyar · Cal Poly SLO · Statistics, Mathematics, Computer Science
Methodology
Feature Engineering: 70 statistical features (34 per accelerometer plus 2 pairwise) engineered from time and frequency domains, designed to capture signal structure rather than raw amplitude. Per-feature linear regression detrending and hyperparameter tuning remove runtime-correlated bias; pairwise cross-accelerometer features capture differences between both sensors, a correction to prior work that used only one.
Dimensionality Reduction: PCA, diffusion maps, and UMAP were evaluated head-to-head. UMAP produced the clearest separation and was selected for the final pipeline.
Clustering: K-Means, HDBSCAN, and spectral clustering were compared. HDBSCAN was chosen for its ability to handle variable-density clusters and naturally identify outliers. Final pipeline: UMAP + HDBSCAN.
Validation & Real-Time Classification: Clusters were validated with leave-one-out resampling, achieving 99% clustering stability across parameter changes. A Gaussian Mixture Model fit on a frozen UMAP embedding classifies incoming data in real time without retraining, supporting early boiling detection in space-based fuel systems.
Results: Six Boiling Regimes
- Lack of Boiling: filtered by noise-floor threshold before clustering
- Sporadic: rare, isolated spikes with no consistent timing
- Chaotic: irregular bursts of varying height and frequency
- Rhythmic: consistent, repeating spike patterns
- Dense / Double Rhythmic: distinct subclusters within rhythmic boiling
- Single Rhythmic with Climax: rhythmic behavior with a distinctive amplitude peak
UMAP + HDBSCAN: three active boiling clusters
Subclustering within the rhythmic regime
Research Poster
Full Report
Navigate the 13-page final report below.
