VesselSeed external bearing study · 21 November 2026

The frozen study evaluates one full bearing-specific implementation of the
waveform → prime field → coherent peaks → score path in US20260220223A1.
The patent allows multiple implementations. The bearing-specific input scale,
peak readout, hyperparameters and trained classifier are choices documented in
protocol.txt and field.py. This is newly written research code. Patent scope,
optional embodiments and industrial reliability are separate questions.

Data and attribution
Calibration: C. Schnur et al., Saarland University bearing dataset.
https://doi.org/10.5281/zenodo.15376390 (CC BY 4.0).
External test: Hoang Si Hong and Nguyen Thuan, HUST bearing dataset, v3.
https://doi.org/10.17632/cbv7jyx4p9.3 (CC BY 4.0).
HUST article, dimensions and acquisition: https://doi.org/10.1186/s13104-023-06400-4
Reference algorithm: https://patents.google.com/patent/US20260220223A1/en
File hashes identify each processed recording. The recordings are available
from the original repositories. No HUST training or threshold tuning was done.
The operating-point comparison marked "external matched" uses the test labels
only to choose cutoff ranks for evaluation. The training-locked cutoffs
represent direct deployment from the original rig.

Reproduce from this extracted directory (Python 3.12).
  python -m pip install -r requirements.txt
  python training/fetch.py
  curl -fL https://data.mendeley.com/public-api/zip/cbv7jyx4p9/download/3 -o hust-v3.zip
  python study.py training hust-v3.zip replay
  python verify.py replay

The fetch step produces the first study's hashed two-second samples. The
HUST ZIP is about 697 MB. Choose a new output directory for another run.
verify.py checks calibration hashes, MAT hashes, per-case scores and decisions
against the published run. Processing time is machine-dependent.
