T6 - kijouneli/EAR-MP-2025 GitHub Wiki
Title
Gearbox Fault-Severity & Confidence Estimation
Summary
Develop a model that classifies gearbox pitting severity (11 grades) and outputs prediction confidence for previously unseen speeds & torque. The training set spans 15 rotational speeds × 6 torque levels and covers seven health states (healthy + six faulty), whereas the hidden test/validation sets extend to 18 unseen speeds and the full eleven-level degradation spectrum. This deliberate coverage gap simulates real PHM practice and demands robust generalization to fault levels and conditions absent during training. Leaderboard scoring combines a proximity-based ordinal penalty with a confidence factor, rewarding accurate, well-calibrated, and trustworthy forecasts.
Deliverables
- Model & code (GitHub)
- 4-page short paper
- Presentation materials
- Demo video
Expected number of team members
4 students
Expected duration in month
5 months
Data sets
- Training and test datasets are available at Data download
- Around 6 GB data set: 3-axis vibration, 15 RPM × 6 torque conditions (train) + extended test/valid sets, CSV/NPY archives.
GPU cloud server
CPU-only baseline feasible in < 2 h
Additional Information
- Official data set of the PHM North America 2023 Conference Data Challenge (competition closed 25 Aug 2023). The task was to estimate gearbox pitting-fault severity (11 ordinal levels) and output a confidence flag for each record.
- License: Re-use is free for research and educational purposes provided proper attribution
- Official description page (includes problem statement, scoring metric, exemplar submission, and reference papers) : https://data.phmsociety.org/phm2023-conference-data-challenge/