Aerospace Prognostics
Deployable end-to-end PHM MLOps, not another leaderboard notebook: NASA C-MAPSS turbofan RUL and ESA spacecraft-telemetry anomaly detection carried through their real evaluation protocols, wrapped in a FastAPI serving API, an operator console, signed release evidence (model card, SBOM, provenance), drift monitoring, and 453 tests. The evaluation layer proved general enough to extract as telemeval, a standalone library on PyPI with a Zenodo DOI.
- Python
- FastAPI
- Streamlit
- Docker
- GitHub Actions
- pytest
- PyPI
What it is
Most predictive-maintenance repos are a notebook and a leaderboard score. This one is an operations system: a FastAPI inference service, a Streamlit operator console, signed release evidence (model card, SBOM, provenance), and drift monitoring — around models trained and evaluated under honest, documented protocols. The engineering envelope is the point; the models are held to it.
Two benchmark tracks
- C-MAPSS turbofan RUL — the familiar baseline anyone can cross-check: sequence models, diagnostics, and calibrated prediction artifacts.
- ESA-ADB spacecraft anomaly detection — a far less picked-over benchmark, run through its real event-wise protocol (plus SMAP/MSL plumbing).
The leakage audit — and the corrected number
The first ESA-ADB run reported recall 0.24. Auditing the evaluation showed the chronological split was counting training-window events — which have no test-window predictions — as missed, deflating recall. Restricting scoring to test-window events (the correct protocol) gives the honest 0.42, and the corrected number is the one reported, with the audit written up. Every artifact records its protocol deviations and is stamped "event-wise detection only — not a full ESA-ADB leaderboard claim."
telemeval — the extracted library
The anomaly-evaluation layer (including the leakage guard born from that audit) proved general enough to extract: telemeval is now a standalone Apache-2.0 library on PyPI with a Zenodo DOI for leakage-safe, event-wise and affiliation-based evaluation of spacecraft-telemetry anomaly detection — and this repository is its reference pipeline, consuming it as a dependency.
Delivery
FastAPI inference service, Streamlit review console, Docker Compose stack, and CI running ruff, pytest, a dependency audit, SBOM generation, and serving-image smoke tests. Release evidence covers model inspection, validation, benchmark, model card, SBOM, release bundle, promotion report, and provenance.
Honest limits
A reference implementation of end-to-end PHM MLOps under MIT — not a product launch track (productization docs are frozen as archived context). The hosted review console is token-gated rather than public, so the visual here is the console itself rather than a click-through demo. Raw telemetry and generated model artifacts stay out of Git.