NASA CMAPSS FD001 — Predictive Maintenance System

Predictive maintenance system built on NASA CMAPSS FD001, designed for systems-oriented and safety-critical data roles, It is a training → artifacts → inference pipeline made for gauging engine health and outputting a recommendation and status based on model prediction.

---

What this system does

• Trains a baseline Remaining Useful Life (RUL) predictor using FD001
• Produces versioned artifacts (model, scaler, schema, metrics) for reproducible inference
• Runs inference with:
  • strict input schema validation (no missing/extra features)
  • feature ordering enforcement
  • numeric validity checks
  • distribution-based anomaly and extrapolation risk flags
  • conservative decision policy using a lower confidence bound
  • structured JSONL logging for audit trails

---

Dataset and scope

• Dataset: CMAPSS FD001 (single operating condition, single fault mode)
• Raw data: immutable (not tracked here)
• Processed dataset (single source of truth for all runs):
• data/processed/fd001_processed.csv
• Shared loader (all files load the same processed data):
• src/data_loader.py -> load_fd001(data_path, rul_col='RUL', cap_rul=125)
• no sensor transformations

Selected features (from EDA): sensor_4, sensor_11, sensor_15

---

Baseline model

Trains a linear regression baseline using an engine-level split and writes metrics and metadata to artifacts/.

• Split: engine-level GroupShuffleSplit (leakage prevented and assertions added)
• Scaling: StandardScaler fit on train only
• Model: OLS via LinearRegression
• Performance (validation):
• MAE ≈ 19 | RMSE ≈ 23 | R² ≈ 0.68
• Remaining error is structural (missing time context), not overfitting.

---

Repository layout

nasa-cmaps-pipeline/
├── src/
│   ├── __init__.py
│   ├── data_loader.py
│   ├── training/
│   │   ├── __init__.py
│   │   └── train_baseline.py
│   └── inference/
│       ├── __init__.py
│       ├── predict.py
│       └── test_predict.py
├── data/processed/fd001_processed.csv
├── artifacts/
│   ├── feature_schema.json
│   ├── metrics.json
│   ├── model.joblib
│   └── scaler.joblib
├── logs/inference.jsonl
├── notebooks/
│   ├── eda.ipynb
│   └── modeling_baseline.ipynb
├── README.md
└── requirements.txt

Artifacts & auditability

Training produces a versioned artifact bundle under artifacts/ that defines the inference contract and enables reproducible, auditable runs.

Artifacts

• model.joblib (gitignored) — trained OLS baseline model
• scaler.joblib (gitignored) — train-only StandardScaler
• feature_schema.json (committed) — locked feature names and ordering
• metrics.json (committed) — model metadata and validation results

Each artifact bundle contains

• artifact_version = "baseline_v1"
• explicit feature list and target (RUL_capped)
• held-out validation metrics
• model coefficients and intercept
• training distribution statistics (mean / std / min / max) used for runtime safety checks
• SHA256 fingerprint of fd001_processed.csv to guarantee data lineage and reproducibility

Artifacts are treated as immutable once written. All inference runs against an explicit, versioned snapshot.

---

Inference contract observability

Inference is intentionally defensive and assumes inputs may be malformed, out-of-distribution, or unsafe.

src/inference/predict.py enforces:

• exact feature schema (no missing features, no extras)
• strict numeric validity checks (np.isfinite)
• extrapolation risk flags when inputs fall outside training min/max
• z-score anomaly detection relative to training mean/std (with zero-std guards)
• conservative lower-bound estimate:
  rul_lower = rul_pred - k × MAE

Operational decisions are based on rul_lower, not the point estimate:

• GREEN — Continue operation
• AMBER — Inspect
• RED — Schedule maintenance
• CRITICAL — Remove from service

Every inference outputs a structured JSON record to:

• logs/inference.jsonl

providing a complete audit trail for post-hoc analysis.

---

How to run

1. Install dependencies

pip install -r requirements.txt

2. Train baseline and write artifacts

python -m src.training.train_baseline

3. Run a single inference example

python -m src.inference.test_predict

---

License

MIT License © 2026 Christopher Orellana