Portfolio Prototype
Predictive Maintenance
Predictive-maintenance prototype on simulated sensor data, with failure classification, RUL regression, Prefect workflows, FastAPI endpoints and a Dash demo.
GitHub ↗Overview
Explores failure-risk classification and remaining-useful-life (RUL) regression using simulated sensor trajectories.
Architecture
- The sensor simulator generates engine degradation trajectories (src/data/simulator.py).
- Prefect flows connect validation, labels, features, classification/regression training and model saving (src/pipeline/flows.py).
- FastAPI exposes inference endpoints; Dash provides a demonstration interface (src/api/ and src/dashboards/web_dashboard.py).
Engineering decisions
- Separate ingestion, feature engineering and training into Prefect flows.
- Use separate models for failure classification and RUL regression.
- Keep demonstration behavior available through heuristic fallback inference when model artifacts are absent.
What is implemented
- Sensor simulation, labeling and feature engineering.
- Model training and evaluation in src/models/training.py.
- Failure, RUL, batch and what-if endpoints in src/api/routes.py.
- Dash demonstration and CI/deployment configuration.
Current limitations
- The data is simulated; no industrial deployment is verified.
- The API uses heuristic predictions when saved models are absent, and the anomaly endpoint uses simple sensor thresholds.
- RUL intervals and what-if cost savings are illustrative rather than calibrated uncertainty or demonstrated business results.
- Cloud deployment files do not establish a verified GCP deployment.