Predictive maintenance
Anomaly detection and failure prediction on sensor time series, wired to alerts and work orders.
- Where
- Mercedes-Benz · Veeco
- When
- 2024 – now
- Context
- Line and tool sensor data
- Type
- ML
−15%unplanned downtime on pilot lines
6 hearlier drift detection than static alarms
0.86F1 on failure prediction
How it works
- 1SenseTemperature, pressure, flow
- 2ShapeWindows and rolling features
- 3ScoreAnomaly and failure risk
- 4AlertThresholds set with maintenance
- 5FixWork order opened
The problem
Static alarm thresholds fire late, or not at all, when a process drifts slowly. By the time a failure is obvious the line is already down.
What I built
- Trained Isolation Forest and autoencoder models on high-frequency sensor data (temperature, pressure, flow).
- Trained XGBoost and LightGBM models to predict component failures.
- Tracked experiments and models in MLflow.
- Wired model outputs to alerts and work orders so maintenance teams act before a failure.
Stack
Isolation ForestAutoencodersXGBoostLightGBMscikit-learnPyTorchMLflow