Predict Turbine Failures from Sensor Data

Business Context

AeroGrid operates 1,200 industrial gas turbines across power plants in North America. Unplanned turbine downtime is expensive, so the reliability team wants a model that predicts whether a turbine will require a maintenance intervention within the next 7 days, using historical sensor and maintenance data.

Dataset

The training data is built at the turbine-day level from SCADA logs, alarm events, and maintenance records.

• Size: 410K turbine-days across 36 months, 55 features
• Target: Binary label indicating whether a turbine has a failure or maintenance-triggering fault within the next 7 days
• Class balance: Highly imbalanced, about 4.6% positive
• Missing data: 12% missing in some sensor windows due to telemetry gaps; 6% missing in maintenance-history fields for newly onboarded turbines

Success Criteria

A production-ready model should achieve strong early-warning performance: PR-AUC above 0.45, recall above 75% at precision of at least 35%, and top-decile lift above 4x versus random ranking.

Constraints

• Predictions run every hour in batch and must score all active turbines in under 5 minutes
• Reliability engineers need interpretable drivers behind each alert
• Data leakage from future maintenance actions or post-failure signals is not allowed
• Model retraining should be feasible monthly on standard cloud CPU instances

Deliverables

1. Build a binary classification pipeline to predict 7-day turbine failure risk.
2. Explain feature engineering choices for time-windowed sensor data.
3. Justify model selection and validation strategy for temporal data.
4. Evaluate the model with metrics appropriate for rare-event prediction.
5. Describe how you would set an alert threshold for maintenance operations.