Business Context

Northline Components operates 18 manufacturing lines across 3 plants and loses significant production time to unplanned machine stoppages. The operations team wants a model that predicts whether a machine will fail in the next 7 days so maintenance can be scheduled before breakdowns occur.

Dataset

You are given historical machine-level daily records built from sensor streams, maintenance logs, and machine metadata.

• Size: 420K machine-day rows covering 30 months, 52 features
• Target: Binary label indicating whether the machine fails within the next 7 days
• Class balance: Failures are rare; about 4.6% positive class
• Missing data: 12% missing in some sensor fields due to telemetry dropouts; 6% missing in maintenance fields for newly installed machines

Success Criteria

A good solution should identify at least 75% of upcoming failures while keeping precision high enough that maintenance teams are not overwhelmed. The model should produce ranked risk scores and explain the main drivers of each alert.

Constraints

• Batch scoring runs every 6 hours for ~9,000 active machines
• Predictions must be explainable to reliability engineers
• False negatives are costly, but excessive false positives create unnecessary maintenance work
• Retraining should be feasible on a monthly cadence

Deliverables

1. Build a binary classification pipeline for failure prediction within 7 days.
2. Explain feature engineering, leakage prevention, and validation strategy.
3. Compare at least one baseline model against a stronger tree-based model.
4. Choose an operating threshold based on maintenance capacity and failure recall.
5. Report final test metrics and the most important features.